* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

# Download Quantum Techniques for Stochastic Mechanics

EPR paradox wikipedia , lookup

Self-adjoint operator wikipedia , lookup

Bell's theorem wikipedia , lookup

Compact operator on Hilbert space wikipedia , lookup

Bra–ket notation wikipedia , lookup

Molecular Hamiltonian wikipedia , lookup

Quantum state wikipedia , lookup

Interpretations of quantum mechanics wikipedia , lookup

Hydrogen atom wikipedia , lookup

Scalar field theory wikipedia , lookup

Theoretical and experimental justification for the Schrödinger equation wikipedia , lookup

Schrödinger equation wikipedia , lookup

History of quantum field theory wikipedia , lookup

Path integral formulation wikipedia , lookup

Perturbation theory wikipedia , lookup

Dirac equation wikipedia , lookup

Quantum electrodynamics wikipedia , lookup

Probability amplitude wikipedia , lookup

Hidden variable theory wikipedia , lookup

Symmetry in quantum mechanics wikipedia , lookup

Noether's theorem wikipedia , lookup

Canonical quantization wikipedia , lookup

Quantum Techniques for Stochastic Mechanics John C. Baez1,2 and Jacob D. Biamonte3 1 2 3 Department of Mathematics University of California Riverside, CA 92521, USA Centre for Quantum Technologies National University of Singapore Singapore 117543 Quantum Complexity Science Initiative Department of Physics University of Malta MSD 2080, Malta email: [email protected], [email protected] Abstract Some ideas from quantum theory are just beginning to percolate back to classical probability theory. For example, there is a widely used and successful theory of ‘chemical reaction networks’, which describes the interaction of molecules in a stochastic rather than quantum way. Computer scientists use a different but equivalent formalism called ‘stochastic Petri nets’ to describe collections of randomly interacting entities. These two equivalent formalisms also underlie many models in population biology and epidemiology. Surprisingly, the mathematics underlying these formalisms is very much like that used in the quantum theory of interacting particles—but modified to use probabilities instead of complex amplitudes. In this text, we explain this fact as part of a detailed analogy between quantum mechanics and the theory of random processes. To heighten the analogy, we call the latter ‘stochastic mechanics’. We use this analogy to explain two major results in the theory of chemical reaction networks. First, the ‘deficiency zero theorem’ gives conditions for the existence of equilibria in the approximation where the number of molecules of each kind is treated as varying continuously in a deterministic way. Our proof uses tools borrowed from quantum mechanics, including a stochastic analogue of Noether’s theorem relating symmetries and conservation laws. Second, the ‘Anderson–Craciun–Kurtz theorem’ gives conditions under which these equilibria continue to exist when we treat the number of molecules of each kind as discrete, varying in a random way. We prove this using another tool borrowed from quantum mechanics: coherent states. We also investigate the overlap of quantum and stochastic mechanics. Some Hamiltonians can describe either quantum-mechanical or stochastic processes. These are called ‘Dirichlet operators’, and they have an intriguing connection to the theory of electrical circuits. In a section on further directions for research, we describe how the stochastic Noether theorem simplifies for Hamiltonians that are Dirichlet operators, and explain some connections between stochastic Petri nets and computation. 1 Foreword This book is about a curious relation between two ways of describing situations that change randomly with the passage of time. The old way is probability theory and the new way is quantum theory. Quantum theory is based, not on probabilities, but on amplitudes. We can use amplitudes to compute probabilities. However, the relation between them is nonlinear: we take the absolute value of an amplitude and square it to get a probability. It thus seems odd to treat amplitudes as directly analogous to probabilities. Nonetheless, if we do this, some good things happen. In particular, we can take techniques devised in quantum theory and apply them to probability theory. This gives new insights into old problems. There is, in fact, a subject eager to be born, which is mathematically very much like quantum mechanics, but which features probabilities in the same equations where quantum mechanics features amplitudes. We call this subject stochastic mechanics. Plan of the book In Section 1 we introduce the basic object of study here: a ‘stochastic Petri net’. A stochastic Petri net describes in a very general way how collections of things of different kinds can randomly interact and turn into other things. If we consider large numbers of things, we obtain a simplified deterministic model called the ‘rate equation’, discussed in Section 2. More fundamental, however, is the ‘master equation’, introduced in Section 3. This describes how the probability of having various numbers of things of various kinds changes with time. In Section 4 we consider a very simple stochastic Petri net and notice that in this case, we can solve the master equation using techniques taken from quantum mechanics. In Section 5 we sketch how to generalize this: for any stochastic Petri net, we can write down an operator called a ‘Hamiltonian’ built from ‘creation and annihilation operators’, which describes the rate of change of the probability of having various numbers of things. In Section 6 we illustrate this with an example taken from population biology. In this example the rate equation is just the logistic equation, one of the simplest models in population biology. The master equation describes reproduction and competition of organisms in a stochastic way. In Section 7 we sketch how time evolution as described by the master equation can be written as a sum over Feynman diagrams. We do not develop this in detail, but illustrate it with a predator–prey model from population biology. In the process, we give a slicker way of writing down the Hamiltonian for any stochastic Petri net. In Section 8 we enter into a main theme of this course: the study of equilibrium solutions of the master and rate equations. We present the Anderson– Craciun–Kurtz theorem, which shows how to get equilibrium solutions of the master equation from equilibrium solutions of the rate equation, at least if a cer- 2 tain technical condition holds. Brendan Fong has translated Anderson, Craciun and Kurtz’s original proof into the language of annihilation and creation operators, and we give Fong’s proof here. In this language, it turns out that the equilibrium solutions are mathematically just like ‘coherent states’ in quantum mechanics. In Section 9 we give an example of the Anderson–Craciun–Kurtz theorem coming from a simple reversible reaction in chemistry. This example leads to a puzzle that is resolved by discovering that the presence of ‘conserved quantities’—quantitites that do not change with time—let us construct many equilibrium solutions of the rate equation other than those given by the Anderson– Craciun–Kurtz theorem. Conserved quantities are very important in quantum mechanics, and they are related to symmetries by a result called Noether’s theorem. In Section 10 we describe a version of Noether’s theorem for stochastic mechanics, which was proved with the help of Brendan Fong. This applies, not just to systems described by stochastic Petri nets, but a much more general class of processes called ‘Markov processes’. In the analogy to quantum mechanics, Markov processes are analogous to arbitrary quantum systems whose time evolution is given by a Hamiltonian. Stochastic Petri nets are analogous to a special case of these: the case where the Hamiltonian is built from annihilation and creation operators. In Section 11 we state the analogy between quantum mechanics and stochastic mechanics more precisely, and with more attention to mathematical rigor. This allows us to set the quantum and stochastic versions of Noether’s theorem side by side and compare them in Section 12. In Section 13 we take a break from the heavy abstractions and look at a fun example from chemistry, in which a highly symmetrical molecule randomly hops between states. These states can be seen as vertices of a graph, with the transitions as edges. In this particular example we get a famous graph with 20 vertices and 30 edges, called the ‘Desargues graph’. In Section 14 we note that the Hamiltonian in this example is a ‘graph Laplacian’, and, following a computation done by Greg Egan, we work out the eigenvectors and eigenvalues of this Hamiltonian explicitly. One reason graph Laplacians are interesting is that we can use them as Hamiltonians to describe time evolution in both stochastic and quantum mechanics. Operators with this special property are called ‘Dirichlet operators’, and we discuss them in Section 15. As we explain, they also describe electrical circuits made of resistors. Thus, in a peculiar way, the intersection of quantum mechanics and stochastic mechanics is the study of electrical circuits made of resistors! In Section 16, we study the eigenvectors and eigenvalues of an arbitrary Dirichlet operator. We introduce a famous result called the Perron–Frobenius theorem for this purpose. However, we also see that the Perron–Frobenius theorem is important for understanding the equilibria of Markov processes. This becomes important later when we prove the ‘deficiency zero theorem’. We introduce the deficiency zero theorem in Section 17. This result, proved by the chemists Feinberg, Horn and Jackson, gives equilibrium solutions for the rate equation for a large class of stochastic Petri nets. Moreover, these 3 equilibria obey the extra condition that lets us apply the Anderson–Craciun– Kurtz theorem and obtain equilibrium solutions of the master equations as well. However, the deficiency zero theorem is best stated, not in terms of stochastic Petri nets, but in terms of another, equivalent, formalism: ‘chemical reaction networks’. So, we explain chemical reaction networks here, and use them heavily throughout the rest of the course. However, because they are applicable to such a large range of problems, we call them simply ‘reaction networks’. Like stochastic Petri nets, they describe how collections of things of different kinds randomly interact and turn into other things. In Section 18 we consider a simple example of the deficiency zero theorem taken from chemistry: a diatomic gas. In Section 19 we apply the Anderson– Craciun–Kurtz theorem to the same example. In Section 20 we begin working toward a proof of the deficiency zero theorem, or at least a part of this theorem. In this section we discuss the concept of ‘deficiency’, which had been introduced before, but not really explained: the definition that makes the deficiency easy to compute is not the one that says what this concept really means. In Section 21 we show how to rewrite the rate equation of a stochastic Petri net—or equivalently, of a reaction network— in terms of a Markov process. This is surprising because the rate equation is nonlinear, while the equation describing a Markov process is linear in the probabilities involved. The trick is to use a nonlinear operation called ‘matrix exponentiation’. In Section 22 we study equilibria for Markov processes. In Section 23, we use these equilibria to obtain equilibrium solutions of the rate equation, completing our treatment of the deficiency zero theorem. We conclude the book by sketching some optional further topics: Noether’s theorem for Markov processes generated by Dirichlet operators in Section 24, and connections between computation and Petri nets in Section 25. Acknowledgements These course notes are based on a series of articles on the Azimuth blog. The original articles are available via this webpage: • Network Theory, http://math.ucr.edu/home/baez/networks/. On the blog you can read discussion of these articles, and also make your own comments or ask questions. We thank the readers of Azimuth for many helpful online discussions, including David Corfield, Manoj Gopalkrishnan, Greg Egan, Arjun Jain and Blake Stacey, but also many others we apologize for not listing here. We especially thank Brendan Fong for his invaluable help in finding a ‘quantum proof’ of the Anderson–Craciun–Kurtz theorem and proving the stochastic version of Noether’s theorem. We thank David Doty, David Soloveichik, Erik Winfree, and others at the for inviting the first author to Programming with Chemical Reaction Networks: Mathematical Foundations, a workshop at the Banff International Research Station held in June 2014, and for educating him about chemical reaction networks 4 and their computational power. In particular, Section 25.3 owes its existence to a talk David Soloveichik gave in this workshop [Sol14]. We thank Jim Stuttard of the Azimuth Project for patiently sifting through a lot of Petri net software. This made Section 25.5 possible, though we used only a fraction of his work. We thank Federica Ferraris for drawing the ‘Baez-magician’, the ‘rabbit archer’ and the ‘rabbit melee’ as well as helping format several of the other pictures appearing in this book. We also thank Wikimedia Commons for the use of many pictures. Finally, we thank the Centre for Quantum Technologies for its hospitality and support, not only for ourselves but also for Brendan Fong. 5 CONTENTS CONTENTS Contents 1 Stochastic Petri nets 10 1.1 The rate equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.2 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 1.3 Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2 The 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 2.10 rate equation Rate equations: the general recipe The formation of water (1) . . . . The formation of water (2) . . . . The dissociation of water (1) . . . The dissociation of water (2) . . . The SI model . . . . . . . . . . . . The SIR model . . . . . . . . . . . The SIRS model . . . . . . . . . . The general recipe revisited . . . . Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 19 20 21 22 23 24 25 26 27 28 3 The master equation 30 3.1 The master equation . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.2 Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 4 Probabilities vs amplitudes 4.1 A Poisson process . . . . . . . . . . . . . . . . . . . . 4.2 Probability theory vs quantum theory . . . . . . . . 4.3 Stochastic vs unitary operators . . . . . . . . . . . . 4.4 Infinitesimal stochastic versus self-adjoint operators . 4.5 The moral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 39 41 43 44 47 5 Annihilation and creation operators 5.1 Rabbits and quantum mechanics . . 5.2 Catching rabbits . . . . . . . . . . . 5.3 Dying rabbits . . . . . . . . . . . . . 5.4 Breeding rabbits . . . . . . . . . . . 5.5 Dueling rabbits . . . . . . . . . . . . 5.6 Brawling rabbits . . . . . . . . . . . 5.7 The general rule . . . . . . . . . . . 5.8 Kissing rabbits . . . . . . . . . . . . 5.9 References . . . . . . . . . . . . . . . 5.10 Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 49 51 52 53 54 55 55 56 58 59 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Population biology 63 6.1 Amoeba fission and competition . . . . . . . . . . . . . . . . . . 63 6.2 The rate equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 6.3 The master equation . . . . . . . . . . . . . . . . . . . . . . . . . 65 6 CONTENTS CONTENTS 6.4 6.5 An equilibrium state . . . . . . . . . . . . . . . . . . . . . . . . . Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Feynman diagrams 7.1 Stochastic Petri nets revisited 7.2 The rate equation . . . . . . . 7.3 The master equation . . . . . 7.4 Feynman diagrams . . . . . . 8 The 8.1 8.2 8.3 8.4 8.5 8.6 8.7 68 71 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 73 74 76 78 Anderson–Craciun–Kurtz theorem The rate equation . . . . . . . . . . . . . Complex balance . . . . . . . . . . . . . The master equation . . . . . . . . . . . Coherent states . . . . . . . . . . . . . . The proof . . . . . . . . . . . . . . . . . An example . . . . . . . . . . . . . . . . Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 82 83 84 84 86 87 88 89 89 89 90 92 . . . . . . . . . . . . . . . . . . . . 9 A reversible reaction 9.1 A reversible reaction . . . . 9.2 Equilibria . . . . . . . . . . 9.3 Complex balanced equilibria 9.4 Conserved quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Noether’s theorem 10.1 Markov processes . 10.2 Noether’s theorem 10.3 Markov chains . . 10.4 Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 . 94 . 97 . 98 . 101 mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 103 105 106 107 109 110 112 12 Noether’s theorem: quantum vs stochastic 12.1 Two versions of Noether’s theorem . . . . . . . . . . . . . . . . . 12.2 Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.3 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 115 115 117 . . . . . . . . . . . . . . . . . . . . 11 Quantum mechanics vs stochastic 11.1 States . . . . . . . . . . . . . . . 11.2 Observables . . . . . . . . . . . . 11.3 Symmetries . . . . . . . . . . . . 11.4 Quantum evolution . . . . . . . . 11.5 Stochastic evolution . . . . . . . 11.6 The Hille–Yosida theorem . . . . 11.7 Answers . . . . . . . . . . . . . . 7 CONTENTS CONTENTS 13 Chemistry and the Desargues graph 13.1 The ethyl cation . . . . . . . . . . . 13.2 The Desargues graph . . . . . . . . . 13.3 The ethyl cation, revisited . . . . . . 13.4 Trigonal bipyramidal molecules . . . 13.5 Drawing the Desargues graph . . . . 13.6 Desargues’ theorem . . . . . . . . . . 13.7 Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 118 120 121 122 124 127 128 14 Graph Laplacians 135 14.1 A random walk on the Desargues graph . . . . . . . . . . . . . . 136 14.2 Graph Laplacians . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 14.3 The Laplacian of the Desargues graph . . . . . . . . . . . . . . . 141 15 Dirichlet operators 145 15.1 Dirichlet operators and electrical circuits . . . . . . . . . . . . . . 145 15.2 Circuits made of resistors . . . . . . . . . . . . . . . . . . . . . . 146 15.3 The big picture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 16 Perron–Frobenius theory 16.1 At the intersection of two theories . . . . . . . . 16.2 Stochastic mechanics versus quantum mechanics 16.3 From graphs to matrices . . . . . . . . . . . . . . 16.4 Perron’s theorem . . . . . . . . . . . . . . . . . . 16.5 From matrices to graphs . . . . . . . . . . . . . . 16.6 The Perron–Frobenius theorem . . . . . . . . . . 16.7 Irreducible Dirichlet operators . . . . . . . . . . . 16.8 An example . . . . . . . . . . . . . . . . . . . . . 16.9 Problems . . . . . . . . . . . . . . . . . . . . . . 16.10Answers . . . . . . . . . . . . . . . . . . . . . . . 17 The 17.1 17.2 17.3 . . . . . . . . . . 153 153 155 158 159 160 161 162 163 165 166 deficiency zero theorem Reaction networks . . . . . . . . . . . . . . . . . . . . . . . . . . The deficiency zero theorem . . . . . . . . . . . . . . . . . . . . . References and remarks . . . . . . . . . . . . . . . . . . . . . . . 170 170 173 177 18 Example of the deficiency 18.1 Diatomic molecules . . . 18.2 A reaction network . . . 18.3 The rate equation . . . . 18.4 Deficiency zero theorem zero . . . . . . . . . . . . theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Example of the Anderson–Craciun–Kurtz 19.1 The master equation . . . . . . . . . . . . 19.2 Equilibrium solutions . . . . . . . . . . . . 19.3 Noether’s theorem . . . . . . . . . . . . . 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 178 179 181 183 theorem 187 . . . . . . . . . . . . . 187 . . . . . . . . . . . . . 189 . . . . . . . . . . . . . 192 CONTENTS CONTENTS 19.4 The Anderson–Craciun–Kurtz theorem . . . . . . . . . . . . . . . 194 19.5 Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 20 The 20.1 20.2 20.3 20.4 20.5 20.6 deficiency of a reaction network Reaction networks revisited . . . . . The concept of deficiency . . . . . . How to compute the deficiency . . . Examples . . . . . . . . . . . . . . . Different kinds of graphs . . . . . . . Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 201 203 205 206 209 211 21 Rewriting the rate equation 212 21.1 The rate equation and matrix exponentiation . . . . . . . . . . . 213 21.2 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 22 Markov processes 221 22.1 The Markov process of a weighted graph . . . . . . . . . . . . . . 223 22.2 Equilibrium solutions of the master equation . . . . . . . . . . . 224 22.3 The Hamiltonian, revisited . . . . . . . . . . . . . . . . . . . . . 229 23 Proof of the deficiency zero theorem 231 23.1 Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234 23.2 The proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236 24 Noether’s theorem for Dirichlet operators 24.1 The quantum case . . . . . . . . . . . . . . 24.2 The stochastic case . . . . . . . . . . . . . . 24.3 A Noether theorem for Dirichlet operators . 24.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 242 242 244 245 25 Computation and Petri nets 25.1 The reachability problem . . . 25.2 Symmetric monoidal categories 25.3 Computation with Petri nets . 25.4 Software . . . . . . . . . . . . . 25.5 Answers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248 248 253 255 259 260 26 Summary table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 9 1 1 STOCHASTIC PETRI NETS Stochastic Petri nets Stochastic Petri nets are one of many different diagrammatic languages people have evolved to study complex systems. We’ll see how they’re used in chemistry, molecular biology, population biology and queuing theory, which is roughly the science of waiting in line. Here’s an example of a Petri net taken from chemistry: It shows some chemicals and some reactions involving these chemicals. To make it into a stochastic Petri net, we’d just label each reaction by a positive real number: the reaction rate constant, or Petri net for short. Chemists often call different kinds of chemicals ‘species’. In general, a Petri net will have a set of species, which we’ll draw as yellow circles, and a set of transitions, which we’ll draw as blue rectangles. Here’s a Petri net from population biology: Now, instead of different chemicals, the species really are different species of animals! And instead of chemical reactions, the transitions are processes involving these species. This Petri net has two species: rabbit and wolf. It has three transitions: • In birth, one rabbit comes in and two go out. This is a caricature of reality: these bunnies reproduce asexually, splitting in two like amoebas. • In predation, one wolf and one rabbit come in and two wolves go out. This is a caricature of how predators need to eat prey to reproduce. Biologists might use ‘biomass’ to make this sort of idea more precise: a certain amount of mass will go from being rabbit to being wolf. 10 1 STOCHASTIC PETRI NETS • In death, one wolf comes in and nothing goes out. Note that we’re pretending rabbits don’t die unless they’re eaten by wolves. If we labelled each transition with a number called a rate constant, we’d have a ‘stochastic’ Petri net. To make this Petri net more realistic, we’d have to make it more complicated. We’re trying to explain general ideas here, not realistic models of specific situations. Nonetheless, this Petri net already leads to an interesting model of population dynamics: a special case of the so-called ‘Lotka–Volterra predatorprey model’. We’ll see the details soon. More to the point, this Petri net illustrates some possibilities that our previous example neglected. Every transition has some ‘input’ species and some ‘output’ species. But a species can show up more than once as the output (or input) of some transition. And as we see in ‘death’, we can have a transition with no outputs (or inputs) at all. But let’s stop beating around the bush, and give you the formal definitions. They’re simple enough: Definition 1. A Petri net consists of a set S of species and a set T of transitions, together with a function i: S × T → N saying how many copies of each species shows up as input for each transition, and a function o: S × T → N saying how many times it shows up as output. Definition 2. A stochastic Petri net is a Petri net together with a function r : T → (0, ∞) giving a rate constant for each transition. Starting from any stochastic Petri net, we can get two things. First: • The master equation. This says how the probability that we have a given number of things of each species changes with time. Since stochastic means ‘random’, the master equation is what gives stochastic Petri nets their name. The master equation is the main thing we’ll be talking about in future chapters. But not right away! Why not? In chemistry, we typically have a huge number of things of each species. For example, a gram of water contains about 3 × 1022 water molecules, and a smaller but still enormous number of hydroxide ions (OH− ), hydronium ions (H3 O+ ), and other scarier things. These things blunder around randomly, bump into each other, and sometimes react and turn into other things. There’s a 11 1 STOCHASTIC PETRI NETS stochastic Petri net describing all this, as we’ll eventually see. But in this situation, we don’t usually want to know the probability that there are, say, exactly 31, 849, 578, 476, 264 hydronium ions. That would be too much information! We’d be quite happy knowing the expected value of the number of hydronium ions, so we’d be delighted to have a differential equation that says how this changes with time. And luckily, such an equation exists; and it’s much simpler than the master equation. So, in this section we’ll talk about: • The rate equation. This says how the expected number of things of each species changes with time. But first, we hope you get the overall idea. The master equation is stochastic: at each time the number of things of each species is a random variable taking values in N, the set of natural numbers. The rate equation is deterministic: at each time the expected number of things of each species is a non-random variable taking values in [0, ∞), the set of nonnegative real numbers. If the master equation is the true story, the rate equation is only approximately true; but the approximation becomes good in some limit where the expected value of the number of things of each species is large, and the standard deviation is comparatively small. If you’ve studied physics, this should remind you of other things. The master equation should remind you of the quantum harmonic oscillator, where energy levels are discrete, and probabilities are involved. The rate equation should remind you of the classical harmonic oscillator, where energy levels are continuous, and everything is deterministic. When we get to the ‘original research’ part of our story, we’ll see this analogy is fairly precise! We’ll take a bunch of ideas from quantum mechanics and quantum field theory, and tweak them a bit, and show how we can use them to describe the master equation for a stochastic Petri net. Indeed, the random processes that the master equation describes can be drawn as pictures: 12 1 1.1 STOCHASTIC PETRI NETS The rate equation This looks like a Feynman diagram, with animals instead of particles! It’s pretty funny, but the resemblance is no joke: the math will back it up. We’re dying to explain all the details. But just as classical field theory is easier than quantum field theory, the rate equation is simpler than the master equation. So we should start there. 1.1 The rate equation If you handed over a stochastic Petri net, we can write down its rate equation. Instead of telling you the general rule, which sounds rather complicated at first, let’s do an example. Take the Petri net we were just looking at: 13 1 1.1 STOCHASTIC PETRI NETS The rate equation We can make it into a stochastic Petri net by choosing a number for each transition: • the birth rate constant β • the predation rate constant γ • the death rate constant δ Let x(t) be the number of rabbits and let y(t) be the number of wolves at time t. Then the rate equation looks like this: dx = βx − γxy dt dy = γxy − δy dt It’s really a system of equations, but we’ll call the whole thing ‘the rate equation’ because later we may get smart and write it as a single equation. See how it works? • We get a term βx in the equation for rabbits, because rabbits are born at a rate equal to the number of rabbits times the birth rate constant β. • We get a term −δy in the equation for wolves, because wolves die at a rate equal to the number of wolves times the death rate constant δ. • We get a term −γxy in the equation for rabbits, because rabbits die at a rate equal to the number of rabbits times the number of wolves times the predation rate constant γ. • We also get a term γxy in the equation for wolves, because wolves are born at a rate equal to the number of rabbits times the number of wolves times γ. Of course we’re not claiming that this rate equation makes any sense biologically! For example, think about predation. The γxy terms in the above equation would make sense if rabbits and wolves roamed around randomly, and whenever a wolf and a rabbit came within a certain distance, the wolf had a certain probability of eating the rabbit and giving birth to another wolf. At least it 14 1 1.1 STOCHASTIC PETRI NETS The rate equation would make sense in the limit of large numbers of rabbits and wolves, where we can treat x and y as varying continuously rather than discretely. That’s a reasonable approximation to make sometimes. Unfortunately, rabbits and wolves don’t roam around randomly, and a wolf doesn’t spit out a new wolf each time it eats a rabbit. Despite that, the equations dx = βx − γxy dt dy = γxy − δy dt are actually studied in population biology. As we said, they’re a special case of the Lotka–Volterra predator-prey model, which looks like this: dx = βx − γxy dt dy = xy − δy dt The point is that while these models are hideously oversimplified and thus quantitatively inaccurate, they exhibit interesting qualititative behavior that’s fairly robust. Depending on the rate constants, these equations can show either a stable equilibrium or stable periodic behavior. And we go from one regime to another, we see a kind of catastrophe called a ‘Hopf bifurcation’. You can read about this in week308 and week309 of This Week’s Finds. Those consider some other equations, not the Lotka–Volterra equations. But their qualitative behavior is the same! If you want stochastic Petri nets that give quantitatively accurate models, it’s better to retreat to chemistry. Compared to animals, molecules come a lot closer to roaming around randomly and having a chance of reacting when they come within a certain distance. So in chemistry, rate equations can be used to make accurate predictions. But we’re digressing. We should be explaining the general recipe for getting a rate equation from a stochastic Petri net! You might not be able to guess it from just one example. In the next section, we’ll do more examples, and maybe even write down a general formula. But if you’re feeling ambitious, you can try this now: Problem 1. Can you write down a stochastic Petri net whose rate equation is the Lotka–Volterra predator-prey model: dx = βx − γxy dt dy = xy − δy dt for arbitrary β, γ, δ, > 0? If not, for which values of these rate constants can you do it? 15 1 1.2 STOCHASTIC PETRI NETS References 1.2 References Here is free online introduction to stochastic Petri nets and their rate equations: [GP98] Peter J. E. Goss and Jean Peccoud, Quantitative modeling of stochastic systems in molecular biology by using stochastic Petri nets, Proc. Natl. Acad. Sci. USA 95 (1988), 6750–6755. We should admit that Petri net people say place where we’re saying species! The term ‘species’ is used in the literature on chemical reaction networks, which we discuss starting in Section 17. Here are some other introductions to the subject: [Haa02] Peter J. Haas, Stochastic Petri Nets: Modelling, Stability, Simulation, Springer, Berlin, 2002. [Koc10] Ina Koch, Petri nets—a mathematical formalism to analyze chemical reaction networks, Molecular Informatics 29, 838–843, 2010. [Wil06] Darren James Wilkinson, Stochastic Modelling for Systems Biology, Taylor and Francis, New York, 2006. 1.3 Answers Here is the answer to the problem: Problem 1. Can you write down a stochastic Petri net whose rate equation is the Lotka–Volterra predator-prey model: dx = βx − γxy dt dy = xy − δy dt for arbitrary β, γ, δ, > 0? If not, for which values of these rate constants can you do it? Answer. We can find a stochastic Petri net that does the job for any β, γ, δ, > 0. In fact we can find one that does the job for any possible value of β, γ, δ, . But to keep things simple, let’s just solve the original problem. We’ll consider a stochastic Petri net with two species, rabbit and wolf, and four transitions: • birth (1 rabbit in, 2 rabbits out), with rate constant β • death (1 wolf in, 0 wolves out), with rate constant δ • jousting (1 wolf and 1 rabbit in, R rabbits and W wolves out, where R, W are arbitrary natural numbers), with rate constant κ 16 1 1.3 STOCHASTIC PETRI NETS Answers • archery (1 wolf and 1 rabbit in, R0 rabbits and W 0 wolves out, where R0 , W 0 are arbitrary natural numbers) with rate constant κ0 . All these rate constants are positive. This gives the rate equation: dx = βx + (R − 1)κxy + (R0 − 1)κ0 xy dt dy = (W − 1)κxy + (W 0 − 1)κ0 xy − δy dt This is flexible enough to do the job. For example, let’s assume that when they joust, the massive, powerful wolf always kills the rabbit, and then eats the rabbit and has one offspring (R = 0 and W = 2). And let’s assume that in an archery battle, fighting with bow and arrow, the quick, crafty rabbit always kills the wolf, but does not reproduce afterward (R0 = 1, W 0 = 0). Then we get dx = βx − κxy dt dy = (κ − κ0 )xy − δy dt This handles the equations dx = βx − γxy dt 17 1 1.3 STOCHASTIC PETRI NETS Answers dy = xy − δy dt where β, γ, δ, > 0 and γ > . In other words, the cases where more rabbits die due to combat than wolves get born! We’ll let you handle the cases where fewer rabbits die than wolves get born. If we also include a death process for rabbits and birth process for wolves, we can get the fully general Lotka–Volterra equations: dx = βx − γxy dt dy = xy − δy dt It’s worth noting that biologists like to study these equations with different choices of sign for the constants involved: the predator-prey Lotka–Volterra equations and the competitive Lotka–Volterra equations. 18 2 2 THE RATE EQUATION The rate equation As we saw previously in Section 1, a Petri net is a picture that shows different kinds of things and processes that turn bunches of things into other bunches of things, like this: The kinds of things are called species and the processes are called transitions. We see such transitions in chemistry: H + OH → H2 O and population biology: amoeba → amoeba + amoeba and the study of infectious diseases: infected + susceptible → infected + infected and many other situations. A ‘stochastic’ Petri net says the rate at which each transition occurs. We can think of these transitions as occurring randomly at a certain rate—and then we get a stochastic process described by something called the ‘master equation’. But for starters, we’ve been thinking about the limit where there are very many things of each species. Then the randomness washes out, and the expected number of things of each species changes deterministically in a manner described by the ‘rate equation’. It’s time to explain the general recipe for getting this rate equation! It looks complicated at first glance, so we’ll briefly state it, then illustrate it with tons of examples, and then state it again. One nice thing about stochastic Petri nets is that they let you dabble in many sciences. Last time we got a tiny taste of how they show up in population biology. This time we’ll look at chemistry and models of infectious diseases. We won’t dig very deep, but trust us: you can do a lot with stochastic Petri nets in these subjects! We’ll give some references in case you want to learn more. 2.1 Rate equations: the general recipe Here’s the recipe, really quickly: 19 2 2.2 THE RATE EQUATION The formation of water (1) A stochastic Petri net has a set of species and a set of transitions. Let’s concentrate our attention on a particular transition. Then the ith species will appear mi times as the input to that transition, and ni times as the output. Our transition also has a reaction rate 0 < r < ∞. The rate equation answers this question: dxi =??? dt where xi (t) is the number of things of the ith species at time t. The answer is a sum of terms, one for each transition. Each term works the same way. For the transition we’re looking at, it’s mk 1 r(ni − mi )xm 1 · · · xk The factor of (ni −mi ) shows up because our transition destroys mi things of the mk 1 ith species and creates ni of them. The big product over all species, xm 1 · · · xk , shows up because our transition occurs at a rate proportional to the product of the numbers of things it takes as inputs. The constant of proportionality is the reaction rate r. 2.2 The formation of water (1) But let’s do an example. Here’s a naive model for the formation of water from atomic hydrogen and oxygen: This Petri net has just one transition: two hydrogen atoms and an oxygen atom collide simultaneously and form a molecule of water. That’s not really how it goes... but what if it were? Let’s use [H] for the number of hydrogen atoms, and so on, and let the reaction rate be α. Then we get this rate equation: d[H] dt = −2α[H]2 [O] d[O] dt = −α[H]2 [O] d[H2 O] dt = α[H]2 [O] 20 2 2.3 THE RATE EQUATION The formation of water (2) See how it works? The reaction occurs at a rate proportional to the product of the numbers of things that appear as inputs: two H’s and one O. The constant of proportionality is the rate constant α. So, the reaction occurs at a rate equal to α[H]2 [O]. Then: • Since two hydrogen atoms get used up in this reaction, we get a factor of −2 in the first equation. • Since one oxygen atom gets used up, we get a factor of −1 in the second equation. • Since one water molecule is formed, we get a factor of +1 in the third equation. 2.3 The formation of water (2) Let’s do another example. Chemical reactions rarely proceed by having three things collide simultaneously—it’s too unlikely. So, for the formation of water from atomic hydrogen and oxygen, there will typically be an intermediate step. Maybe something like this: Here OH is called a ‘hydroxyl radical’. We’re not sure this is the most likely pathway, but never mind—it’s a good excuse to work out another rate equation. If the first reaction has rate constant α and the second has rate constant β, here’s what we get: d[H] = −α[H][O] − β[H][OH] dt d[OH] dt d[O] dt = α[H][O] − β[H][OH] = −α[H][O] d[H2 O] = β[H][OH] dt See how it works? Each reaction occurs at a rate proportional to the product of the numbers of things that appear as inputs. We get minus signs when a 21 2 2.4 THE RATE EQUATION The dissociation of water (1) reaction destroys one thing of a given kind, and plus signs when it creates one. We don’t get factors of 2 as we did last time, because now no reaction creates or destroys two of anything. 2.4 The dissociation of water (1) In chemistry every reaction comes with a reverse reaction. So, if hydrogen and oxygen atoms can combine to form water, a water molecule can also ‘dissociate’ into hydrogen and oxygen atoms. The rate constants for the reverse reaction can be different than for the original reaction... and all these rate constants depend on the temperature. At room temperature, the rate constant for hydrogen and oxygen to form water is a lot higher than the rate constant for the reverse reaction. That’s why we see a lot of water, and not many lone hydrogen or oxygen atoms. But at sufficiently high temperatures, the rate constants change, and water molecules become more eager to dissociate. Calculating these rate constants is a big subject. We’re just starting to read this book, which looked like the easiest one on the library shelf: [Log96] S. R. Logan, Chemical Reaction Kinetics, Longman, Essex, (1996). But let’s not delve into these mysteries yet. Let’s just take our naive Petri net for the formation of water and turn around all the arrows, to get the reverse reaction: If the reaction rate is α, here’s the rate equation: d[H] dt = 2α[H2 O] d[O] dt = α[H2 O] d[H2 O] dt = −α[H2 O] 22 2 2.5 THE RATE EQUATION The dissociation of water (2) See how it works? The reaction occurs at a rate proportional to [H2 O], since it has just a single water molecule as input. That’s where the α[H2 O] comes from. Then: • Since two hydrogen atoms get formed in this reaction, we get a factor of +2 in the first equation. • Since one oxygen atom gets formed, we get a factor of +1 in the second equation. • Since one water molecule gets used up, we get a factor of +1 in the third equation. 2.5 The dissociation of water (2) Of course, we can also look at the reverse of the more realistic reaction involving a hydroxyl radical as an intermediate. Again, we just turn around the arrows in the Petri net we had: Now the rate equation looks like this: d[H] dt d[OH] dt d[O] dt d[H2 O] dt = +α[OH] + β[H2 O] = −α[OH] + β[H2 O] = +α[OH] = −β[H2 O] Do you see why? Test your understanding of the general recipe. By the way: if you’re a category theorist, when we said “turn around all the arrows” you probably thought “opposite category”. And you’d be right! A Petri net is just a way of presenting of a strict symmetric monoidal category that’s freely generated by some objects (the species) and some morphisms (the 23 2 2.6 THE RATE EQUATION The SI model transitions). When we turn around all the arrows in our Petri net, we’re getting a presentation of the opposite symmetric monoidal category. If you don’t know what this means, don’t worry. We won’t use category theory in this book, even though it’s lurking right beneath the surface. Only at the end, in Section 25.2, will we say a bit more about Petri nets and symmetric monoidal categories. 2.6 The SI model The SI model is an extremely simple model of an infectious disease. We can describe it using this Petri net: There are two species: susceptible and infected. And there’s a transition called infection, where an infected person meets a susceptible person and infects them. Suppose S is the number of susceptible people and I the number of infected ones. If the rate constant for infection is β, the rate equation is dS dt = −βSI dI dt = βSI Do you see why? By the way, it’s easy to solve these equations exactly. The total number of people doesn’t change, so S + I is a conserved quantity. Use this to get rid of one of the variables. You’ll get a version of the famous logistic equation, so the fraction of people infected must grow sort of like this: 24 2 2.7 THE RATE EQUATION The SIR model Problem 2. Is there a stochastic Petri net with just one species whose rate equation is the logistic equation: dP = αP − βP 2 ? dt 2.7 The SIR model The SI model is just a warmup for the more interesting SIR model, which was invented by Kermack and McKendrick in 1927: [KM27] W. O. Kermack and A. G. McKendrick, A contribution to the mathematical theory of epidemics, Proc. Roy. Soc. Lond. A 115, 700-721, (1927). This is the only mathematical model we know to have been knighted : Sir Model. This model has an extra species, called resistant, and an extra transition, called recovery, where an infected person gets better and develops resistance to the disease: If the rate constant for infection is β and the rate constant for recovery is 25 2 2.8 THE RATE EQUATION The SIRS model α, the rate equation for this stochastic Petri net is: dS dt = −βSI dI dt = βSI − αI dR dt = αI See why? We don’t know a ‘closed-form’ solution to these equations. But Kermack and McKendrick found an approximate solution in their original paper. They used this to model the death rate from bubonic plague during an outbreak in Bombay, and got pretty good agreement. Nowadays, of course, we can solve these equations numerically on the computer. 2.8 The SIRS model There’s an even more interesting model of infectious disease called the SIRS model. This has one more transition, called losing resistance, where a resistant person can go back to being susceptible. Here’s the Petri net: Problem 3. If the rate constants for recovery, infection and loss of resistance are α, β, and γ, write down the rate equations for this stochastic Petri net. In the SIRS model we see something new: cyclic behavior! Say you start with a few infected people and a lot of susceptible ones. Then lots of people 26 2 2.9 THE RATE EQUATION The general recipe revisited get infected... then lots get resistant... and then, much later, if you set the rate constants right, they lose their resistance and they’re ready to get sick all over again! You can sort of see it from the Petri net, which looks like a cycle. You can learn about the SI, SIR and SIRS models here: [Man06] Marc Mangel, The Theoretical Biologist’s Toolbox: Quantitative Methods for Ecology and Evolutionary Biology, Cambridge U. Press, Cambridge, (2006). For more models of this type, see: • Compartmental models in epidemiology, Wikipedia. A ‘compartmental model’ is closely related to a stochastic Petri net, but beware: the pictures in this article are not really Petri nets! 2.9 The general recipe revisited Now we’ll remind you of the general recipe and polish it up a bit. So, suppose we have a stochastic Petri net with k species. Let xi be the number of things of the ith species. Then the rate equation looks like: dxi =??? dt It’s really a bunch of equations, one for each 1 ≤ i ≤ k. But what is the right-hand side? The right-hand side is a sum of terms, one for each transition in our Petri net. So, let’s assume our Petri net has just one transition! (If there are more, consider one at a time, and add up the results.) Suppose the ith species appears as input to this transition mi times, and as output ni times. Then the rate equation is dxi mk 1 = r(ni − mi )xm 1 · · · xk dt where r is the rate constant for this transition. That’s really all there is to it! But subscripts make you eyes hurt more and as you get older—this is the real reason for using index-free notation, despite any sophisticated rationales you may have heard—so let’s define a vector x = (x1 , . . . , xk ) that keeps track of how many things there are in each species. Similarly let’s make up an input vector: m = (m1 , . . . , mk ) and an output vector: n = (n1 , . . . , nk ) 27 2 2.10 THE RATE EQUATION Answers for our transition. And a bit more unconventionally, let’s define mk 1 xm = xm 1 · · · xk Then we can write the rate equation for a single transition as dx = r(n − m)xm dt This looks a lot nicer! Indeed, this emboldens us to consider a general stochastic Petri net with lots of transitions, each with their own rate constant. Let’s write T for the set of transitions and r(τ ) for the rate constant of the transition τ ∈ T . Let n(τ ) and m(τ ) be the input and output vectors of the transition τ . Then the rate equation for our stochastic Petri net is X dx = r(τ )(n(τ ) − m(τ ))xm(τ ) dt τ ∈T That’s the fully general recipe in a nutshell. We’re not sure yet how helpful this notation will be, but it’s here whenever we want it. In Section 2 we’ll get to the really interesting part, where ideas from quantum theory enter the game! We’ll see how things of different species randomly transform into each other via the transitions in our Petri net. And someday we’ll check that the expected number of things in each state evolves according to the rate equation we just wrote down... at least in the limit where there are lots of things in each state. 2.10 Answers Here are the answers to the problems: Problem 2. Is there a stochastic Petri net with just one state whose rate equation is the logistic equation: dP = αP − βP 2 ? dt Answer. Yes. Use the Petri net with one species, say amoeba, and two transitions: • fission, with one amoeba as input and two as output, with rate constant α. • competition, with two amoebas as input and one as output, with rate constant β. 28 2 2.10 THE RATE EQUATION Answers The idea of ‘competition’ is that when two amoebas are competing for limited resources, one may die. Problem 3. If the rate constants for recovery, infection and loss of resistance are α, β, and γ, write down the rate equations for this stochastic Petri net: Answer. The rate equation is: dS dt = −βSI + γR dI dt = βSI − αI dR dt = αI − γR 29 3 3 THE MASTER EQUATION The master equation In Section 2 we explained the rate equation of a stochastic Petri net. But now let’s get serious: let’s see what’s stochastic—that is, random— about a stochastic Petri net. For this we need to forget the rate equation (temporarily) and learn about the ‘master equation’. This is where ideas from quantum field theory start showing up! A Petri net has a bunch of species and a bunch of transitions. Here’s an example we’ve already seen, from chemistry: The species are in yellow, the transitions in blue. A labelling of our Petri net is a way of having some number of things of each species. We can draw these things as little black dots, often called tokens: In this example there are only 0 or 1 things of each species: we’ve got one atom of carbon, one molecule of oxygen, one molecule of sodium hydroxide, one molecule of hydrochloric acid, and nothing else. But in general, we can have any natural number of things of each species. In a stochastic Petri net, the transitions occur randomly as time passes. For example, as time passes we might see a sequence of transitions like this: 30 3 3.1 THE MASTER EQUATION The master equation Each time a transition occurs, the number of things of each species changes in an obvious way. 3.1 The master equation Now, we said the transitions occur ‘randomly’, but that dosen’t mean there’s no rhyme or reason to them! The miracle of probability theory is that it lets us state precise laws about random events. The law governing the random behavior of a stochastic Petri net is called the ‘master equation’. In a stochastic Petri net, each transition has a rate constant, a positive real number. Roughly speaking, this determines the probability of that transition. A bit more precisely: suppose we have a Petri net that is labelled in some way at some moment. Then the probability that a given transition occurs in a short time ∆t is approximately: • the rate constant for that transition, times 31 3 3.1 THE MASTER EQUATION The master equation • the time ∆t, times • the number of ways the transition can occur. More precisely still: this formula is correct up to terms of order (∆t)2 . So, taking the limit as ∆t → 0, we get a differential equation describing precisely how the probability of the Petri net having a given labelling changes with time! And this is the master equation. Now, you might be impatient to actually see the master equation, but that would be rash. The true master does’t need to see the master equation. It sounds like a Zen proverb, but it’s true. The raw beginner in mathematics wants to see the solutions of an equation. The more advanced student is content to prove that the solution exists. But the master is content to prove that the equation exists. A bit more seriously: what matters is understanding the rules that inevitably lead to some equation: actually writing it down is then straightforward. And you see, there’s something we haven’t explained yet: “the number of ways the transition can occur”. This involves a bit of counting. Consider, for example, this Petri net: Suppose there are 10 rabbits and 5 wolves. • How many ways can the birth transition occur? Since birth takes one rabbit as input, it can occur in 10 ways. • How many ways can predation occur? Since predation takes one rabbit and one wolf as inputs, it can occur in 10 × 5 = 50 ways. • How many ways can death occur? Since death takes one wolf as input, it can occur in 5 ways. Or consider this one: 32 3 3.1 THE MASTER EQUATION The master equation Suppose there are 10 hydrogen atoms and 5 oxygen atoms. How many ways can they form a water molecule? There are 10 ways to pick the first hydrogen, 9 ways to pick the second hydrogen, and 5 ways to pick the oxygen. So, there are 10 × 9 × 5 = 450 ways. Note that we’re treating the hydrogen atoms as distinguishable, so there are 10 × 9 ways to pick them, not 10×9 = 10 2 2 . In general, the number of ways to choose M distinguishable things from a collection of L is the falling power LM = L · (L − 1) · · · (L − M + 1) where there are M factors in the product, but each is 1 less than the preceding one—hence the term ‘falling’. Okay, now we’ve given you all the raw ingredients to work out the master equation for any stochastic Petri net. The previous paragraph was a big fat hint. One more nudge and you’re on your own: Problem 4. Suppose we have a stochastic Petri net with k species and one transition with rate constant r. Suppose the ith species appears mi times as the input of this transition and ni times as the output. A labelling of this stochastic Petri net is a k-tuple of natural numbers ` = (`1 , . . . , `k ) saying how many things are in each species Let ψ` (t) be the probability that the labelling is ` at time t. Then the master equation looks like this: X d ψ`0 (t) = H`0 ` ψ` (t) dt ` for some matrix of real numbers H`0 ` . What is this matrix? You can write down a formula for this matrix using what we’ve told you. And then, if you have a stochastic Petri net with more transitions, you can just compute the matrix for each transition using this formula, and add them all up. 33 3 3.1 THE MASTER EQUATION The master equation There’s a straightforward way to solve this problem, but we want to get the solution by a strange route: we want to guess the master equation using ideas from quantum field theory! Why? Well, if we think about a stochastic Petri net whose labelling undergoes random transitions as we’ve described, you’ll see that any possible ‘history’ for the labelling can be drawn in a way that looks like a Feynman diagram. In quantum field theory, Feynman diagrams show how things interact and turn into other things. But that’s what stochastic Petri nets do, too! For example, if our Petri net looks like this: then a typical history can be drawn like this: 34 3 3.1 THE MASTER EQUATION The master equation Some rabbits and wolves come in on top. They undergo some transitions as time passes, and go out on the bottom. The vertical coordinate is time, while the horizontal coordinate doesn’t really mean anything: it just makes the diagram easier to draw. If we ignore all the artistry that makes it cute, this Feynman diagram is just a graph with species as edges and transitions as vertices. Each transition occurs at a specific time. We can use these Feynman diagrams to compute the probability that if we start it off with some labelling at time t1 , our stochastic Petri net will wind up with some other labelling at time t2 . To do this, we just take a sum over Feynman diagrams that start and end with the given labellings. For each Feynman diagram, we integrate over all possible times at which the transitions occur. And what do we integrate? Just the product of the rate constants for those transitions! That was a bit of a mouthful, and it doesn’t really matter if you followed it in detail. What matters is that it sounds a lot like stuff you learn when you 35 3 3.1 THE MASTER EQUATION The master equation study quantum field theory! That’s one clue that something cool is going on here. Another is the master equation itself: X d ψ`0 (t) = H`0 ` ψ` (t) dt ` This looks a lot like Schrödinger’s equation, the basic equation describing how a quantum system changes with the passage of time. We can make it look even more like Schrödinger’s equation if we create a vector space with the labellings ` as a basis. The numbers ψ` (t) will be the components of some vector ψ(t) in this vector space. The numbers H`0 ` will be the matrix entries of some operator H on that vector space. And the master equation becomes: d ψ(t) = Hψ(t) dt Compare this with Schrödinger’s equation: i d ψ(t) = Hψ(t) dt The only visible difference is that factor of i! But of course this is linked to another big difference: in the master equation ψ describes probabilities, so it’s a vector in a real vector space. In quantum theory ψ describes amplitudes, so it’s a vector in a complex Hilbert space. Apart from this huge difference, everything is a lot like quantum field theory. In particular, our vector space is a lot like the Fock space one sees in quantum field theory. Suppose we have a quantum particle that can be in k different states. Then its Fock space is the Hilbert space we use to describe an arbitrary collection of such particles. It has an orthonormal basis denoted |`1 · · · `k i where `1 , . . . , `k are natural numbers saying how many particles there are in each state. So, any vector in Fock space looks like this: X ψ= ψ`1 ,...,`k |`1 · · · `k i `1 ,...,`k But if write the whole list `1 , . . . , `k simply as `, this becomes X ψ= ψ` |`i ` This is almost like what we’ve been doing with Petri nets!—except we hadn’t gotten around to giving names to the basis vectors. In quantum field theory class, you would learn lots of interesting operators on Fock space: annihilation and creation operators, number operators, and so on. So, when considering this master equation d ψ(t) = Hψ(t) dt 36 3 3.2 THE MASTER EQUATION Answers it seemed natural to take the operator H and write it in terms of these. There was an obvious first guess, which didn’t quite work... but thinking a bit harder eventually led to the right answer. Later, it turned out people had already thought about similar things. So, we want to explain this. When we first started working on this stuff, we focused on the difference between collections of indistinguishable things, like bosons or fermions, and collections of distinguishable things, like rabbits or wolves. But with the benefit of hindsight, it’s even more important to think about the difference between quantum theory, which is all about probability amplitudes, and the game we’re playing now, which is all about probabilities. So, in the next Section, we’ll explain how we need to modify quantum theory so that it’s about probabilities. This will make it easier to guess a nice formula for H. 3.2 Answers Here is the answer to the problem: Problem 4. Suppose we have a stochastic Petri net with k species and just one transition, whose rate constant is r. Suppose the ith species appears mi times as the input of this transition and ni times as the output. A labelling of this stochastic Petri net is a k-tuple of natural numbers ` = (`1 , . . . , `k ) saying how many things there are of each species. Let ψ` (t) be the probability that the labelling is ` at time t. Then the master equation looks like this: X d ψ`0 (t) = H`0 ` ψ` (t) dt ` for some matrix of real numbers H`0 ` . What is this matrix? Answer. To compute H`0 ` it’s enough to start the Petri net in a definite labelling ` and see how fast the probability of being in some labelling `0 changes. In other words, if at some time t we have ψ` (t) = 1 then d ψ`0 (t) = H`0 ` dt at this time. Now, suppose we have a Petri net that is labelled in some way at some moment. Then the probability that the transition occurs in a short time ∆t is approximately: • the rate constant r, times • the time ∆t, times • the number of ways the transition can occur, which is the product of falling mk m1 powers `1 · · · `k . Let’s call this product `m for short. 37 3 3.2 THE MASTER EQUATION Answers Multiplying these 3 things we get r`m ∆t So, the rate at which the transition occurs is just: r`m And when the transition occurs, it eats up mi things of the i th species, and produces ni things of that species. So, it carries our system from the original labelling ` to the new labelling `0 = ` + n − m So, in this case we have d ψ`0 (t) = r`m dt and thus H`0 ` = r`m However, that’s not all: there’s another case to consider! Since the probability of the Petri net being in this new labelling `0 is going up, the probability of it staying in the original labelling ` must be going down by the same amount. So we must also have H`` = −r`m We can combine both cases into one formula like this: H`0 ` = r`m (δ`0 ,`+n−m − δ`0 ,` ) Here the first term tells us how fast the probability of being in the new labelling is going up. The second term tells us how fast the probability of staying in the original labelling is going down. Note: each column in the matrix H`0 ` sums to zero, and all the off-diagonal entries are nonnegative. That’s good: in the next section we’ll show that this matrix must be ‘infinitesimal stochastic’, meaning precisely that it has these properties! 38 4 4 PROBABILITIES VS AMPLITUDES Probabilities vs amplitudes In Section 3 we saw clues that stochastic Petri nets are a lot like quantum field theory, but with probabilities replacing amplitudes. There’s a powerful analogy at work here, which can help us a lot. It’s time to make that analogy precise. But first, let us quickly sketch why it could be worthwhile. 4.1 A Poisson process Consider this stochastic Petri net with rate constant r: It describes an inexhaustible supply of fish swimming down a river, and getting caught when they run into a fisherman’s net. In any short time ∆t there’s a chance of about r∆t of a fish getting caught. There’s also a chance of two or more fish getting caught, but this becomes negligible by comparison as ∆t → 0. Moreover, the chance of a fish getting caught during this interval of time is independent of what happens before or afterwards. This sort of process is called a Poisson process. Problem 5. Suppose we start out knowing for sure there are no fish in the fisherman’s net. What’s the probability that he has caught n fish at time t? Answer. At any time there will be some probability of having caught n fish; let’s call this probability ψ(n, t). We can summarize all these probabilities in a single power series, called a generating function: Ψ(t) = ∞ X ψ(n, t) z n n=0 Here z is a formal variable—don’t ask what it means, for now it’s just a trick. In quantum theory we use this trick when talking about collections of photons rather than fish, but then the numbers ψ(n, t) are complex ‘amplitudes’. Now they are real probabilities, but we can still copy what the physicists do, and use this trick to rewrite the master equation as follows: d Ψ(t) = HΨ(t) dt This describes how the probability of having caught any given number of fish changes with time. 39 4 4.1 PROBABILITIES VS AMPLITUDES A Poisson process What’s the operator H? Well, in quantum theory we describe the creation of photons using a certain operator on power series called the creation operator: a† Ψ = zΨ We can try to apply this to our fish. If at some time we’re 100% sure we have n fish, we have Ψ = zn so applying the creation operator gives a† Ψ = z n+1 One more fish! That’s good. So, an obvious wild guess is H = ra† where r is the rate at which we’re catching fish. Let’s see how well this guess works. If you know how to exponentiate operators, you know to solve this equation: d Ψ(t) = HΨ(t) dt It’s easy: Ψ(t) = exp(tH)Ψ(0) Since we start out knowing there are no fish in the net, we have Ψ(0) = 1 so with our guess for H we get Ψ(t) = exp(rta† )1 But a† is the operator of multiplication by z, so exp(rta† ) is multiplication by ertz , and ∞ X (rt)n n Ψ(t) = ertz = z n! n=0 So, if our guess is right, the probability of having caught n fish at time t is (rt)n n! Unfortunately, this can’t be right, because these probabilities don’t sum to 1! Instead their sum is ∞ X (rt)n = ert n! n=0 We can try to wriggle out of the mess we’re in by dividing our answer by this fudge factor. It sounds like a desperate measure, but we’ve got to try something! 40 4 4.2 PROBABILITIES VS AMPLITUDES Probability theory vs quantum theory This amounts to guessing that the probability of having caught n fish by time t is (rt)n −rt e n! And this guess is right! This is called the Poisson distribution: it’s famous for being precisely the answer to the problem we’re facing. So on the one hand our wild guess about H was wrong, but on the other hand it was not so far off. We can fix it as follows: H = r(a† − 1) The extra −1 gives us the fudge factor we need. So, a wild guess corrected by an ad hoc procedure seems to have worked! But what’s really going on? What’s really going on is that a† , or any multiple of this, is not a legitimate Hamiltonian for a master equation: if we define a time evolution operator exp(tH) using a Hamiltonian like this, probabilities won’t sum to 1! But a† − 1 is okay. So, we need to think about which Hamiltonians are okay. In quantum theory, self-adjoint Hamiltonians are okay. But in probability theory, we need some other kind of Hamiltonian. Let’s figure it out. 4.2 Probability theory vs quantum theory Suppose we have a system of any kind: physical, chemical, biological, economic, whatever. The system can be in different states. In the simplest sort of model, we say there’s some set X of states, and say that at any moment in time the system is definitely in one of these states. But we want to compare two other options: • In a probabilistic model, we may instead say that the system has a probability ψ(x) of being in any state x ∈ X. These probabilities are nonnegative real numbers with X ψ(x) = 1 x∈X • In a quantum model, we may instead say that the system has an amplitude ψ(x) of being in any state x ∈ X. These amplitudes are complex numbers with X |ψ(x)|2 = 1 x∈X Probabilities and amplitudes are similar yet strangely different. Of course given an amplitude we can get a probability by taking its absolute value and squaring it. This is a vital bridge from quantum theory to probability theory. In the present section, however, we don’t want to focus on the bridges, but rather the parallels between these theories. 41 4 4.2 PROBABILITIES VS AMPLITUDES Probability theory vs quantum theory We often want to replace the sums above by integrals. For that we need to replace our set X by a measure space, which is a set equipped with enough structure that you can integrate real or complex functions defined on it. Well, at least you can integrate so-called ‘integrable’ functions—but we’ll neglect all issues of analytical rigor here. Then: • In a probabilistic model, the system has a probability distribution ψ : X → R, which obeys ψ ≥ 0 and Z ψ(x) dx = 1 X • In a quantum model, the system has a wavefunction ψ : X → C, which obeys Z |ψ(x)|2 dx = 1 X In probability theory, we integrate ψ over a set S ⊂ X to find out the probability that our systems state is in this set. In quantum theory we integrate |ψ|2 over the set to answer the same question. We don’t need to think about sums over sets and integrals over measure spaces separately: there’s a way to make any set X into a measure space such that by definition, Z X ψ(x) dx = ψ(x) X x∈X In short, integrals are more general than sums! So, we’ll mainly talk about integrals, until the very end. In probability theory, we want our probability distributions to be vectors in some vector space. Ditto for wave functions in quantum theory! So, we make up some vector spaces: • In probability theory, the probability distribution ψ is a vector in the space Z L1 (X) = {ψ : X → C : |ψ(x)| dx < ∞} X • In quantum theory, the wavefunction ψ is a vector in the space Z L2 (X) = {ψ : X → C : |ψ(x)|2 dx < ∞} X You may wonder why we defined L1 (X) to consist of complex functions when probability distributions are real. We’re just struggling to make the analogy seem as strong as possible. In fact probability distributions are not just real but nonnegative. We need to say this somewhere... but we can, if we like, start by saying they’re complex-valued functions, but then whisper that they must in fact be nonnegative (and thus real). It’s not the most elegant solution, but that’s what we’ll do for now. Now: 42 4 4.3 PROBABILITIES VS AMPLITUDES Stochastic vs unitary operators • The main thing we can do with elements of L1 (X), besides what we can do with vectors in any vector space, is integrate one. This gives a linear map: Z : L1 (X) → C • The main thing we can with elements of L2 (X), besides the things we can do with vectors in any vector space, is take the inner product of two: Z hψ, φi = ψ(x)φ(x) dx X This gives a map that’s linear in one slot and conjugate-linear in the other: h−, −i : L2 (X) × L2 (X) → C First came probability theory with L1 (X); then came quantum theory with L (X). Naive extrapolation would say it’s about time for someone to invent an even more bizarre theory of reality based on L3 (X). In this, you’d have to integrate the product of three wavefunctions to get a number! The math of Lp spaces is already well-developed, so give it a try if you want. We’ll stick to L1 and L2 . 2 4.3 Stochastic versus unitary operators Now let’s think about time evolution: • In probability theory, the passage of time is described by a map sending probability distributions to probability distributions. This is described using a stochastic operator U : L1 (X) → L1 (X) meaning a linear operator such that Z Z Uψ = ψ and ψ≥0 ⇒ Uψ ≥ 0 • In quantum theory the passage of time is described by a map sending wavefunction to wavefunctions. This is described using an isometry U : L2 (X) → L2 (X) meaning a linear operator such that hU ψ, U φi = hψ, φi 43 4 4.4 PROBABILITIES VS AMPLITUDES Infinitesimal stochastic versus self-adjoint operators In quantum theory we usually want time evolution to be reversible, so we focus on isometries that have inverses: these are called unitary operators. In probability theory we often consider stochastic operators that are not invertible. 4.4 Infinitesimal stochastic versus self-adjoint operators Sometimes it’s nice to think of time coming in discrete steps. But in theories where we treat time as continuous, to describe time evolution we usually need to solve a differential equation. This is true in both probability theory and quantum theory. In probability theory we often describe time evolution using a differential equation called the master equation: d ψ(t) = Hψ(t) dt whose solution is ψ(t) = exp(tH)ψ(0) In quantum theory we often describe time evolution using a differential equation called Schrödinger’s equation: i d ψ(t) = Hψ(t) dt whose solution is ψ(t) = exp(−itH)ψ(0) In both cases, we call the operator H the Hamiltonian. In fact the appearance of i in the quantum case is purely conventional; we could drop it to make the analogy better, but then we’d have to work with ‘skew-adjoint’ operators instead of self-adjoint ones in what follows. Let’s guess what properties an operator H should have to make exp(−itH) unitary for all t. We start by assuming it’s an isometry: hexp(−itH)ψ, exp(−itH)φi = hψ, φi Then we differentiate this with respect to t and set t = 0, getting h−iHψ, φi + hψ, −iHφi = 0 or in other words: hHψ, φi = hψ, Hφi Physicists call an operator obeying this condition self-adjoint. Mathematicians know there’s more to it, but now is not the time to discuss such subtleties, intriguing though they be. All that matters now is that there is, indeed, a correspondence between self-adjoint operators and well-behaved ‘1-parameter unitary groups’ exp(−itH). This is called Stone’s Theorem. 44 4 4.4 PROBABILITIES VS AMPLITUDES Infinitesimal stochastic versus self-adjoint operators But now let’s copy this argument to guess what properties an operator H must have to make exp(tH) stochastic. We start by assuming exp(tH) is stochastic, so Z Z exp(tH)ψ = ψ and ψ≥0 ⇒ exp(tH)ψ ≥ 0 We can differentiate the first equation with respect to t and set t = 0, getting Z Hψ = 0 for all ψ. But what about the second condition, ψ≥0 ⇒ exp(tH)ψ ≥ 0? It seems easier to deal with this in the special case when integrals over X reduce to sums. So let’s suppose that happens... and let’s start by seeing what the first condition says in this case. In this case, L1 (X) has a basis of ‘Kronecker delta functions’: The Kronecker delta function δi vanishes everywhere except at one point i ∈ X, where it equals 1. Using this basis, we can write any operator on L1 (X) as a matrix. As a warmup, let’s see what it means for an operator U : L1 (X) → L1 (X) to be stochastic in this case. We’ll take the conditions Z Z Uψ = ψ and ψ≥0 ⇒ Uψ ≥ 0 and rewrite them using matrices. For both, it’s enough to consider the case where ψ is a Kronecker delta, say δj . In these terms, the first condition says X Uij = 1 i∈X for each column j. The second says Uij ≥ 0 for all i, j. So in this case, a stochastic operator is just a square matrix where each column sums to 1 and all the entries are nonnegative. (Such matrices are often called left stochastic.) 45 4 4.4 PROBABILITIES VS AMPLITUDES Infinitesimal stochastic versus self-adjoint operators Next, let’s see what we need for an operator H to have the property that exp(tH)is stochastic for all t ≥ 0. It’s enough to assume t is very small, which lets us use the approximation exp(tH) = 1 + tH + · · · and work to first order in t. Saying that each column of this matrix sums to 1 then amounts to X δij + tHij + · · · = 1 i∈X which requires X Hij = 0 i∈X Saying that each entry is nonnegative amounts to δij + tHij + · · · ≥ 0 When i = j this will be automatic when t is small enough, so the meat of this condition is Hij ≥ 0 if i 6= j So, let’s say H is infinitesimal stochastic if its columns sum to zero and its off-diagonal entries are nonnegative. The idea is that any infinitesimal stochastic operator should be the infinitesimal generator of a stochastic process. The adjective ‘infinitesimal stochastic’ doesn’t roll off the tongue, but we don’t know one we like better. Some people call an infinitesimal stochastic operator an intensity matrix, while others call it a stochastic Hamiltonian, but we really want an adjective. Anyway, when we get the details straightened out, any 1-parameter family of stochastic operators U (t) : L1 (X) → L1 (X) t≥0 obeying U (0) = I U (t)U (s) = U (t + s) and continuity: ti → t ⇒ U (ti )ψ → U (t)ψ should be of the form U (t) = exp(tH) for a unique ‘infinitesimal stochastic operator’ H. When X is a finite set, this is true—and an infinitesimal stochastic operator is just a square matrix whose columns sum to zero and whose off-diagonal entries are nonnegative. But we don’t know a really good theorem characterizing 46 4 4.5 PROBABILITIES VS AMPLITUDES The moral infinite stochastic operators for general measure spaces X. We shall say what we know about this in Section 11.6. Luckily, for our work on stochastic Petri nets we only need to understand the case where X is a countable set and our integrals are really just sums. This should be almost like the case where X is a finite set—but we’ll need to take care that all our sums converge. 4.5 The moral Now we can see why a Hamiltonian like a† is no good, while a† − 1 is good. (We’ll ignore the rate constant r since it’s irrelevant here.) The first one is not infinitesimal stochastic, while the second one is! In this example, our set of states is the natural numbers: X=N The probability distribution ψ: N → C tells us the probability of having caught any specific number of fish. The creation operator is not infinitesimal stochastic: in fact, it’s stochastic! Why? Well, when we apply the creation operator, what was the probability of having n fish now becomes the probability of having n + 1 fish. So, the probabilities remain nonnegative, and their sum over all n is unchanged. Those two conditions are all we need for a stochastic operator. Using our fancy abstract notation, these conditions say: Z Z † a ψ= ψ and ψ ≥ 0 ⇒ a† ψ ≥ 0 So, precisely by virtue of being stochastic, the creation operator fails to be infinitesimal stochastic: Z a† ψ 6= 0 Thus it’s a bad Hamiltonian for our stochastic Petri net. On the other hand, a† − 1 is infinitesimal stochastic. Its off-diagonal entries are the same as those of a† , so they’re nonnegative. Moreover: Z (a† − 1)ψ = 0 precisely because Z a† ψ = Z ψ You may be thinking: all this fancy math just to understand a single stochastic Petri net, the simplest one of all! 47 4 4.5 PROBABILITIES VS AMPLITUDES The moral But next we’ll explain a general recipe which will let you write down the Hamiltonian for any stochastic Petri net. The lessons we’ve just learned will make this much easier. And pondering the analogy between probability theory and quantum theory will also be good for our bigger project of unifying the applications of network diagrams to dozens of different subjects. 48 5 5 ANNIHILATION AND CREATION OPERATORS Annihilation and creation operators Now comes the fun part. Let’s see how tricks from quantum theory can be used to describe random processes. We’ll try to make this section completely self-contained, except at the very end. So, even if you skipped a bunch of the previous ones, this should make sense. You’ll need to know a bit of math: calculus, a tiny bit probability theory, and linear operators on vector spaces. You don’t need to know quantum theory, though you’ll have more fun if you do. What we’re doing here is very similar, but also strangely different—for reasons explained in Section 4. 5.1 Rabbits and quantum mechanics Suppose we have a population of rabbits in a cage and we’d like to describe its growth in a stochastic way, using probability theory. Let ψn be the probability of having n rabbits. We can borrow a trick from quantum theory, and summarize all these probabilities in a formal power series like this: Ψ= ∞ X ψn z n n=0 The variable z doesn’t mean anything in particular, and we don’t care if the power series converges. See, in math ‘formal’ means “it’s only symbols on the page, just follow the rules”. It’s like if someone says a party is ‘formal’, so need to wear a white tie: you’re not supposed to ask what the tie means. However, there’s a good reason for this trick. We can define two operators on formal power series, called the annihilation operator: aΨ = d Ψ dz and the creation operator: a† Ψ = zΨ They’re just differentiation and multiplication by z, respectively. So, for example, suppose we start out being 100% sure we have n rabbits for some particular number n. Then ψn = 1, while all the other probabilities are 0, so: Ψ = zn If we then apply the creation operator, we obtain a† Ψ = z n+1 Voilà! One more rabbit! 49 5 5.1 ANNIHILATION AND CREATION OPERATORS Rabbits and quantum mechanics The annihilation operator is more subtle. If we start out with n rabbits: Ψ = zn and then apply the annihilation operator, we obtain aΨ = nz n−1 What does this mean? The z n−1 means we have one fewer rabbit than before. But what about the factor of n? It means there were n different ways we could pick a rabbit and make it disappear! This should seem a bit mysterious, for various reasons... but we’ll see how it works soon enough. The creation and annihilation operators don’t commute: (aa† − a† a)Ψ = d d (zΨ) − z Ψ = Ψ dz dz so for short we say: aa† − a† a = 1 or even shorter: [a, a† ] = 1 where the commutator of two operators is [S, T ] = ST − T S. The noncommutativity of operators is often claimed to be a special feature of quantum physics, and the creation and annihilation operators are fundamental to understanding the quantum harmonic oscillator. There, instead of rabbits, we’re studying quanta of energy, which are peculiarly abstract entities obeying rather counterintuitive laws. So, it’s cool that the same math applies to purely classical entities, like rabbits! In particular, the equation [a, a† ] = 1 just says that there’s one more way to put a rabbit in a cage of rabbits, and then take one out, than to take one out and then put one in. 50 5 5.2 ANNIHILATION AND CREATION OPERATORS Catching rabbits But how do we actually use this setup? We want to describe how the probabilities ψn change with time, so we write Ψ(t) = ∞ X ψn (t)z n n=0 Then, we write down an equation describing the rate of change of Ψ: d Ψ(t) = HΨ(t) dt Here H is an operator called the Hamiltonian, and the equation is called the master equation. The details of the Hamiltonian depend on our problem! But we can often write it down using creation and annihilation operators. Let’s do some examples, and then we’ll tell you the general rule. 5.2 Catching rabbits In Section 4 we told you what happens when we stand in a river and catch fish as they randomly swim past. Let us remind you of how that works. But now let’s use rabbits. So, suppose an inexhaustible supply of rabbits are randomly roaming around a huge field, and each time a rabbit enters a certain area, we catch it and add it to our population of caged rabbits. Suppose that on average we catch one rabbit per unit time. Suppose the chance of catching a rabbit during any interval of time is independent of what happens before or afterwards. What is the Hamiltonian describing the probability distribution of caged rabbits, as a function of time? There’s an obvious dumb guess: the creation operator! However, we saw last time that this doesn’t work, and we saw how to fix it. The right answer is H = a† − 1 To see why, suppose for example that at some time t we have n rabbits, so: Ψ(t) = z n Then the master equation says that at this moment, d Ψ(t) = (a† − 1)Ψ(t) = z n+1 − z n dt 51 5 5.3 ANNIHILATION AND CREATION OPERATORS Dying rabbits P∞ Since Ψ = n=0 ψn (t)z n , this implies that the coefficients of our formal power series are changing like this: d ψn+1 (t) = 1 dt d ψn (t) = −1 dt while all the rest have zero derivative at this moment. And that’s exactly right! See, ψn+1 (t) is the probability of having one more rabbit, and this is going up at rate 1. Meanwhile, ψn (t) is the probability of having n rabbits, and this is going down at the same rate. Problem 6. Show that with this Hamiltonian and any initial conditions, the master equation predicts that the expected number of rabbits grows linearly. 5.3 Dying rabbits Don’t worry: no rabbits are actually injured in the research that we’re doing here at the Centre for Quantum Technologies. This is just a thought experiment. Suppose a mean nasty guy had a population of rabbits in a cage and didn’t feed them at all. Suppose that each rabbit has a unit probability of dying per unit time. And as always, suppose the probability of this happening in any interval of time is independent of what happens before or after that time. What is the Hamiltonian? Again there’s a dumb guess: the annihilation operator! And again this guess is wrong, but it’s not far off. As before, the right answer includes a ‘correction term’: H =a−N This time the correction term is famous in its own right. It’s called the number operator: N = a† a The reason is that if we start with n rabbits, and apply this operator, it amounts to multiplication by n: d N z n = z z n = nz n dz Let’s see why this guess is right. Again, suppose that at some particular time t we have n rabbits, so Ψ(t) = z n 52 5 5.4 ANNIHILATION AND CREATION OPERATORS Breeding rabbits Then the master equation says that at this time d Ψ(t) = (a − N )Ψ(t) = nz n−1 − nz n dt So, our probabilities are changing like this: d ψn−1 (t) = n dt d ψn (t) = −n dt while the rest have zero derivative. And this is good! We’re starting with n rabbits, and each has a unit probability per unit time of dying. So, the chance of having one less should be going up at rate n. And the chance of having the same number we started with should be going down at the same rate. Problem 7. Show that with this Hamiltonian and any initial conditions, the master equation predicts that the expected number of rabbits decays exponentially. 5.4 Breeding rabbits Suppose we have a strange breed of rabbits that reproduce asexually. Suppose that each rabbit has a unit probability per unit time of having a baby rabbit, thus effectively duplicating itself. As you can see from the cryptic picture above, this ‘duplication’ process takes one rabbit as input and has two rabbits as output. So, if you’ve been paying attention, you should be ready with a dumb guess for the Hamiltonian: a† a† a. This operator annihilates one rabbit and then creates two! But you should also suspect that this dumb guess will need a ‘correction term’. And you’re right! As always, the correction terms makes the probability of things staying the same go down at exactly the rate that the probability of things changing goes up. You should guess the correction term... but we’ll just tell you: H = a† a† a − N 53 5 5.5 ANNIHILATION AND CREATION OPERATORS Dueling rabbits We can check this in the usual way, by seeing what it does when we have n rabbits: d Hz n = z 2 z n − nz n = nz n+1 − nz n dz That’s good: since there are n rabbits, the rate of rabbit duplication is n. This is the rate at which the probability of having one more rabbit goes up... and also the rate at which the probability of having n rabbits goes down. Problem 8. Show that with this Hamiltonian and any initial conditions, the master equation predicts that the expected number of rabbits grows exponentially. 5.5 Dueling rabbits Let’s do some stranger examples, just so you can see the general pattern. Here each pair of rabbits has a unit probability per unit time of fighting a duel with only one survivor. You might guess the Hamiltonian a† aa, but in fact: H = a† aa − N (N − 1) Let’s see why this is right! Let’s see what it does when we have n rabbits: Hz n = z d2 n z − n(n − 1)z n = n(n − 1)z n−1 − n(n − 1)z n dz 2 That’s good: since there are n(n − 1) ordered pairs of rabbits, the rate at which duels take place is n(n − 1). This is the rate at which the probability of having one less rabbit goes up... and also the rate at which the probability of having n rabbits goes down. (If you prefer unordered pairs of rabbits, just divide the Hamiltonian by 2. We should talk about this more, but not now.) 54 5 5.6 5.6 ANNIHILATION AND CREATION OPERATORS Brawling rabbits Brawling rabbits Now each triple of rabbits has a unit probability per unit time of getting into a fight with only one survivor! We don’t know the technical term for a three-way fight, but perhaps it counts as a small ‘brawl’ or ‘melee’. Now the Hamiltonian is: H = a† a3 − N (N − 1)(N − 2) You can check that: Hz n = n(n − 1)(n − 2)z n−2 − n(n − 1)(n − 2)z n and this is good, because n(n − 1)(n − 2) is the number of ordered triples of rabbits. You can see how this number shows up from the math, too: a3 z n = 5.7 d3 n z = n(n − 1)(n − 2)z n−3 dz 3 The general rule Suppose we have a process taking k rabbits as input and having j rabbits as output: 55 5 5.8 ANNIHILATION AND CREATION OPERATORS Kissing rabbits By now you can probably guess the Hamiltonian we’ll use for this: j H = a† ak − N (N − 1) · · · (N − k + 1) This works because dk n z = n(n − 1) · · · (n − k + 1)z n−k dz k so that if we apply our Hamiltonian to n rabbits, we get ak z n = Hz n = n(n − 1) · · · (n − k + 1)(z n+j−k − z n ) See? As the probability of having n + j − k rabbits goes up, the probability of having n rabbits goes down, at an equal rate. This sort of balance is necessary for H to be a sensible Hamiltonian in this sort of stochastic theory ‘infinitesimal stochastic operator’, to be precise). And the rate is exactly the number of ordered k-tuples taken from a collection of n rabbits. This is called the kth falling power of n, and written as follows: nk = n(n − 1) · · · (n − k + 1) Since we can apply functions to operators as well as numbers, we can write our Hamiltonian as: j H = a† ak − N k 5.8 Kissing rabbits 56 5 5.8 ANNIHILATION AND CREATION OPERATORS Kissing rabbits Let’s do one more example just to test our understanding. This time each pair of rabbits has a unit probability per unit time of bumping into each other, exchanging a friendly kiss and walking off. This shouldn’t affect the rabbit population at all! But let’s follow the rules and see what they say. According to our rules, the Hamiltonian should be: 2 H = a† a2 − N (N − 1) However, d2 n z = n(n − 1)z n = N (N − 1)z n dz 2 and since z n form a ‘basis’ for the formal power series, we see that: 2 a† a2 z n = z 2 2 a† a2 = N (N − 1) so in fact: H=0 That’s good: if the Hamiltonian is zero, the master equation will say d Ψ(t) = 0 dt so the population, or more precisely the probability of having any given number of rabbits, will be constant. There’s another nice little lesson here. Copying the calculation we just did, it’s easy to see that: k a† ak = N k This is a cute formula for falling powers of the number operator in terms of annihilation and creation operators. It means that for the general transition we saw before: we can write the Hamiltonian in two equivalent ways: j j k H = a† ak − N k = a† ak − a† ak 57 5 5.9 ANNIHILATION AND CREATION OPERATORS References Okay, that’s it for now! We can, and will, generalize all this stuff to stochastic Petri nets where there are things of many different kinds—not just rabbits. And we’ll see that the master equation we get matches the answer to the problem in Section 3. That’s pretty easy. 5.9 References For a general introduction to stochastic processes, try this: [Van07] Nico Van Kampen, Stochastic Processes in Physics and Chemistry, NorthHolland, New York, 2007. There has been a lot of work on using annihilation and creation operators for stochastic systems. Here are some papers on the foundations: [Doi76a] M. Doi, Second-quantization representation for classical many-particle systems, Jour. Phys. A 9 1465–1477, 1976. [Doi76b] M. Doi, Stochastic theory of diffusion-controlled reactions. Jour. Phys. A 9 1479–1495, 1976. [Pel85] L. Peliti, Path integral approach to birth-death processes on a lattice, Jour. Physique 46 1469-83, 1985. [Car] J. Cardy, Renormalization group approach to reaction-diffusion problems. Available as arXiv:cond-mat/9607163. [MG98] D. C. Mattis and M. L. Glasser, The uses of quantum field theory in diffusion-limited reactions, Rev. Mod. Phys. 70, 979–1001, 1998. Here are some papers on applications: [THVL78] Uwe Täuber, Martin Howard and Benjamin P. Vollmayr-Lee, Applications of field-theoretic renormalization group methods to reaction-diffusion problems, Jour. Phys. A 38 (2005), R79. Also available as arXiv:condmat/0501678. [BC07] M. A. Buice, and J. D. Cowan, Field-theoretic approach to fluctuation effects in neural networks, Phys. Rev. E 75 (2007), 051919. [BC09] M. A. Buice and J. D. Cowan, Statistical mechanics of the neocortex, Prog. Biophysics & Mol. Bio. 99, 53–86, 2009. [DF09] Peter J. Dodd and Neil M. Ferguson, A many-body field theory approach to stochastic models in population biology, PLoS ONE 4, e6855, 2009. 58 5 5.10 5.10 ANNIHILATION AND CREATION OPERATORS Answers Answers Here are the answers to the problems: Problem 6. Show that with the Hamiltonian H = a† − 1 and any initial conditions, the master equation predicts that the expected number of rabbits grows linearly. Answer. Here is one answer, thanks to David Corfield on Azimuth. If at some time t we have n rabbits, so that Ψ(t) = z n , we have seen that the probability of having any number of rabbits changes as follows: d d d ψn+1 (t) = 1, ψn (t) = −1, ψm (t) = 0 otherwise dt dt dt Thus the rate of increase of the expected number of rabbits is (n + 1) − n = 1. But any probability distribution is a linear combination of these basis vector z n , so the rate of increase of the expected number of rabbits is always X ψn (t) = 1 n so the expected number grows linearly. Here is a second solution, using some more machinery. This machinery is overkill here, but it will be useful for solving the next two problems and also many other problems. R Answer. In the general formalism described in Section 4, we used ψ to mean the integral of a function over some measure space, so that probability distributions are the functions obeying Z ψ=1 and ψ≥0 In the examples we’ve been looking at, this integral is really a sum over n = 0, 1, 2, . . . , and it might be confusing to use integral notation since we’re using derivatives for a completely different purpose. So let us define a sum notation as follows: ∞ X X Ψ= ψn n=0 This may be annoying, since after all we really have ∞ X Ψ|z=1 = ψn n=0 but please humor us. To work with this sum notation, two rules are very handy. 59 5 5.10 ANNIHILATION AND CREATION OPERATORS Answers Rule 1: For any formal power series Φ, X a† Φ = X Φ We mentioned this rule in Section 4: it’s part of the creation operator being a stochastic operator. It’s easy to check: X X a† Φ = zΦ|z=1 = Φ|z=1 = Φ Rule 2: For any formal power series Φ, X aΦ = X NΦ Again this is easy to check: X X X NΦ = a† aΦ = aΦ These rules can be used together with the commutation relation [a, a† ] = 1 and its consequences [a, N ] = a, [a† , N ] = −a† to do many interesting things. Let’s see how! Suppose we have some observable O that we can write as an operator on formal power series: for example, the number operator, or any power of that. The expected value of this observable in the probability distribution Ψ is X OΨ So, if we’re trying to work out the time derivative of the expected value of O, we can start by using the master equation: X d X d X OΨ(t) = O Ψ(t) = OHΨ(t) dt dt Then we can write O and H using annihilation and creation operators and use our rules. For example, in the problem at hand, we have H = a† − 1 and the observable we’re interested in is the number of rabbits O=N 60 5 5.10 ANNIHILATION AND CREATION OPERATORS Answers so we want to compute X OHΨ(t) = X N (a† − 1)Ψ(t) There are many ways to use our rules to evaluate this. For example, Rule 1 implies that X X N (a† − 1)Ψ(t) = a(a† − 1)Ψ(t) but the commutation relations say aa† = a† a + 1 = N + 1, so X X a(a† − 1)Ψ(t) = (N + 1 − a)Ψ(t) and using Rule 1 again we see this equals X Ψ(t) = 1 Thus we have d X N Ψ(t) = 1 dt It follows that the expected number of rabbits grows linearly: X N Ψ(t) = t + c Problem 7. Show that with the Hamiltonian H =a−N and any initial conditions, the master equation predicts that the expected number of rabbits decays exponentially. Answer. We use the machinery developed in our answer to Problem 6. We want to compute the time derivative of the expected number of rabbits: X X d X NΨ = N HΨ = N (a − N )Ψ dt The commutation relation [a, N ] = a implies that N a = aN − a. So: X X N (a − N )Ψ = (aN − N − N 2 )Ψ but now Rule 2 says: X X X (aN − N − N 2 )Ψ = (N 2 − N − N 2 )Ψ = − NΨ so we see X d X NΨ = − NΨ dt It follows that the expected number of rabbits decreases exponentially: X N Ψ(t) = ce−t 61 5 5.10 ANNIHILATION AND CREATION OPERATORS Answers Problem 8. Show that with the Hamiltonian 2 H = a† a − N and any initial conditions, the master equation predicts that the expected number of rabbits grows exponentially. Answer. We use the same technique to compute X X d X 2 N Ψ(t) = N HΨ(t) = (N a† a − N )Ψ(t) dt First use the commutation relations to note: 2 N (a† a − N ) = N a† N − N 2 = a† (N + 1)N − N 2 Then: X (a† (N + 1)N − N 2 )Ψ(t) = X ((N + 1)N − N 2 )Ψ(t) = X N Ψ(t) So, we have X d X N Ψ(t) = N Ψ(t) dt It follows that the expected number of rabbits grows exponentially: X N Ψ(t) = cet 62 6 6 POPULATION BIOLOGY An example from population biology Stochastic Petri nets can be used to model everything from vending machines to chemical reactions. Chemists have proven some powerful theorems about when these systems have equilibrium states. We’d like to fit these results into our framework. We’ll do this soon. But first, let’s look at an example from population biology. 6.1 Amoeba fission and competition Here’s a stochastic Petri net: It shows a world with one species, amoeba, and two transitions: • fission, where one amoeba turns into two. Let’s call the rate constant for this transition α. • competition, where two amoebas battle for resources and only one survives. Let’s call the rate constant for this transition β. We are going to analyse this example in several ways. First we’ll study the deterministic dynamics it describes: we’ll look at its rate equation, which turns out to be the logistic equation, familiar in population biology. Then we’ll study the stochastic dynamics, meaning its master equation. That’s where the ideas from quantum field theory come in. 6.2 The rate equation If P (t) is the population of amoebas at time t, we can follow the rules explained in Section 2 and crank out this rate equation: dP = αP − βP 2 dt We can rewrite this as dP P = kP (1 − ) dt Q 63 6 6.2 POPULATION BIOLOGY The rate equation where Q= α , β k=α What’s the meaning of Q and k? • Q is the carrying capacity, that is, the maximum sustainable population the environment can support. • k is the growth rate describing the approximately exponential growth of population when P (t) is small. It’s a rare treat to find such an important differential equation that can be solved by analytical methods. Let’s enjoy solving it. We start by separating variables and integrating both sides: Z Z dP = kdt P (1 − P/Q) We need to use partial fractions on the left side above, resulting in Z Z Z dP dP + = kdt P Q−P and so we pick up a constant C, let A = ±e−C , and rearrange things as Q−P = Ae−kt P so the population as a function of time becomes P (t) = Q 1 + Ae−kt At t = 0 we can determine A uniquely. We write P0 := P (0) and A becomes A= Q − P0 P0 The model now becomes very intuitive. Let’s set Q = k = 1 and make a plot for various values of A: 64 6 6.3 POPULATION BIOLOGY The master equation We arrive at three distinct cases: • equilibrium (A = 0). The horizontal blue line corresponds to the case where the initial population P0 exactly equals the carrying capacity. In this case the population is constant. • dieoff (A < 0). The three decaying curves above the horizontal blue line correspond to cases where initial population is higher than the carrying capacity. The population dies off over time and approaches the carrying capacity. • growth (A > 0). The four increasing curves below the horizontal blue line represent cases where the initial population is lower than the carrying capacity. Now the population grows over time and approaches the carrying capacity. 6.3 The master equation Next, let us follow the rules explained in Section 5 to write down the master equation for our example. Remember, now we write: Ψ(t) = ∞ X ψn (t)z n n=0 where ψn (t) is the probability of having n amoebas at time t, and z is a formal variable. The master equation says: d Ψ(t) = HΨ(t) dt where H is an operator on formal power series called the Hamiltonian. To get the Hamiltonian we take each transition in our Petri net and build an operator from creation and annihilation operators, as follows. Fission works like this: while competition works like this: 65 6 6.3 POPULATION BIOLOGY The master equation Here a is the annihilation operator, a† is the creation operator and N = a† a is the number operator. In Section 5 we explained precisely how the N ’s arise. So the theory is already in place, and we arrive at this Hamiltonian: H = α(a† a† a − N ) + β(a† aa − N (N − 1)) Remember, α is the rate constant for fission, while β is the rate constant for competition. d The master equation can be solved: it’s equivalent to dt (e−tH Ψ(t)) = 0 so that e−tH Ψ(t) is constant, and so Ψ(t) = etH Ψ(0) and that’s it! We can calculate the time evolution starting from any initial probability distribution of populations. Maybe everyone is already used to this, but we find it rather remarkable. Here’s how it works. We pick a population, say n amoebas at t = 0. This would mean Ψ(0) = z n . We then evolve this state using etH . We expand this operator as P∞ 1 n n etH = n=0 n! t H = 1 + tH + 21 t2 H 2 + · · · This operator contains the full information for the evolution of the system. It contains the histories of all possible amoeba populations—an amoeba mosaic if you will. From this, we can construct amoeba Feynman diagrams. To do this, we work out each of the H n terms in the expansion above. The first-order terms correspond to the Hamiltonian acting once. These are proportional to either α or β. The second-order terms correspond to the Hamiltonian acting twice. These are proportional to either α2 , αβ or β 2 . And so on. This is where things start to get interesting! To illustrate how it works, we will consider two possibilities for the second-order terms: 1. We start with a lone amoeba, so Ψ(0) = z. It reproduces and splits into two. In the battle of the century, the resulting amoebas compete and one dies. At the end we have: αβ † (a aa)(a† a† a)z 2 66 6 6.3 POPULATION BIOLOGY The master equation We can draw this as a Feynman diagram: You might find this tale grim, and you may not like the odds either. It’s true, the odds could be better, but people are worse off than amoebas! The great Japanese swordsman Miyamoto Musashi quoted the survival odds of fair sword duels as 1/3, seeing that 1/3 of the time both participants die. A remedy is to cheat, but these amoeba are competing honestly. 2. We start with two amoebas, so the initial state is Ψ(0) = z 2 . One of these amoebas splits into two. One of these then gets into an argument about Petri nets with the original amoeba. The amoeba who solved all the problems in this book survives. At the end we have αβ † (a aa)(a† a† a)z 2 2 with corresponding Feynman diagram: 67 6 6.4 POPULATION BIOLOGY An equilibrium state This should give an idea of how this all works. The exponential of the Hamiltonian gives all possible histories, and each of these can be translated into a Feynman diagram. 6.4 An equilibrium state We’ve seen the equilibrium solution for the rate equation; now let’s look for equilibrium solutions of the master equation. This paper: [ACK] David F. Anderson, Georghe Craciun and Thomas G. Kurtz, Product-form stationary distributions for deficiency zero chemical reaction networks. Bulletin of Mathematical Biology, Volume 72, Issue 8, pages 1947-1970 (2010). Available as arXiv:0803.3042. proves that for a large class of stochastic Petri nets, there exists an equilibrium solution of the master equation where the number of things of each species is distributed according to a Poisson distribution. Even more remarkably, these probability distributions are independent, so knowing how many things are in one species tells you nothing about how many are in another! Here’s a nice quote from this paper: “The surprising aspect of the deficiency zero theorem is that the assumptions of the theorem are completely related to the network of the system whereas the conclusions of the theorem are related to the dynamical properties of the system.” The ‘deficiency zero theorem’ is a result of Feinberg, Horn and Jackson, which says that for a large class of stochastic Petri nets, the rate equation has a unique equilibrium solution. Anderson, Craciun and Kurtz showed how to use this fact to get equilibrium solutions of the master equation! 68 6 6.4 POPULATION BIOLOGY An equilibrium state Though they did not put it this way, their result relies crucially on the stochastic analogue of ‘coherent states’. Coherent states are important in quantum theory. Legend (or at least Wikipedia) has it that Erwin Schrödinger himself discovered them when he was looking for states of a quantum system that look ‘as classical as possible’. Suppose you have a quantum harmonic oscillator. Then the uncertainty principle says that ∆p∆q ≥ ~/2 where ∆p is the uncertainty in the momentum and ∆q is the uncertainty in position. Suppose we want to make ∆p∆q as small as possible, and suppose we also want ∆p = ∆q. Then we need our particle to be in a ‘coherent state’. That’s the definition. For the quantum harmonic oscillator, there’s a way to write quantum states as formal power series Ψ= ∞ X ψn z n n=0 where ψn is the amplitude for having n quanta of energy. A coherent state then looks like this: ∞ X cn n z Ψ = ecz = n! n=0 where c can be any complex number. Here we have omitted a constant factor necessary to normalize the state. We can also use coherent states in classical stochastic systems like collections of amoebas! Now the coefficient of z n tells us the probability of having n amoebas, so c had better be real. And probabilities should sum to 1, so we really should normalize Ψ as follows: ∞ X ecz cn n −c Ψ= c =e z e n! n=0 Now, the probability distribution ψn = e−c cn n! is called a Poisson distribution. So, for starters you can think of a ‘coherent state’ as an over-educated way of talking about a Poisson distribution. Let’s work out the expected number of amoebas in this Poisson distribution. In Section 5.10, we started using this abbreviation: X Ψ= ∞ X ψn n=0 We also saw that the expected number of amoebas in the probability distribution Ψ is X NΨ 69 6 6.4 POPULATION BIOLOGY An equilibrium state What does this equal? Remember that N = a† a. The annihilation operator a d is just dz , so aΨ = cΨ and we get X NΨ = X a† aΨ = c X a† Ψ But we saw in Section 4 that a† is stochastic, meaning X X a† Ψ = Ψ for any Ψ. Furthermore, our Ψ here has X Ψ=1 since it’s a probability distribution. So: X X X NΨ = c a† Ψ = c Ψ=c The expected number of amoebas is just c. Problem 9. This calculation must be wrong if c is negative: there can’t be a negative number of amoebas. What goes wrong then? Problem 10. Use the same tricks to calculate the standard deviation of the number of amoebas in the Poisson distribution Ψ. Now let’s return to our problem and consider the initial amoeba state Ψ = ecz Here we aren’t bothering to normalize it, because we’re going to look for equilibrium solutions to the master equation, meaning solutions where Ψ(t) doesn’t change with time. So, we want to solve HΨ = 0 Since this equation is linear, the normalization of Ψ doesn’t really matter. Remember, HΨ = α(a† a† a − N )Ψ + β(a† aa − N (N − 1))Ψ Let’s work this out. First consider the two α terms: a† a† aΨ = cz 2 Ψ and −N Ψ = −a† aΨ = −czΨ Likewise for the β terms we find a† aaΨ = c2 zΨ 70 6 6.5 POPULATION BIOLOGY Answers and −N (N − 1)ψ = −a† a† aaΨ = −c2 z 2 Ψ 2 Here we’re using a result from Section 5: the product a† a2 equals the ‘falling power’ N (N − 1). The sum of all four terms must vanish. This happens whenever α(cz 2 − cz) + β(c2 z − c2 z 2 ) = 0 which is satisfied for c= α β Yipee! We’ve found an equilibrium solution, since we found a value for c that makes HΨ = 0. Even better, we’ve seen that the expected number of amoebas in this equilibrium state is α β This is just the same as the equilibrium population we saw in the rate equation —that is, the logistic equation! That’s pretty cool, but it’s no coincidence: as we shall see, Anderson, Craciun and Kurtz proved it works like this for lots of stochastic Petri nets. 6.5 Answers Here are the answers to the problems, provided by David Corfield: Problem 9. We calculated that the expected number of amoebas in the Poisson distribution with parameter c is equal to c. But this can’t be true if c is negative: there can’t be a negative number of amoebas. What goes wrong then? Answer. If the probability of having n amoebas is given by the Poisson distribution cn ψn = e−c n! then c had better be nonnegative for the probability to be nonnegative when n = 1. Problem 10. Calculate the standard deviation of the number of amoebas in the Poisson distribution. Answer. The standard deviation is the square root of the variance, which is X X N 2Ψ − ( N Ψ)2 We have seen that for the Poisson distribution, X NΨ = c 71 6 6.5 POPULATION BIOLOGY Answers and using the same tricks we see P 2 N Ψ = = = = = P † † Pa aa aΨ c P a† aa† Ψ c P aa† Ψ c (a† a + 1)Ψ c(c + 1) So, the variance is c(c + 1) − c2 = c and the standard deviation is 72 √ c. 7 7 FEYNMAN DIAGRAMS Feynman diagrams We’ve already begun to see how Feynman diagrams show up in the study of stochastic Petri nets: in Section 6.3 we saw an example from population biology, involving amoebas. Now let’s dig a bit deeper, and look at an example involving more than one species. For this we need to generalize some of our notation a bit. 7.1 Stochastic Petri nets revisited First, recall that a Petri net consists of a set S of species and a set T of transitions, together with a function i: S × T → N saying how many things of each species appear in the input for each transition, and a function indexPetri net!transition o: S × T → N saying how many things of each species appear in the output. A stochastic Petri net is a Petri net together with a function r : T → (0, ∞) giving a rate constant for each transition. In this section, we’ll consider a Petri net with two species and three transitions: It should be clear that the transition ‘predation’ has one wolf and one rabbit as input, and two wolves as output. As we’ve seen, starting from any stochastic Petri net we can get two things. First: • The master equation. This says how the probability that we have a given number of things of each species changes with time. • The rate equation. This says how the expected number of things of each species changes with time. 73 7 7.2 FEYNMAN DIAGRAMS The rate equation The master equation is stochastic: it describes how probabilities change with time. The rate equation is deterministic. The master equation is more fundamental. It’s like the equations of quantum electrodynamics, which describe the amplitudes for creating and annihilating individual photons. The rate equation is less fundamental. It’s like the classical Maxwell equations, which describe changes in the electromagnetic field in a deterministic way. The classical Maxwell equations are an approximation to quantum electrodynamics. This approximation gets good in the limit where there are lots of photons all piling on top of each other to form nice waves. Similarly, the rate equation can be derived from the master equation in the limit where the number of things of each species become large, and the fluctuations in these numbers become negligible. But we won’t do this derivation! Nor will we probe more deeply into the analogy with quantum field theory, even though that’s the ultimate goal. For now we’ll be content to write down the master equation in a terse notation, and say a bit about Feynman diagrams. But to set up our notation, let’s start by recalling the rate equation. 7.2 The rate equation Suppose we have a stochastic Petri net with k different species. Let xi be the number of things of the ith species. Then the rate equation looks like this: dxi =??? dt It’s really a bunch of equations, one for each 1 ≤ i ≤ k. But what is the right-hand side? The right-hand side is a sum of terms, one for each transition in our Petri net. So, let’s start by assuming our Petri net has just one transition. Suppose the ith species appears as input to this transition mi times, and as output ni times. Then the rate equation is dxi mk 1 = r(ni − mi )xm 1 · · · xk dt where r is the rate constant for this transition. That’s really all there is to it! But we can make it look nicer. Let’s make up a vector x = (x1 , . . . , xk ) ∈ [0, ∞)k that says how many things there are of each species. Similarly let’s make up an input vector m = (m1 , . . . , mk ) ∈ Nk and an output vector n = (n1 , . . . , nk ) ∈ Nk 74 7 7.2 FEYNMAN DIAGRAMS The rate equation for our transition. To be cute, let’s also define mk 1 xm = xm 1 · · · xk Then we can write the rate equation for a single transition like this: dx = r(n − m)xm dt Next let’s do a general stochastic Petri net, with lots of transitions. Let’s write T for the set of transitions and r(τ ) for the rate constant of the transition τ ∈ T . Let n(τ ) and m(τ ) be the input and output vectors of the transition τ . Then the rate equation is: X dx = r(τ ) (n(τ ) − m(τ )) xm(τ ) dt τ ∈T For example, consider our rabbits and wolves: Suppose: • the rate constant for ‘birth’ is β, • the rate constant for ‘predation’ is γ, • the rate constant for ‘death’ is δ. Let x1 (t) be the number of rabbits and x2 (t) the number of wolves at time t. Then the rate equation looks like this: dx1 = βx1 − γx1 x2 dt dx2 = γx1 x2 − δx2 dt If you stare at this, and think about it, it should make perfect sense. If it doesn’t, go back and read Section 2. 75 7 7.3 FEYNMAN DIAGRAMS The master equation 7.3 The master equation Now let’s do something new. In Section 5 we explained how to write down the master equation for a stochastic Petri net with just one species. Now let’s generalize that. Luckily, the ideas are exactly the same. So, suppose we have a stochastic Petri net with k different species. Let ψn1 ,...,nk be the probability that we have n1 things of the first species, n2 of the second species, and so on. The master equation will say how all these probabilities change with time. To keep the notation clean, let’s introduce a vector n = (n1 , . . . , nk ) ∈ Nk and let ψn = ψn1 ,...,nk Then, let’s take all these probabilities and cook up a formal power series that has them as coefficients: as we’ve seen, this is a powerful trick. To do this, we’ll bring in some variables z1 , . . . , zk and write z n = z1n1 · · · zknk as a convenient abbreviation. Then any formal power series in these variables looks like this: X Ψ= ψn z n n∈Nk We call Ψ a state if the probabilities sum to 1 as they should: X ψn = 1 n The simplest example of a state is a monomial: z n = z1n1 · · · zknk This is a state where we are 100% sure that there are n1 things of the first species, n2 of the second species, and so on. We call such a state a pure state, since physicists use this term to describe a state where we know for sure exactly what’s going on. Sometimes a general state, one that might not be pure, is called mixed. The master equation says how a state evolves in time. It looks like this: d Ψ(t) = HΨ(t) dt So, we just need to tell you what H is! It’s called the Hamiltonian. It’s a linear operator built from special operators that annihilate and create things of various species. Namely, for each state 1 ≤ i ≤ k we have a annihilation operator: ai Ψ = d Ψ dzi 76 7 7.3 FEYNMAN DIAGRAMS The master equation and a creation operator: a†i Ψ = zi Ψ How do we build H from these? Suppose we’ve got a stochastic Petri net whose set of transitions is T . As before, write r(τ ) for the rate constant of the transition τ ∈ T , and let n(τ ) and m(τ ) be the input and output vectors of this transition. Then: X n(τ ) m(τ ) H= r(τ ) (a† − a† ) am(τ ) τ ∈T where as usual we’ve introduce some shorthand notations to keep from going insane. For example: m (τ ) m (τ ) am(τ ) = a1 1 · · · ak k and a† m(τ ) = a†1 m1 (τ ) · · · a†k mk (τ ) Now, it’s not surprising that each transition τ contributes a term to H. It’s also not surprising that this term is proportional to the rate constant r(τ ). The only tricky thing is the expression (a† n(τ ) − a† m(τ ) )am(τ ) How can we understand it? The basic idea is this. We’ve got two terms here. The first term: n(τ ) m(τ ) a† a describes how mi (τ ) things of the ith species get annihilated, and ni (τ ) things of the ith species get created. Of course this happens thanks to our transition τ . The second term: m(τ ) m(τ ) −a† a is a bit harder to understand, but it says how the probability that nothing happens—that we remain in the same pure state—decreases as time passes. Again this happens due to our transition τ . In fact, the second term must take precisely the form it does to ensure ‘conservation of total probability’. In other words: if the probabilities ψn sum to 1 at time zero, we want these probabilities to still sum to 1 at any later time. And for this, we need that second term to be what it is! In Section 5 we saw this in the special case where there’s only one species. The general case works the same way. Let’s look at an example. Consider our rabbits and wolves yet again: 77 7 7.4 FEYNMAN DIAGRAMS Feynman diagrams and again suppose the rate constants for birth, predation and death are β, γ and δ, respectively. We have X Ψ= ψn z n n where z n = z1n1 z2n2 and ψn = ψn1 ,n2 is the probability of having n1 rabbits and n2 wolves. These probabilities evolve according to the equation d Ψ(t) = HΨ(t) dt where the Hamiltonian is H = βB + γC + δD and B, C and D are operators describing birth, predation and death, respectively. (B stands for birth, D stands for death... and you can call predation ‘consumption’ if you want something that starts with C. Besides, ‘consumer’ is a nice euphemism for ‘predator’.) What are these operators? Just follow the rules we described: 2 B = a†1 a1 − a†1 a1 2 C = a†2 a1 a2 − a†1 a†2 a1 a2 D = a2 − a†2 a2 In each case, the first term is easy to understand: • Birth annihilates one rabbit and creates two rabbits. • Predation annihilates one rabbit and one wolf and creates two wolves. • Death annihilates one wolf. The second term is trickier, but we told you how it works. 7.4 Feynman diagrams How do we solve the master equation? If we don’t worry about mathematical rigor too much, it’s easy. The solution of d Ψ(t) = HΨ(t) dt should be Ψ(t) = etH Ψ(0) 78 7 7.4 FEYNMAN DIAGRAMS Feynman diagrams and we can hope that etH = 1 + tH + (tH)2 + ··· 2! so that t2 2 H Ψ(0) + · · · 2! Of course there’s always the question of whether this power series converges. In many contexts it doesn’t, but that’s not necessarily a disaster: the series can still be asymptotic to the right answer, or even better, Borel summable to the right answer. But let’s not worry about these subtleties yet! Let’s just imagine our rabbits and wolves, with Hamiltonian Ψ(t) = Ψ(0) + tHΨ(0) + H = βB + γC + δD Now, imagine working out Ψ(t) = Ψ(0) + tHΨ(0) + t3 t2 2 H Ψ(0) + H 3 Ψ(0) + · · · 2! 3! We’ll get lots of terms involving products of B, C and D hitting our original state Ψ(0). And we can draw these as diagrams! For example, suppose we start with one rabbit and one wolf. Then Ψ(0) = z1 z2 And suppose we want to compute H 3 Ψ(0) = (βB + γC + δD)3 Ψ(0) as part of the task of computing Ψ(t). Then we’ll get lots of terms: 27, in fact, though many will turn out to be zero. Let’s take one of these terms, for example the one proportional to: DCBΨ(0) We can draw this as a sum of Feynman diagrams, including this: 79 7 7.4 FEYNMAN DIAGRAMS Feynman diagrams In this diagram, we start with one rabbit and one wolf at top. As we read the diagram from top to bottom, first a rabbit is born (B), then predation occur (C), and finally a wolf dies (D). The end result is again a rabbit and a wolf. This is just one of four Feynman diagrams we should draw in our sum for DCBΨ(0), since either of the two rabbits could have been eaten, and either wolf could have died. So, the end result of computing H 3 Ψ(0) will involve a lot of Feynman diagrams... and of course computing Ψ(t) = Ψ(0) + tHΨ(0) + t3 t2 2 H Ψ(0) + H 3 Ψ(0) + · · · 2! 3! will involve even more, even if we get tired and give up after the first few terms. So, this Feynman diagram business may seem quite tedious... and it may not be obvious how it helps. But it does, sometimes! 80 7 7.4 FEYNMAN DIAGRAMS Feynman diagrams Now is not the time to describe ‘practical’ benefits of Feynman diagrams. Instead, we’ll just point out one conceptual benefit. We started with what seemed like a purely computational chore, namely computing Ψ(t) = Ψ(0) + tHΨ(0) + t2 2 H Ψ(0) + · · · 2! But then we saw how this series has a clear meaning! It can be written as a sum over diagrams, each of which represents a possible history of rabbits and wolves. So, it’s what physicists call a ‘sum over histories’. Feynman invented the idea of a sum over histories in the context of quantum field theory. At the time this idea seemed quite mind-blowing, for various reasons. First, it involved elementary particles instead of everyday things like rabbits and wolves. Second, it involved complex ‘amplitudes’ instead of real probabilities. Third, it actually involved integrals instead of sums. And fourth, a lot of these integrals diverged, giving infinite answers that needed to be ‘cured’ somehow. Now we’re seeing a sum over histories in a more down-to-earth context without all these complications. A lot of the underlying math is analogous... but now there’s nothing mind-blowing about it: it’s quite easy to understand. So, we can use this analogy to demystify quantum field theory a bit. On the other hand, thanks to this analogy, all sorts of clever ideas invented by quantum field theorists will turn out to have applications to biology and chemistry! So it’s a double win. 81 8 8 THE ANDERSON–CRACIUN–KURTZ THEOREM The Anderson–Craciun–Kurtz theorem In Section 7 we reviewed the rate equation and the master equation. Both of them describe processes where things of various kinds can react and turn into other things. But: • In the rate equation, we assume the number of things varies continuously and is known precisely. • In the master equation, we assume the number of things varies discretely and is known only probabilistically. This should remind you of the difference between classical mechanics and quantum mechanics. But the master equation is not quantum, it’s stochastic: it involves probabilities, but there’s no uncertainty principle going on. Still, a lot of the math is similar. Now, given an equilibrium solution to the rate equation—one that doesn’t change with time—we’ll try to find a solution to the master equation with the same property. We won’t always succeed—but we often can! The theorem saying how was proved here: [ACK] David F. Anderson, Georghe Craciun and Thomas G. Kurtz, Product-form stationary distributions for deficiency zero chemical reaction networks. Available as arXiv:0803.3042. To emphasize the analogy to quantum mechanics, Brendan Fong has translated their proof into the language of annihilation and creation operators. In particular, our equilibrium solution of the master equation is just like what people call a ‘coherent state’ in quantum mechanics. In what follows, we’ll present Fong’s ‘quantum proof’ of the Anderson– Craciun–Kurtz theorem. So if you know about quantum mechanics and coherent states, you should be happy. But if you don’t, fear not!—we’re not assuming you do. 8.1 The rate equation To construct our equilibrium solution of the master equation, we need a special type of solution to our rate equation. We call this type a ‘complex balanced solution’. This means that not only is the net rate of production of each species zero, but the net rate of production of each possible bunch of species is zero. Before we make this more precise, let’s remind ourselves of the basic setup. We’ll consider a stochastic Petri net with a finite set S of species and a finite set T of transitions. For convenience let’s take S = {1, . . . , k}, so our species are numbered from 1 to k. Then each transition τ has an input vector m(τ ) ∈ Nk and output vector n(τ ) ∈ Nk . These say how many things of each species go in, and how many go out. Each transition also has rate constant r(τ ) ∈ (0, ∞), which says how rapidly it happens. 82 8 8.2 THE ANDERSON–CRACIUN–KURTZ THEOREM Complex balance The rate equation concerns a vector x(t) ∈ [0, ∞)k whose ith component is the number of things of the ith species at time t. Note: we’re assuming this number of things varies continuously and is known precisely! This should remind you of classical mechanics. So, we’ll call x(t), or indeed any vector in [0, ∞)k , a classical state. The rate equation says how the classical state x(t) changes with time: X dx = r(τ ) (n(τ ) − m(τ )) xm(τ ) dt τ ∈T You may wonder what xm(τ ) means: after all, we’re taking a vector to a vector power! It’s just an abbreviation, which we’ve seen plenty of times before. If x ∈ Rk is a list of numbers and m ∈ Nk is a list of natural numbers, we define mk 1 xm = xm 1 · · · xk We’ll also use this notation when x is a list of operators. 8.2 Complex balance The vectors m(τ ) and n(τ ) are examples of what chemists call complexes. A complex is a bunch of things of each species. For example, if the set S consists of three species, the complex (1, 0, 5) is a bunch consisting of one thing of the first species, none of the second species, and five of the third species. For our Petri net, the set of complexes is the set Nk , and the complexes of particular interest are the input complex m(τ ) and the output complex n(τ ) of each transition τ . We say a classical state c ∈ [0, ∞)k is complex balanced if for all complexes κ ∈ Nk we have X X r(τ )cm(τ ) = r(τ )cm(τ ) {τ :m(τ )=κ} {τ :n(τ )=κ} The left hand side of this equation, which sums over the transitions with input complex κ, gives the rate of consumption of the complex κ. The right hand side, which sums over the transitions with output complex κ, gives the rate of production of κ. So, this equation requires that the net rate of production of the complex κ is zero in the classical state c. Problem 11. Show that if a classical state c is complex balanced, and we set x(t) = c for all t, then x(t) is a solution of the rate equation. Since x(t) doesn’t change with time here, we call it an equilibrium solution of the rate equation. Since x(t) = c is complex balanced, we call it complex balanced equilibrium solution. 83 8 8.3 THE ANDERSON–CRACIUN–KURTZ THEOREM The master equation 8.3 The master equation We’ve seen that any complex balanced classical state gives an equilibrium solution of the rate equation. The Anderson–Craciun–Kurtz theorem says that it also gives an equilibrium solution of the master equation. The master equation concerns a formal power series X Ψ(t) = ψn (t)z n n∈Nk where z n = z1n1 · · · zknk and ψn (t) = ψn1 ,...,nk (t) is the probability that at time t we have n1 things of the first species, n2 of the second species, and so on. Note: now we’re assuming this number of things varies discretely and is known only probabilistically! So, we’ll call Ψ(t), or indeed any formal power series where the coefficients are probabilities summing to 1, a stochastic state. Earlier we just called it a ‘state’, but that would get confusing now: we’ve got classical states and stochastic states, and we’re trying to relate them. The master equation says how the stochastic state Ψ(t) changes with time: d Ψ(t) = HΨ(t) dt where the Hamiltonian H is: n(τ ) X m(τ ) H= r(τ ) a† − a† am(τ ) τ ∈T The notation here is designed to neatly summarize some big products of annihilation and creation operators. For any vector n ∈ Nk , we have an = an1 1 · · · ank k and n a† = a†1 8.4 n1 · · · a†k nk Coherent states Now suppose c ∈ [0, ∞)k is a complex balanced equilibrium solution of the rate equation. We want to get an equilibrium solution of the master equation. How do we do it? For any c ∈ [0, ∞)k there is a stochastic state called a coherent state, defined by ecz Ψc = c e 84 8 8.4 THE ANDERSON–CRACIUN–KURTZ THEOREM Coherent states Here we are using some very terse abbreviations. Namely, we are defining ec = ec1 · · · eck and ecz = ec1 z1 · · · eck zk Equivalently, ecz = X cn zn n! k n∈N n n where c and z are defined as products in our usual way, and n! = n1 ! · · · nk ! Either way, if you unravel the abbrevations, here’s what you get: Ψc = e−(c1 +···+ck ) X cn1 · · · cnk 1 k z1n1 · · · zknk n ! · · · n ! 1 k k n∈N Maybe now you see why we like the abbreviations. The name ‘coherent state’ comes from quantum mechanics. In quantum mechanics, we think of a coherent state Ψc as the ‘quantum state’ that best approximates the classical state c. But we’re not doing quantum mechanics now, we’re doing probability theory. Ψc isn’t a ‘quantum state’, it’s a stochastic state. In probability theory, people like Poisson distributions. In the state Ψc , the probability of having ni things of the ith species is equal to e−ci cni i ni ! This is precisely the definition of a Poisson distribution with mean equal to ci . We can multiply a bunch of factors like this, one for each species, to get e−c cn n! This is the probability of having n1 things of the first species, n2 things of the second, and so on, in the state Ψc . So, the state Ψc is a product of independent Poisson distributions. In particular, knowing how many things there are of one species says nothing all about how many things there are of any other species! It is remarkable that such a simple state can give an equilibrium solution of the master equation, even for very complicated stochastic Petri nets. But it’s true—at least if c is complex balanced. 85 8 8.5 8.5 THE ANDERSON–CRACIUN–KURTZ THEOREM The proof Proof of the Anderson–Craciun–Kurtz theorem In a post on the Azimuth blog, Brendan Fong translated Anderson, Craciun and Kurtz’s proof of their theorem into the language of annihilation and creation operators. Now we’re ready to present Fong’s proof: Theorem 3 (Anderson–Craciun–Kurtz). Suppose c ∈ [0, ∞)k is a complex balanced equilibrium solution of the rate equation. Then HΨc = 0. Of course, it follows that Ψc is an equilibrium solution of the master equation. In other words, if we take Ψ(t) = Ψc for all times t, the master equation holds: d Ψ(t) = HΨ(t) dt since both sides are zero. Proof. We just need to show that HΨc = 0. Since Ψc is a constant times ecz , it suffices to show Hecz = 0. Remember that n(τ ) X m(τ ) Hecz = r(τ ) a† − a† am(τ ) ecz τ ∈T Since the annihilation operator ai is given by differentiation with respect to zi , while the creation operator a†i is just multiplying by zi , we have: X Hecz = r(τ ) cm(τ ) z n(τ ) − z m(τ ) ecz τ ∈T Expanding out ecz we get: He cz = XX r(τ )c m(τ ) z i n(τ ) c i i∈Nk τ ∈T i! z −z m(τ ) c i i! z i Shifting indices and defining negative powers to be zero: i−n(τ ) XX c ci−m(τ ) Hecz = r(τ )cm(τ ) zi − zi (i − n(τ ))! (i − m(τ ))! k i∈N τ ∈T So, to show Hecz = 0, we need to show this: XX i∈Nk τ ∈T r(τ )cm(τ ) XX ci−n(τ ) ci−m(τ ) zi = r(τ )cm(τ ) zi (i − n(τ ))! (i − m(τ ))! k i∈N τ ∈T We do this by splitting the sum over T according to output and then input 86 8 8.6 THE ANDERSON–CRACIUN–KURTZ THEOREM An example complexes, making use of the complex balanced condition: X X i∈Nk κ∈Nk X r(τ )cm(τ ) τ :n(τ )=κ ci−n(τ ) zi (i − n(τ ))! = X X i∈Nk = κ∈Nk X X i∈Nk κ∈Nk = X X ci−κ zi (i − κ)! ci−κ zi (i − κ)! X i∈Nk κ∈Nk τ :m(τ )=κ X r(τ )cm(τ ) τ :n(τ )=κ X r(τ )cm(τ ) τ :m(τ )=κ r(τ )cm(τ ) ci−m(τ ) zi (i − m(τ ))! This completes the proof! We hope you agree how amazing this result is. If you know quantum mechanics and coherent states you’ll understand what we mean. A coherent state is the “best quantum approximation”; to a classical state, but we don’t expect this quantum state to be exactly time-independent when the corresponding classical state is, except in very special cases, like when the Hamiltonian is quadratic in the creation and annihilation operators. Here we are getting a result like that much more generally... but only given the “complex balanced” condition. 8.6 An example We’ve already seen one example of the Anderson–Craciun–Kurtz theorem back in Section 6. We had this stochastic Petri net: We saw that the rate equation is just the logistic equation, familiar from population biology. The equilibrium solution is complex balanced, because pairs of amoebas are getting created at the same rate as they’re getting destroyed, and single amoebas are getting created at the same rate as they’re getting destroyed. So, the Anderson–Craciun–Kurtz theorem guarantees that there’s an equilibrium solution of the master equation where the number of amoebas is distributed according to a Poisson distribution. And, we actually checked that this was true! In the next section, we’ll look at another example. 87 8 8.7 8.7 THE ANDERSON–CRACIUN–KURTZ THEOREM Answers Answers Here is the answer to the problem, provided by David Corfield: Problem 11. Show that if a classical state c is complex balanced, and we set x(t) = c for all t, then x(t) is a solution of the rate equation. Answer. Assuming c is complex balanced, we have: X X X r(τ )m(τ )cm(τ ) = r(τ )m(τ )cm(τ ) κ τ :m(τ )=κ τ ∈T = X X r(τ )κcm(τ ) κ τ :m(τ )=κ = X X r(τ )κcm(τ ) κ τ :n(τ )=κ = X X r(τ )n(τ )cm(τ ) κ τ :n(τ )=κ = X r(τ )n(τ )cm(τ ) τ ∈T So, we have X r(τ ) (n(τ ) − m(τ )) cm(τ ) = 0 τ ∈T and thus if x(t) = c for all t then x(t) is a solution of the rate equation: X dx =0= r(τ ) (n(τ ) − m(τ )) xm(τ ) dt τ ∈T 88 9 9 A REVERSIBLE REACTION An example of the Anderson–Craciun–Kurtz theorem Let’s look at an example that illustrates the Anderson–Craciun–Kurtz theorem on equilibrium states. This example brings up an interesting ‘paradox’—or at least a problem. Resolving this will get us ready to think about a version of Noether’s theorem relating conserved quantities and symmetries. In Section 9 we’ll discuss Noether’s theorem in more detail. 9.1 A reversible reaction In chemistry a type of atom, molecule, or ion is called a chemical species, or species for short. Since we’re applying our ideas to both chemistry and biology, it’s nice that ‘species’ is also used for a type of organism in biology. This stochastic Petri net describes the simplest reversible reaction of all, involving two species: We have species 1 turning into species 2 with rate constant β, and species 2 turning back into species 1 with rate constant γ. So, the rate equation is: dx1 dt = −βx1 + γx2 dx2 dt = βx1 − γx2 where x1 and x2 are the amounts of species 1 and 2, respectively. 9.2 Equilibrium solutions of the rate equation Let’s look for equilibrium solutions of the rate equation, meaning solutions where the amount of each species doesn’t change with time. Equilibrium occurs when each species is getting created at the same rate at which it’s getting destroyed. 89 9 9.3 A REVERSIBLE REACTION Complex balanced equilibria So, let’s see when dx1 dx2 = =0 dt dt Clearly this happens precisely when βx1 = γx2 This says the rate at which 1’s are turning into 2’s equals the rate at which 2’s are turning back into 1’s. That makes perfect sense. 9.3 Complex balanced equilibria In general, a chemical reaction involves a bunch of species turning into a bunch of species. Since ‘bunch’ is not a very dignified term, a bunch of species is usually called a complex. We saw in Section 8 that it’s very interesting to study a strong version of equilibrium: complex balanced equilibrium, in which each complex is being created at the same rate at which it’s getting destroyed. However, in the Petri net we’re studying now, all the complexes being produced or destroyed consist of a single species. In this situation, any equilibrium solution is automatically complex balanced. This is great, because it means we can apply the Anderson–Craciun–Kurtz theorem from Section 8! This says how to get from a complex balanced equilibrium solution of the rate equation to an equilibrium solution of the master equation. First remember what the master equation says. Let ψn1 ,n2 (t) be the probability that we have n1 things of species 1 and n2 things of species 2 at time t. We summarize all this information in a formal power series: Ψ(t) = ∞ X ψn1 ,n2 (t)z1n1 z2n2 n1 ,n2 =0 Then the master equation says d Ψ(t) = HΨ(t) dt where following the general rules laid down in Section 7, = β(a†2 − a†1 )a1 + γ(a†1 − a†2 )a2 H This may look scary, but the annihilation operator ai and the creation operator a†i are just funny ways of writing the partial derivative ∂/∂zi and multiplication by zi , so H = β(z2 − z1 ) ∂ ∂ + γ(z1 − z2 ) ∂z1 ∂z2 or if you prefer, H = (z2 − z1 ) (β 90 ∂ ∂ −γ ) ∂z1 ∂z2 9 9.3 A REVERSIBLE REACTION Complex balanced equilibria The first term describes species 1 turning into species 2. The second describes species 2 turning back into species 1. Now, the Anderson–Craciun–Kurtz theorem says that whenever (x1 , x2 ) is a complex balanced solution of the rate equation, this recipe gives an equilibrium solution of the master equation: Ψ= ex1 z1 +x2 z2 ex1 +x2 In other words: whenever βx1 = γx2 , we have we have HΨ = 0 Let’s check this! For starters, the constant in the denominator of Ψ doesn’t matter here, since H is linear. It’s just a normalizing constant, put in to make sure that our probabilities ψn1 ,n2 sum to 1. So, we just need to check that (z2 − z1 )(β ∂ ∂ x1 z1 +x2 z2 −γ )e =0 ∂z1 ∂z2 If we do the derivatives on the left hand side, it’s clear we want (z2 − z1 )(βx1 − γx2 )ex1 z1 +x2 z2 = 0 and this is indeed true when βx1 = γx2 . So, the theorem works as advertised. And now we can work out the probability ψn1 ,n2 of having n1 things of species 1 and n2 of species 2 in our equilibrium state Ψ. To do this, we just expand the function Ψ as a power series and look at the coefficient of z1n1 z2n2 . We have Ψ= ∞ X (x1 z1 )n1 (x2 z2 )n2 1 ex1 z1 +x2 z2 = x +x x x e 1 2 e 1 e 2 n ,n n1 ! n2 ! 1 so we get ψn1 ,n2 = 2 1 xn1 1 1 xn1 2 · ex1 n1 ! ex2 n2 ! This is just a product of two independent Poisson distributions! In case you forget, a Poisson distribution says the probability of k events occurring in some interval of time if they occur with a fixed average rate and independently of the time since the last event. If the expected number of events is λ, the Poisson distribution is 1 λk eλ k! and it looks like this for various values of λ: 91 9 9.4 A REVERSIBLE REACTION Conserved quantities It looks almost like a Gaussian when λ is large, but when λ is small it becomes very lopsided. Anyway: we’ve seen that in our equilibrium state, the number of things of species i = 1, 2 is given by a Poisson distribution with mean xi . That’s very nice and simple... but the amazing thing is that these distributions are independent. Mathematically, this means we just multiply them to get the probability of finding n1 things of species 1 and n2 of species 2. But it also means that knowing how many things there are of one species says nothing about the number of the other. But something seems odd here. One transition in our Petri net consumes a 1 and produces a 2, while the other consumes a 2 and produces a 1. The total number of particles in the system never changes. The more 1’s there are, the fewer 2’s there should be. But we just said knowing how many 1’s we have tells us nothing about how many 2’s we have! At first this seems like a paradox. Have we made a mistake? Not exactly. But we’re neglecting something. 9.4 Conserved quantities Namely: the equilibrium solutions of the master equation we’ve found so far are not the only ones! There are other solutions that fit our intuitions better. Suppose we take any of our equilibrium solutions Ψ and change it like this: set the probability ψn1 ,n2 equal to 0 unless n1 + n2 = n but otherwise leave it unchanged. Of course the probabilities no longer sum to 1, but we can rescale them so they do. 92 9 9.4 A REVERSIBLE REACTION Conserved quantities The result is a new equilibrium solution, say Ψn . Why? Because, as we’ve already seen, no transitions will carry us from one value of n1 + n2 to another. And in this new solution, the number of 1’s is clearly not independent from the number of 2’s. The bigger one is, the smaller the other is. Problem 12. Show that this new solution Ψn depends only on n and the ratio x1 /x2 = γ/β, not on anything more about the values of x1 and x2 in the original solution ex1 z1 +x2 z2 Ψ= ex1 +x2 Problem 13. What is this new solution like when β = γ? (This particular choice makes the problem symmetrical when we interchange species 1 and 2.) What’s happening here is that this particular stochastic Petri net has a ‘conserved quantity’: the total number of things never changes with time. So, we can take any equilibrium solution of the master equation and—in the language of quantum mechanics—‘project down to the subspace’ where this conserved quantity takes a definite value, and get a new equilibrium solution. In the language of probability theory, we say it a bit differently: we’re ‘conditioning on’ the conserved quantity taking a definite value. But the idea is the same. This important feature of conserved quantities suggests that we should try to invent a new version of Noether’s theorem. This theorem links conserved quantities and symmetries of the Hamiltonian. There are already a couple versions of Noether’s theorem for classical mechanics, and for quantum mechanics... but now we want a version for stochastic mechanics. And indeed one exists, and it’s relevant to what we’re doing here. So let us turn to that. 93 10 10 NOETHER’S THEOREM A stochastic version of Noether’s theorem Noether proved lots of theorems, but when people talk about Noether’s theorem, they always seem to mean her result linking symmetries to conserved quantities. Her original result applied to classical mechanics, but now we’d like to present a version that applies to ‘stochastic mechanics’—or in other words, Markov processes. This theorem was proved here: [BF] John Baez and Brendan Fong, A Noether theorem for Markov processes, J. Math. Phys. 54 (2013), 013301. Also available as arXiv:1203.2035. What’s a Markov process? We’ll say more in a minute—but in plain English, it’s a physical system where something hops around randomly from state to state, where its probability of hopping anywhere depends only on where it is now, not its past history. Markov processes include, as a special case, the stochastic Petri nets we’ve been talking about. Our stochastic version of Noether’s theorem is copied after a well-known quantum version. It’s yet another example of how we can exploit the analogy between stochastic mechanics and quantum mechanics. But for now we’ll just present the stochastic version. In the next section, we’ll compare it to the quantum one. 10.1 Markov processes We should and probably will be more general, but let’s start by considering a finite set of states, say X. To describe a Markov process we then need a matrix of real numbers H = (Hij )i,j∈X . The idea is this: suppose right now our system is in the state i. Then the probability of being in some state j changes as time goes by—and Hij is defined to be the time derivative of this probability right now. So, if ψi (t) is the probability of being in the state i at time t, we want the master equation to hold: X d ψi (t) = Hij ψj (t) dt j∈X This motivates the definition of ‘infinitesimal stochastic’, which we gave in a more general context back in Section 4: Definition 4. Given a finite set X, a matrix of real numbers H = (Hij )i,j∈X is infinitesimal stochastic if i 6= j ⇒ Hij ≥ 0 and X Hij = 0 i∈X for all j ∈ X. 94 10 10.1 NOETHER’S THEOREM Markov processes The inequality says that if we start in the state i, the probability of being in some other state, which starts at 0, can’t go down, at least initially. The equation says that the probability of being somewhere or other doesn’t change. Together, these facts imply that: Hii ≤ 0 That makes sense: the probability of being in the state i, which starts at 1, can’t go up, at least initially. Using the magic of matrix multiplication, we can rewrite the master equation as follows: d ψ(t) = Hψ(t) dt and we can solve it like this: ψ(t) = exp(tH)ψ(0) If H is an infinitesimal stochastic operator, we will call exp(tH) a Markov process, and H its Hamiltonian. (Actually, most people call exp(tH) a Markov semigroup, and reserve the term Markov process for another way of looking at the same idea. So, be careful.) Noether’s theorem is about ‘conserved quantities’, that is, observables whose expected values don’t change with time. To understand this theorem, you need to know a bit about observables. In stochastic mechanics an observable is simply a function assigning a number Oi to each state i ∈ X. However, in quantum mechanics we often think of observables as matrices, so it’s nice to do that here, too. It’s easy: we just create a matrix whose diagonal entries are the values of the function O. And just to confuse you, we’ll also call this matrix O. So: Oi if i = j Oij = 0 if i 6= j One advantage of this trick is that it lets us ask whether an observable commutes with the Hamiltonian. Remember, the commutator of matrices is defined by [O, H] = OH − HO Noether’s theorem will say that [O, H] = 0 if and only if O is ‘conserved’ in some sense. What sense? First, recall that a stochastic state is just our fancy name for a probability distribution ψ on the set X. Second, the expected value of an observable O in the stochastic state ψ is defined to be X O i ψi i∈X In Section 4 we introduced the notation Z X φ= φi i∈X 95 10 10.1 NOETHER’S THEOREM Markov processes for any function φ on X. The reason is that later, when we generalize X from a finite set to a measure space, the sum at right will become an integral over X. Indeed, a sum is just a special sort of integral! Using this notation and the magic of matrix multiplication, we can write the expected value of O in the stochastic state ψ as Z Oψ We can calculate how this changes in time if ψ obeys the master equation... and we can write the answer using the commutator [O, H]: Lemma 5. Suppose H is an infinitesimal stochastic operator and O is an observable. If ψ(t) obeys the master equation, then Z Z d Oψ(t) = [O, H]ψ(t) dt Proof. Using the master equation we have Z Z Z d d Oψ(t) = O ψ(t) = OHψ(t) dt dt But since H is infinitesimal stochastic, X Hij = 0 i∈X so for any function φ on X we have Z X Hφ = Hij φj = 0 i,j∈X and in particular Z HOψ(t) = 0 Since [O, H] = OH − HO, we conclude from (1) and (2) that Z Z d Oψ(t) = [O, H]ψ(t) dt as desired. The commutator doesn’t look like it’s doing much here, since we also have Z Z d Oψ(t) = OHψ(t) dt which is even simpler. But the commutator will become useful when we get to Noether’s theorem! 96 10 10.2 10.2 NOETHER’S THEOREM Noether’s theorem Noether’s theorem Here’s a version of Noether’s theorem for Markov processes. It says an observable commutes with the Hamiltonian iff the expected values of that observable and its square don’t change as time passes: Theorem 6. Suppose H is an infinitesimal stochastic operator and O is an observable. Then [O, H] = 0 if and only if and d dt Z d dt Z Oψ(t) = 0 O2 ψ(t) = 0 for all ψ(t) obeying the master equation. If you know Noether’s theorem from quantum mechanics, you might be surprised that in this version we need not only the observable but also its square to have an unchanging expected value! We’ll explain this, but first let’s prove the theorem. R d Proof. The easy part is showing that if [O, H] = 0 then dt Oψ(t) = 0 and R 2 d O ψ(t) = 0. In fact there’s nothing special about these two powers of t; dt we’ll show that Z d On ψ(t) = 0 dt for all n. The point is that since H commutes with O, it commutes with all powers of O: [On , H] = 0 So, applying the Lemma to the observable On , we see Z Z d On ψ(t) = [On , H]ψ(t) = 0 dt The backward direction is a bit trickier. We now assume that Z Z d d Oψ(t) = O2 ψ(t) = 0 dt dt for all solutions ψ(t) of the master equation. This implies Z Z OHψ(t) = O2 Hψ(t) = 0 or since this holds for all solutions, X X Oi Hij = Oi2 Hij = 0 i∈X i∈X 97 10 10.3 NOETHER’S THEOREM Markov chains We wish to show that [O, H] = 0. First, recall that we can think of O as a diagonal matrix with: Oi if i = j Oij = 0 if i 6= j So, we have [O, H]ij = X (Oik Hkj − Hik Okj ) = Oi Hij − Hij Oj = (Oi − Oj )Hij k∈X To show this is zero for each pair of elements i, j ∈ X, it suffices to show that when Hij 6= 0, then Oj = Oi . That is, we need to show that if the system can move from state j to state i, then the observable takes the same value on these two states. In fact, it’s enough to show that this sum is zero for any j ∈ X: X (Oj − Oi )2 Hij i∈X Why? When i = j, Oj − Oi = 0, so that term in the sum vanishes. But when i 6= j, (Oj − Oi )2 and Hij are both non-negative—the latter because H is infinitesimal stochastic. So if they sum to zero, they must each be individually zero. Thus for all i 6= j, we have (Oj − Oi )2 Hij = 0. But this means that either Oi = Oj or Hij = 0, which is what we need to show. So, let’s take that sum and expand it: X X (Oj − Oi )2 Hij = (Oj2 Hij − 2Oj Oi Hij + Oi2 Hij ) i X i∈X X X = Oj2 Hij − 2Oj Oi Hij + Oi2 Hij i i i The three terms here are each zero: the first because H is infinitesimal stochastic, and the latter two by equation (3). So, we’re done! 10.3 Markov chains So that’s the proof... but why do we need both O and its square to have an expected value that doesn’t change with time to conclude [O, H] = 0? There’s an easy counterexample if we leave out the condition involving O2 . However, the underlying idea is clearer if we work with Markov chains instead of Markov processes. In a Markov process, time passes by continuously. In a Markov chain, time comes in discrete steps! We get a Markov process by forming exp(tH) where H is an infinitesimal stochastic operator. We get a Markov chain by forming the operator U, U 2 , U 3 , . . . where U is a ‘stochastic operator’. Remember: 98 10 10.3 NOETHER’S THEOREM Markov chains Definition 7. Given a finite set X, a matrix of real numbers U = (Uij )i,j∈X is stochastic if Uij ≥ 0 for all i, j ∈ X and X Uij = 1 i∈X for all j ∈ X. The idea is that U describes a random hop, with Uij being the probability of hopping to the state i if you start at the state j. These probabilities are nonnegative and sum to 1. Any stochastic operator gives rise to a Markov chain U, U 2 , U 3 , . . . . And in case it’s not clear, that’s how we’re defining a Markov chain: the sequence of powers of a stochastic operator. There are other definitions, but they’re equivalent. We can draw a Markov chain by drawing a bunch of states and arrows labelled by transition probabilities, which are the matrix elements Uij : Here is Noether’s theorem for Markov chains: Theorem 8. Suppose U is a stochastic operator and O is an observable. Then [O, U ] = 0 if and only if Z Z OU ψ = and Z O2 U ψ = for all stochastic states ψ. 99 Oψ Z O2 ψ 10 10.3 NOETHER’S THEOREM Markov chains In other words, an observable commutes with U iff the expected values of that observable and its square don’t change when we evolve our state one time step using U . You can prove this theorem by copying the proof for Markov processes: Problem 14. Prove Noether’s theorem for Markov chains. But let’s see why we need the condition on the square of observable! That’s the intriguing part. Here’s a nice little Markov chain: where we haven’t drawn arrows labelled by 0. So, state 1 has a 50% chance of hopping to state 0 and a 50% chance of hopping to state 2; the other two states just sit there. Now, consider the observable O with Oi = i It’s easy to check that the expected value of this observable doesn’t change with time: Z Z OU ψ = Oψ for all ψ. The reason, in plain English, is this. Nothing at all happens if you start at states 0 or 2: you just sit there, so the expected value of O doesn’t change. If you start at state 1, the observable equals 1. You then have a 50% chance of going to a state where the observable equals 0 and a 50% chance of going to a state where it equals 2, so its expected value doesn’t change: it still equals 1. On the other hand, we do not have [O, U ] = 0 in this example, because we can hop between states where O takes different values. Furthermore, Z Z O2 U ψ 6= O2 ψ After all, if you start at state 1, O2 equals 1 there. You then have a 50% chance of going to a state where O2 equals 0 and a 50% chance of going to a state where it equals 4, so value changes! R its expected R So, that’s why OU ψ = Oψ for all ψ is not enough to guarantee [O, U ] = 0. The same sort of counterexample works for Markov processes, too. Finally, we should add that there’s nothing terribly sacred about the square of the observable. For example, we have: Theorem 9. Suppose H is an infinitesimal stochastic operator and O is an observable. Then [O, H] = 0 100 10 10.4 NOETHER’S THEOREM Answers if and only if d dt Z f (O)ψ(t) = 0 for all smooth f : R → R and all ψ(t) obeying the master equation. Theorem 10. Suppose U is a stochastic operator and O is an observable. Then [O, U ] = 0 if and only if Z Z f (O)U ψ = f (O)ψ for all smooth f : R → R and all stochastic states ψ. These make the ‘forward direction’ of Noether’s theorem stronger... and in fact, the forward direction, while easier, is probably more useful! However, if you ever use Noether’s theorem in the ‘reverse direction’, it might be easier to check a condition involving only O and its square. 10.4 Answers Here’s the answer to the problem: Problem 14. Suppose U is a stochastic operator and O is an observable. Show that O commutes with U iff the expected values of O and its square don’t change when we evolve our state one time step using U . In other words, show that [O, U ] = 0 if and only if Z Z OU ψ = and Z O2 U ψ = Oψ Z O2 ψ for all stochastic states ψ. Answer. One direction is easy: if [O, U ] = 0 then [On , U ] = 0 for all n, so Z Z Z On U ψ = U On ψ = On ψ where in the last step we use the fact that U is stochastic. For the converse we can use the same tricks that worked for Markov processes. Assume that Z Z OU ψ = Oψ 101 10 10.4 NOETHER’S THEOREM Answers and Z Z 2 O Uψ = O2 ψ for all stochastic states ψ. These imply that X Oi Uij = Oj (1) i∈X and X Oi2 Uij = Oj2 (2) i∈X We wish to show that [O, U ] = 0. Note that X [O, U ]ij = (Oik Ukj − Uik Okj ) k∈X (Oi − Oj )Uij = To show this is always zero, we’ll show that when Uij 6= 0, then Oj = Oi . This says that when our system can hop from one state to another, the observable O must take the same value on these two states. For this, in turn, it’s enough to show that the following sum vanishes for any j ∈ X: X (Oj − Oi )2 Uij i∈X Why? The matrix elements Uij are nonnegative since U is stochastic. Thus the sum can only vanish if each term vanishes, meaning that Oj = Oi whenever Uij 6= 0. To show the sum vanishes, let’s expand it: X X (Oj − Oi )2 Uij = (Oj2 Uij − 2Oj Oi Uij + Oi2 Uij ) i i∈X = Oj2 X Uij − 2Oj i X Oi Uij + i X Oi2 Uij i Now, since (1) and (2) hold for all stochastic states ψ, this equals X Oj2 Uij − 2Oj2 + Oj2 i But this is zero because U is stochastic, which implies X Uij = 1 i So, we’re done! 102 11 11 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Quantum mechanics vs stochastic mechanics In Section 10 we proved a version of Noether’s theorem for stochastic mechanics. Now we want to compare that to the more familiar quantum version. But to do this, we need to say more about the analogy between stochastic mechanics and quantum mechanics. So far, whenever we’ve tried, we feel pulled toward explaining some technical issues involving analysis: whether sums converge, whether derivatives exist, and so on. We’ve been trying to avoid such stuff— not because we dislike it, but because we’re afraid you might. But the more we put off discussing these issues, the more they fester. So, this time we will gently explore some of these technical issues. But don’t be scared: we’ll mainly talk about some simple big ideas. In the next section, we’ll use these to compare the two versions of Noether’s theorem. To beging with, we need to recall the analogy we began sketching in Section 4, and push it a bit further. The idea is that stochastic mechanics differs from quantum mechanics in two big ways: • First, instead of complex amplitudes, stochastic mechanics uses nonnegative real probabilities. The complex numbers form a ring; the nonnegative real numbers form a mere rig, which is a ‘ring without negatives’. Rigs are much neglected in the typical math curriculum, but unjustly so: they’re almost as good as rings in many ways, and there are lots of important examples, like the natural numbers N and the nonnegative real numbers, [0, ∞). For probability theory, we should learn to love rigs. But there are, alas, situations where we need to subtract probabilities, even when the answer comes out negative: namely when we’re taking the time derivative of a probability. So sometimes we need R instead of just [0, ∞). • Second, while in quantum mechanics a state is described using a ‘wavefunction’, meaning a complex-valued function obeying Z |ψ|2 = 1 in stochastic mechanics it’s described using a ‘probability distribution’, meaning a nonnegative real function obeying Z ψ=1 So, let’s try our best to present the theories in close analogy, while respecting these two differences. 11.1 States We’ll start with a set X whose points are states that a system can be in. So far, we assumed X was a finite set, but this section is so mathematical we might 103 11 11.1 QUANTUM MECHANICS VS STOCHASTIC MECHANICS States as well assume it’s a measure space. A measure space lets you do integrals, but a finite set is a special case, and then these integrals are just sums. So, we’ll write things like Z f and mean the integral of the function f over the measure space X, but if X is a finite set this just means X f (x) x∈X Now, we’ve already defined the word ‘state’, but both quantum and stochastic mechanics need a more general concept of state. Let’s call these ‘quantum states’ and ‘stochastic states’: • In quantum mechanics, the system has an amplitude ψ(x) of being in any state x ∈ X. These amplitudes are complex numbers with Z |ψ|2 = 1 We call ψ : X → C obeying this equation a quantum state. • In stochastic mechanics, the system has a probability ψ(x) of being in any state x ∈ X. These probabilities are nonnegative real numbers with Z ψ=1 We call ψ : X → [0, ∞) obeying this equation a stochastic state. In quantum mechanics we often use this abbreviation: Z hφ, ψi = φψ so that a quantum state has hψ, ψi = 1 Similarly, we could introduce this notation in stochastic mechanics: Z hψi = ψ so that a stochastic state has hψi = 1 But this notation is a bit risky, since angle brackets of this Rsort often stand for expectation values of observables. So, we’ve been writing ψ, and we’ll keep on doing this most of the time. 104 11 11.2 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Observables In quantum mechanics, hφ, ψi is well-defined whenever both φ and ψ live in the vector space Z L2 (X) = {ψ : X → C : |ψ|2 < ∞} In stochastic mechanics, hψi is well-defined whenever ψ lives in the vector space Z L1 (X) = {ψ : X → R : |ψ| < ∞} You’ll notice we wrote R rather than [0, ∞) here. That’s because in some calculations we’ll need functions that take negative values, even though our stochastic states are nonnegative. 11.2 Observables A state is a way our system can be. An observable is something we can measure about our system. They fit together: we can measure an observable when our system is in some state. If we repeat this we may get different answers, but there’s a nice formula for average or ‘expected’ answer. In quantum mechanics, an observable is a self-adjoint operator A on L2 (X). In the case where X is a finite set, a self-adjoint operator on L2 (X) is just one with hψ, Aφi = hAψ, φi for all ψ, φ ∈ L2 (X). In general there are some extra subtleties, which we prefer to sidestep here. The expected value of an observable A in the state ψ is Z hψ, Aψi = ψAψ Here we’re assuming that we can apply A to ψ and get a new vector Aψ ∈ L2 (X). This is automatically true when X is a finite set, but in general we need to be more careful. In stochastic mechanics, an observable is a real-valued function A on X. The expected value of A in the state ψ is Z hAψi = Aψ Here we’re using the fact that we can multiply A and ψ and get a new vector Aψ ∈ L1 (X), at least if A is bounded. Again, this is automatic if X is a finite set, but not otherwise. The left side of the equation above is just an abbreviation for the right side, where we write the integral sign using angle brackets. We mentioned in the previous section that this use of angle brackets is a bit risky, so we won’t use it often. It does make the analogy between quantum mechanics and stochastic mechanics rather pretty, though. There are a couple of other equivalent ways to write the expected value of an observable in stochastic mechanics: 105 11 11.3 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Symmetries • Sometimes the stochastic state ψ ∈ L1 (X) and the constant function 1 both lie L2 (X). This R happens whenever X is a finite measure space, meaning one with 1 < ∞. For example, it happens when X is just a finite set with integration being summation. In this situation, we have Z Aψ = h1, Aψi This gives the expected value of stochastic observables a bit more of a quantum flavor! We consider this formula to be more of a ‘trick’ than a deep insight. Sometimes it may be useful, though. • Sometimes the stochastic state ψ ∈ L1 (X) and the observable A, which is a bounded function on X, both lie in L2 (X). This is always the case when X is a finite set with integration being summation. In this situation we can think of A as vector Â ∈ L2 (X), and we have Z Aψ = hÂ, ψi Again, this formula has a kind of quantum flavor. We shall use this formula in Section ??. 11.3 Symmetries Besides states and observables, we need ‘symmetries’, which are transformations that map states to states. For example, we use these to describe how our system changes when we wait a while. In quantum mechanics, an isometry is a linear map U : L2 (X) → L2 (X) such that hU φ, U ψi = hφ, ψi for all ψ, φ ∈ L2 (X). If U is an isometry and ψ is a quantum state, then U ψ is again a quantum state. In stochastic mechanics, a stochastic operator is a linear map U : L1 (X) → L1 (X) such that Z Z Uψ = ψ and ψ ≥ 0 ⇒ Uψ ≥ 0 for all ψ ∈ L1 (X). If U is stochastic and ψ is a stochastic state, then U ψ is again a stochastic state. In quantum mechanics we are mainly interested in invertible isometries, which are called unitary operators. There are lots of these, and their inverses are always isometries. There are, however, very few stochastic operators whose inverses are stochastic: 106 11 11.4 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Quantum evolution Problem 15. Suppose X is a finite set. Show that any isometry U : L2 (X) → L2 (X) is invertible, and its inverse is again an isometry. Problem 16. Suppose X is a finite set. Which stochastic operators U : L1 (X) → L1 (X) have stochastic inverses? This is why we usually think of time evolution as being reversible quantum mechanics, but not in stochastic mechanics! In quantum mechanics we often describe time evolution using a ‘1-parameter group’, while in stochastic mechanics we describe it using a 1-parameter semi group... meaning that we can run time forwards, but not backwards. But let’s see how this works in detail! 11.4 Time evolution in quantum mechanics In quantum mechanics there’s a beautiful relation between observables and symmetries, which goes like this. Suppose that for each time t we want a unitary operator U (t) : L2 (X) → L2 (X) that describes time evolution. Then it makes a lot of sense to demand that these operators form a 1-parameter group: Definition 11. A collection of linear operators U(t) (t ∈ R) on some vector space forms a 1-parameter group if U (0) = 1 and U (s + t) = U (s)U (t) for all s, t ∈ R. Note that these conditions force all the operators U (t) to be invertible. Now suppose our vector space is a Hilbert space, like L2 (X). Then we call a 1-parameter group a 1-parameter unitary group if the operators involved are all unitary. It turns out that 1-parameter unitary groups are either continuous in a certain way, or so pathological that you can’t even prove they exist without the axiom of choice! So, we always focus on the continuous case: Definition 12. A 1-parameter unitary group is strongly continuous if U (t)ψ depends continuously on t for all ψ, in this sense: ti → t ⇒ kU (ti )ψ − U (t)ψk → 0 Then we get a classic result proved by Marshall Stone back in the early 1930s. You may not know him, but he was so influential at the University of Chicago during this period that it’s often called the ‘Stone Age’. And here’s one reason why: 107 11 11.4 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Quantum evolution Theorem 13 (Stone’s Theorem). There is a one-to-one correspondence between strongly continuous 1-parameter unitary groups on a Hilbert space and self-adjoint operators on that Hilbert space, given as follows. Given a strongly continuous 1-parameter unitary group U (t) we can always write U (t) = exp(−itH) for a unique self-adjoint operator H. Conversely, any self-adjoint operator determines a strongly continuous 1-parameter group this way. For all vectors ψ for which Hψ is well-defined, we have d = −iHψ U (t)ψ dt t=0 Moreover, for any of these vectors, if we set ψ(t) = exp(−itH)ψ we have d ψ(t) = −iHψ(t) dt When U (t) = exp(−itH) describes the evolution of a system in time, H is is called the Hamiltonian, and it has the physical meaning of ‘energy’. The equation we just wrote down is then called Schrödinger’s equation. So, simply put, in quantum mechanics we have a correspondence between observables and nice one-parameter groups of symmetries. Not surprisingly, our favorite observable, energy, corresponds to our favorite symmetry: time evolution! However, if you were paying attention, you noticed that we carefully avoided explaining how we define exp(−itH). We didn’t even say what a self-adjoint operator is. This is where the technicalities come in: they arise when H is unbounded, and not defined on all vectors in our Hilbert space. Luckily, these technicalities evaporate for finite-dimensional Hilbert spaces, such as L2 (X) for a finite set X. Then we get: Theorem 14 (Stone’s Theorem—Baby Version). Suppose we are given a finite-dimensional Hilbert space. In this case, a linear operator H on this space is self-adjoint iff it’s defined on the whole space and hφ, Hψi = hHφ, ψi for all vectors φ, ψ. Given a strongly continuous 1-parameter unitary group U (t) we can always write U (t) = exp(−itH) for a unique self-adjoint operator H, where exp(−itH)ψ = ∞ X (−itH)n ψ n! n=0 108 11 11.5 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Stochastic evolution with the sum converging for all ψ. Conversely, any self-adjoint operator on our space determines a strongly continuous 1-parameter group this way. For all vectors ψ in our space we then have d = −iHψ U (t)ψ dt t=0 and if we set ψ(t) = exp(−itH)ψ we have 11.5 d ψ(t) = −iHψ(t) dt Time evolution in stochastic mechanics We’ve seen that in quantum mechanics, time evolution is usually described by a 1-parameter group of operators that comes from an observable: the Hamiltonian. Stochastic mechanics is different! First, since stochastic operators aren’t usually invertible, we typically describe time evolution by a mere ‘semigroup’: Definition 15. A collection of linear operators U (t) on some vector space, where t ∈ [0, ∞), forms a 1-parameter semigroup if U (0) = 1 and U (s + t) = U (s)U (t) for all s, t ≥ 0. Now suppose this vector space is L1 (X) for some measure space X. We want to focus on the case where the operators U (t) are stochastic and depend continuously on t in the same sense we discussed earlier. Definition 16. A 1-parameter strongly continuous semigroup of stochastic operators U (t) : L1 (X) → L1 (X) is called a Markov semigroup. What’s the analogue of Stone’s theorem for Markov semigroups? We don’t know a fully satisfactory answer! If you know, please tell us. In Section we’ll say what we do know—we’re not completely clueless—but for now let’s look at the ‘baby’ case where X is a finite set. Then the story is neat and complete: Theorem 17. Suppose we are given a finite set X. In this case, a linear operator H on L1 (X) is infinitesimal stochastic iff it’s defined on the whole space, Z Hψ = 0 109 11 11.6 QUANTUM MECHANICS VS STOCHASTIC MECHANICS The Hille–Yosida theorem for all ψ ∈ L1 (X), and the matrix of H in terms of the obvious basis obeys Hij ≥ 0 for all j 6= i. Given a Markov semigroup U (t) on L1 (X), we can always write U (t) = exp(tH) for a unique infinitesimal stochastic operator H, where exp(tH)ψ = ∞ X (tH)n ψ n! n=0 with the sum converging for all ψ. Conversely, any infinitesimal stochastic operator on our space determines a Markov semigroup this way. For all ψ ∈ L1 (X) we then have d U (t)ψ = Hψ dt t=0 and if we set ψ(t) = exp(tH)ψ we have the master equation: d ψ(t) = Hψ(t) dt In short, time evolution in stochastic mechanics is a lot like time evolution in quantum mechanics, except it’s typically not invertible, and the Hamiltonian is typically not an observable. Why not? Because we defined an observable to be a function A : X → R. We can think of this as giving an operator on L1 (X), namely the operator of multiplication by A. That’s a nice trick, which we used to good effect in Section 10. However, at least when X is a finite set, this operator will be diagonal in the obvious basis consisting of functions that equal 1 one point of X and zero elsewhere. So, it can only be infinitesimal stochastic if it’s zero! Problem 17. If X is a finite set, show that any operator on L1 (X) that’s both diagonal and infinitesimal stochastic must be zero. 11.6 The Hille–Yosida theorem We’ve now told you everything you really need to know about the broad analogy between quantum mechanics and stochastic mechanics... but not everything we want to say. What happens when X is not a finite set? What are Markov semigroups like then? We can’t abide letting this question go unresolved! Unfortunately we only know a partial answer. We can get a certain distance using the ‘Hille–Yosida theorem’, but this theorem does not completely characterize Markov semigroups. Still, it’s worth knowing. We need a few concepts to get started: 110 11 11.6 QUANTUM MECHANICS VS STOCHASTIC MECHANICS The Hille–Yosida theorem Definition 18. A Banach space is a vector space with a norm such that any Cauchy sequence converges. Examples include Hilbert spaces like L2 (X) for any measure space, but also other spaces like L1 (X) for any measure space! Definition 19. If V is a Banach space, a 1-parameter semigroup of operators U (t) : V → V is called a contraction semigroup if it is strongly continuous and kU (t)ψk ≤ kψk for all t ≥ 0 and all ψ ∈ V . Examples include strongly continuous 1-parameter unitary groups, but also Markov semigroups! Problem 18. Show any Markov semigroup is a contraction semigroup. The Hille–Yosida theorem generalizes Stone’s theorem to contraction semigroups: Theorem 20 (Hille–Yosida Theorem). Given a contraction semigroup U (t) we can always write U (t) = exp(tH) for some densely defined operator H such that H − λI has an inverse and k(H − λI)−1 ψk ≤ 1 kψk λ for all λ > 0 and ψ ∈ V . Conversely, any such operator determines a strongly continuous 1-parameter group. For all vectors ψ for which Hψ is well-defined, we have d = Hψ U (t)ψ dt t=0 Moreover, for any of these vectors, if we set ψ(t) = U (t)ψ we have d ψ(t) = Hψ(t) dt If you like, you can take the stuff at the end of this theorem to be what we mean by saying U (t) = exp(tH). When U (t) = exp(tH), we say that H generates the semigroup U (t). But now suppose V = L1 (X). Besides the conditions in the Hille–Yosida theorem, what extra conditions on H are necessary and sufficient for H to generate a Markov semigroup? In other words, what’s a definition of ‘infinitesimal stochastic operator’ that’s suitable not only when X is a finite set, but an arbitrary measure space? The best answer we know, not completely satisfactory, seems to be here: 111 11 11.7 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Answers [MR92] Zhi-Ming Ma and Michael Röckner, Introduction to the Theory of (NonSymmetric) Dirichlet Forms, Springer, Berlin, 1992. This book provides a very nice self-contained proof of the Hille–Yosida theorem. On the other hand, it does not answer the question in general, but only when the skew-symmetric part of H is dominated (in a certain sense) by the symmetric part. So, we’re stuck on this front, but that needn’t bring the whole project to a halt. We’ll just sidestep this question. For a good well-rounded introduction to Markov semigroups and what they’re good for, try: [RRTK] Ryszard Rudnicki, Katarzyna Pichòr and Marta Tyran-Kamı̀nska, Markov semigroups and their applications. 11.7 Answers Problem 15. Suppose X is a finite set. Show that any isometry U : L2 (X) → L2 (X) is invertible, and its inverse is again an isometry. Answer. Remember that U being an isometry means that it preserves the inner product: hU ψ, U φi = hψ, φi and thus it preserves the L2 norm kU ψk = kψk given by kψk = hψ, ψi1/2 . It follows that if U ψ = 0, then ψ = 0, so U is one-to-one. Since U is a linear operator from a finite-dimensional vector space to itself, U must therefore also be onto. Thus U is invertible, and because U preserves the inner product, so does its inverse: given ψ, φ ∈ L2 (X) we have hU −1 φ, U −1 ψi = hφ, ψi since we can write φ0 = U −1 φ, ψ 0 = U −1 ψ and then the above equation says hφ0 , ψ 0 i = hU φ0 , U ψ 0 i Problem 16. Suppose X is a finite set. Which stochastic operators U : L1 (X) → L1 (X) have stochastic inverses? Answer. Suppose the set X has n points. Then the set of stochastic states Z S = {ψ : X → R : ψ ≥ 0, ψ = 1} is a simplex. That is, it’s an equilateral triangle when n = 3, a regular tetrahedron when n = 4, and so on. 112 11 11.7 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Answers In general, S has n corners, which are the functions ψ that equal 1 at one point of S and zero elsewhere. Mathematically speaking, S is a convex set, and its corners are its extreme points: the points that can’t be written as convex combinations of other points of S in a nontrivial way. Any stochastic operator U must map S into itself, so if U has an inverse that’s also a stochastic operator, it must give a bijection U : S → S. Any linear transformation acting as a bijection between convex sets must map extreme points to extreme points (this is easy to check), so U must map corners to corners in a bijective way. This implies that it comes from a permutation of the points in X. In other words, any stochastic matrix with an inverse that’s also stochastic is a permutation matrix: a square matrix with every entry 0 except for a single 1 in each row and each column. For another proof, see the answer to Problem 23, which appears in Section 16.10. It is worth adding that there are lots of stochastic operators whose inverses are not, in general, stochastic. We can see this in at least two ways. First, for any measure space X, every stochastic operator U : L1 (X) → 1 L (X) that’s ‘close to the identity’ in this sense: kU − Ik < 1 (where the norm is the operator norm) will be invertible, simply because every operator obeying this inequality is invertible! After all, if this inequality holds, we have a convergent geometric series: U −1 = ∞ X 1 = (I − U )n I − (I − U ) n=0 Second, suppose X is a finite set and H is infinitesimal stochastic operator on L1 (X). Then H is bounded, so the stochastic operator exp(tH) where t ≥ 0 will always have an inverse, namely exp(−tH). But for t sufficiently small, this inverse exp(−tH) will only be stochastic if −H is infinitesimal stochastic, and that’s only true if H = 0. In something more like plain English: when you’ve got a finite set of states, you can formally run any Markov process backwards in time, but a lot of those ‘backwards-in-time’ operators will involve negative probabilities for the system to hop from one state to another! Problem 17. If X is a finite set, show that any operator on L1 (X) that’s both diagonal and infinitesimal stochastic must be zero. 113 11 11.7 QUANTUM MECHANICS VS STOCHASTIC MECHANICS Answers Answer. We are thinking of operators on L1 (X) as matrices with respect to the obvious basis of functions that equal 1 at one point and 0 elsewhere. If Hij is an infinitesimal stochastic matrix, the sum of the entries in each column is zero. If it’s diagonal, there’s at most one nonzero entry in each column. So, we must have H = 0. Problem 18. Show any Markov semigroup U (t) : L1 (X) → L1 (X) is a contraction semigroup. Answer. We need to show kU (t)ψk ≤ kψk for all t ≥ 0 and ψ ∈ L1 (X). Here the norm is the L1 norm, so more explicitly we need to show Z Z |U (t)ψ| ≤ |ψ| We can split ψ into its positive and negative parts: ψ = ψ+ − ψ− where ψ± ≥ 0 Since U (t) is stochastic we have U (t)ψ± ≥ 0 and Z Z U (t)ψ± = so R R |U (t)ψ| = R ≤ R = R = R = ψ± |U (t)ψ+ − U (t)ψ− | |U (t)ψ+ | + |U (t)ψ− | U (t)ψ+ + U (t)ψ− ψ+ + ψ− |ψ| 114 12 12 NOETHER’S THEOREM: QUANTUM VS STOCHASTIC Noether’s theorem: quantum vs stochastic In this chapter we merely want to show you the quantum and stochastic versions of Noether’s theorem, side by side. Having made our sacrificial offering to the math gods last time by explaining how everything generalizes when we replace our finite set X of states by an infinite set or an even more general measure space, we’ll now relax and state Noether’s theorem only for a finite set. If you’re the sort of person who finds that unsatisfactory, you can do the generalization yourself. 12.1 Two versions of Noether’s theorem Let us write the quantum and stochastic Noether’s theorem so they look almost the same: Theorem 21. Let X be a finite set. Suppose H is a self-adjoint operator on L2 (X), and let O be an observable. Then [O, H] = 0 if and only if for all states ψ(t) obeying Schrödinger’s equation d ψ(t) = −iHψ(t) dt the expected value of O in the state ψ(t) does not change with t. Theorem 22. Let X be a finite set. Suppose H is an infinitesimal stochastic operator on L1 (X), and let O be an observable. Then [O, H] = 0 if and only if for all states ψ(t) obeying the master equation d ψ(t) = Hψ(t) dt the expected values of O and O2 in the state ψ(t) do not change with t. This makes the big difference stick out like a sore thumb: in the quantum version we only need the expected value of O, while in the stochastic version we need the expected values of O and O2 ! We’ve already proved the stochastic version of Noether’s theorem. Now let’s do the quantum version. 12.2 Proof of the quantum version Our statement of the quantum version was silly in a couple of ways. First, we spoke of the Hilbert space L2 (X) for a finite set X, but any finite-dimensional 115 12 12.2 NOETHER’S THEOREM: QUANTUM VS STOCHASTIC Proof Hilbert space will do equally well. Second, we spoke of the “self-adjoint operator” H and the “observable” O, but in quantum mechanics an observable is the same thing as a self-adjoint operator! Why did we talk in such a silly way? Because we were attempting to emphasize the similarity between quantum mechanics and stochastic mechanics. But they’re somewhat different. For example, in stochastic mechanics we have two very different concepts: infinitesimal stochastic operators, which generate symmetries, and functions on our set X, which are observables. But in quantum mechanics something wonderful happens: self-adjoint operators both generate symmetries and are observables! So, our attempt was a bit strained. Let us state and prove a less silly quantum version of Noether’s theorem, which implies the one above: Theorem 23. Suppose H and O are self-adjoint operators on a finite-dimensional Hilbert space. Then [O, H] = 0 if and only if for all states ψ(t) obeying Schrödinger’s equation d ψ(t) = −iHψ(t) dt the expected value of O in the state ψ(t) does not change with t: d hψ(t), Oψ(t)i = 0 dt Proof. The trick is to compute the time derivative we just wrote down. Using Schrödinger’s equation, the product rule, and the fact that H is self-adjoint we get: d hψ(t), Oψ(t)i = dt h−iHψ(t), Oψ(t)i + hψ(t), O(−iHψ(t))i = ihψ(t), HOψ(t)i − ihψ(t), OHψ(t)i = −ihψ(t), [O, H]ψ(t)i So, if [O, H] = 0, clearly the above time derivative vanishes. Conversely, if this time derivative vanishes for all states ψ(t) obeying Schrödinger’s equation, we know hψ, [O, H]ψi = 0 for all states ψ and thus all vectors in our Hilbert space. Does this imply [O, H] = 0? Yes, because i times a commutator of a self-adjoint operator is self-adjoint, and for any self-adjoint operator A we have ∀ψ hψ, Aψi = 0 116 ⇒ A=0 12 12.3 NOETHER’S THEOREM: QUANTUM VS STOCHASTIC Comparison This is a well-known fact whose proof goes like this. Assume hψ, Aψi = 0 for all ψ. Then to show A = 0, it is enough to show hφ, Aψi = 0 for all φ and ψ. But we have a marvelous identity: hφ, Aψi = 1 4 (hψ + φ, A(ψ + φ)i − hψ − φ, A(ψ − φ)i +ihψ + iφ, A(ψ + iφ)i − ihψ − iφ, A(ψ − iφ)i) and all four terms on the right vanish by our assumption. The marvelous identity up there is called the polarization identity. In plain English, it says: if you know the diagonal entries of a self-adjoint matrix in every basis, you can figure out all the entries of that matrix in every basis. Why is it called the ‘polarization identity’ ? Presumably because it shows up in optics, in the study of polarized light. 12.3 Comparison In both the quantum and stochastic cases, the time derivative of the expected value of an observable O is expressed in terms of its commutator with the Hamiltonian. In the quantum case we have d hψ(t), Oψ(t)i = −ihψ(t), [O, H]ψ(t)i dt and for the right side to always vanish, we need [O, H] = 0, thanks to the polarization identity. In the stochastic case, a perfectly analogous equation holds: Z Z d Oψ(t) = [O, H]ψ(t) dt but now the right side can always vanish even without [O, H] = 0. We saw a counterexample in Section 10. There is nothing like the polarization identity to save us! To get [O, H] = 0 we need a supplementary hypothesis, for example the vanishing of Z d O2 ψ(t) dt Okay! Starting in the next Section we’ll change gears and look at some more examples of stochastic Petri nets and Markov processes, including some from chemistry. After some more of that, we’ll move on to networks of other sorts. There’s a really big picture here, and we’re afraid we’ve been getting caught up in the details of a tiny corner. 117 13 13 CHEMISTRY AND THE DESARGUES GRAPH Chemistry and the Desargues graph We’ve been doing a lot of hard work lately. Let’s take a break and think about a fun example from chemistry! 13.1 The ethyl cation Suppose you start with a molecule of ethane, which has 2 carbons and 6 hydrogens arranged like this: Then suppose you remove one hydrogen. The result is a positively charged ion, or ‘cation’. It would be cute if the opposite of a cation was called a ‘dogion’. Alas, it’s not. This particular cation, formed from removing one hydrogen from an ethane molecule, is called an ‘ethyl cation’. People used to think it looked like this: They also thought a hydrogen could hop from the carbon with 3 hydrogens attached to it to the carbon with 2. So, they drew a graph with a vertex for each way the hydrogens could be arranged, and an edge for each hop. It looks really cool: 118 13 13.1 CHEMISTRY AND THE DESARGUES GRAPH The ethyl cation The red vertices come from arrangements where the first carbon has 2 hydrogens attached to it, and the blue vertices come from those where the second carbon has 2 hydrogens attached to it. So, each edge goes between a red vertex and a blue vertex. This graph has 20 vertices, which are arrangements or ‘states’ of the ethyl cation. It has 30 edges, which are hops or ‘transitions’. Let’s see why those numbers are right. First we need to explain the rules of the game. The rules say that the 2 carbon atoms are distinguishable: there’s a ‘first’ one and a ‘second’ one. The 5 hydrogen atoms are also distinguishable. But, all we care about is which carbon atom each hydrogen is bonded to: we don’t care about further details of its location. And we require that 2 of the hydrogens are bonded to one carbon, and 3 to the other. If you’re a physicist, you may wonder why the rules work this way: after all, at a fundamental level, identical particles aren’t really distinguishable. We’re afraid we can’t give a fully convincing explanation right now: we’re just reporting the rules! Given these rules, there are 2 choices of which carbon has two hydrogens attached to it. Then there are 5 5×4 = = 10 2 2×1 choices of which two hydrogens are attached to it. This gives a total of 2 × 10 = 20 states. These are the vertices of our graph: 10 red and 10 blue. The edges of the graph are transitions between states. Any hydrogen in the group of 3 can hop over to the group of 2. There are 3 choices for which 119 13 13.2 CHEMISTRY AND THE DESARGUES GRAPH The Desargues graph hydrogen atom makes the jump. So, starting from any vertex in the graph there are 3 edges. This means there are 3 × 20/2 = 30 edges. Why divide by 2? Because each edge touches 2 vertices. We have to avoid double-counting them. 13.2 The Desargues graph The idea of using this graph in chemistry goes back to this paper: [BFB66] A. T. Balaban, D. Fǎrcaşiu, and R. Bǎnicǎ, Graphs of multiple 1,2-shifts in carbonium ions and related systems, Rev. Roum. Chim. 11 (1966), 1205. This paper is famous because it was the first to use graphs in chemistry to describe molecular transitions, as opposed to using them as pictures of molecules! But this particular graph was already famous for other reasons. It’s called the Desargues–Levi graph, or Desargues graph for short: • Desargues graph, Wikipedia. Later we’ll say why it’s called this. There are lots of nice ways to draw the Desargues graph. For example: The reason why we can draw such pretty pictures is that the Desargues graph is very symmetrical. Clearly any permutation of the 5 hydrogens acts as a symmetry of the graph, and so does any permutation of the 2 carbons. This 120 13 13.3 CHEMISTRY AND THE DESARGUES GRAPH The ethyl cation, revisited gives a symmetry group S5 × S2 , which has 5! × 2! = 240 elements. And in fact this turns out to be the full symmetry group of the Desargues graph. The Desargues graph, its symmetry group, and its applications to chemistry are discussed here: [Ran97] Milan Randic, Symmetry properties of graphs of interest in chemistry: II: Desargues–Levi graph, Int. Jour. Quantum Chem. 15 (1997), 663–682. 13.3 The ethyl cation, revisited We can try to describe the ethyl cation using probability theory. If at any moment its state corresponds to some vertex of the Desargues graph, and it hops randomly along edges as time goes by, it will trace out a random walk on the Desargues graph. This is a nice example of a Markov process! We could also try to describe the ethyl cation using quantum mechanics. Then, instead of having a probability of hopping along an edge, it has an amplitude of doing so. But as we’ve seen, a lot of similar math will still apply. It should be fun to compare the two approaches. But we bet you’re wondering which approach is correct. This is a somewhat tricky question. The answer would seem to depend on how much the ethyl cation is interacting with its environment—for example, bouncing off other molecules. When a system is interacting a lot with its environment, a probabilistic approach seems to be more appropriate. The relevant buzzword is ‘environmentally induced decoherence’. However, there’s something much more basic we have tell you about. After the paper by Balaban, Farcasiu, and Banic came out, people gradually realized that the ethyl cation doesn’t really look like the drawing we showed you! It’s what chemists call ‘nonclassical’ ion. What they mean is this: its actual structure is not what you get by taking the traditional ball-and-stick model of an ethane molecule and ripping off a hydrogen. The ethyl cation really looks like this: This picture comes from here: • Stephen Bacharach, Ethyl cation, Computational Organic Chemistry. If you go to this website, you’ll find more details, and pictures that you can actually rotate. 121 13 13.4 13.4 CHEMISTRY AND THE DESARGUES GRAPH Trigonal bipyramidal molecules Trigonal bipyramidal molecules In short: if we stubbornly insist on applying the Desargues graph to realistic chemistry, the ethyl cation won’t do: we need to find some other molecule to apply it to. Luckily, there are lots of options! They’re called trigonal bipyramidal molecules. They look like this: The 5 balls on the outside are called ‘ligands’: they could be atoms or bunches of atoms. In chemistry, ‘ligand’ just means something that’s stuck onto a central thing. For example, in phosphorus pentachloride the ligands are chlorine atoms, all attached to a central phosphorus atom: It’s a colorless solid, but as you might expect, it’s pretty nasty stuff: it’s not flammable, but it reacts with water or heat to produce toxic chemicals like hydrogen chloride. Another example is iron pentacarbonyl, where 5 carbon-oxygen ligands are attached to a central iron atom: 122 13 13.4 CHEMISTRY AND THE DESARGUES GRAPH Trigonal bipyramidal molecules You can make this stuff by letting powdered iron react with carbon monoxide. It’s a straw-colored liquid with a pungent smell! Whenever you’ve got a molecule of this shape, the ligands come in two kinds. There are the 2 ‘axial’ ones, and the 3 ‘equatorial’ ones: The molecule has 20 states... but only if count the states a certain way. We have to treat all 5 ligands as distinguishable, but think of two arrangements of them as the same if we can rotate one to get the other. The trigonal bipyramid has a rotational symmetry group with 6 elements. So, there are 5! / 6 = 20 states. The transitions between states are devilishly tricky. They’re called pseudorotations, and they look like this: If you look very carefully, you’ll see what’s going on. First the 2 axial ligands move towards each other to become equatorial. Now the equatorial ones are no longer in the horizontal plane: they’re in the plane facing us! Then 2 of the 3 123 13 13.5 CHEMISTRY AND THE DESARGUES GRAPH Drawing the Desargues graph equatorial ones swing out to become axial. This fancy dance is called the Berry pseudorotation mechanism. To get from one state to another this way, we have to pick 2 of the 3 equatorial ligands to swing out and become axial. There are 3 choices here. So, if we draw a graph with states as vertices and transitions as edges, it will have 20 vertices and 20 × 3/2 = 30 edges. That sounds suspiciously like the Desargues graph! Problem 19. Show that the graph with states of a trigonal bipyramidal molecule as vertices and pseudorotations as edges is indeed the Desargues graph. It seems this fact was first noticed here: [LR68] Paul C. Lauterbur and Fausto Ramirez, Pseudorotation in trigonal-bipyramidal molecules, J. Am. Chem. Soc. 90 (1968), 6722.6726. In the next section we’ll say more about the Markov process or quantum process corresponding to a random walk on the Desargues graph. But since the Berry pseudorotation mechanism is so hard to visualize, we’ll pretend that the ethyl cation looks like this: and we’ll use this picture to help us think about the Desargues graph. That’s okay: everything we’ll figure out can easily be translated to apply to the real-world situation of a trigonal bipyramidal molecule. The virtue of math is that when two situations are ‘mathematically the same’, or ‘isomorphic’, we can talk about either one, and the results automatically apply to the other. This is true even if the one we talk about doesn’t actually exist in the real world! 13.5 Drawing the Desargues graph Before we quit, let’s think a bit more about how we draw the Desargues graph. For this it’s probably easiest to go back to our naive model of an ethyl cation: 124 13 13.5 CHEMISTRY AND THE DESARGUES GRAPH Drawing the Desargues graph Even though ethyl cations don’t really look like this, and we should be talking about some trigonal bipyramidal molecule instead, it won’t affect the math to come. Mathematically, the two problems are isomorphic! So let’s stick with this nice simple picture. We can be a bit more abstract, though. A state of the ethyl cation is like having 5 balls, with 3 in one pile and 2 in the other. And we can focus on the first pile and forget the second, because whatever isn’t in the first pile must be in the second. Of course a mathematician calls a pile of things a ‘set’, and calls those things ‘elements’. So let’s say we’ve got a set with 5 elements. Draw a red dot for each 2-element subset, and a blue dot for each 3-element subset. Draw an edge between a red dot and a blue dot whenever the 2-element subset is contained in the 3-element subset. We get the Desargues graph. That’s true by definition. But we never proved that this graph looks like this: 125 13 13.5 CHEMISTRY AND THE DESARGUES GRAPH Drawing the Desargues graph We won’t now, either. We’ll just leave it as a problem: Problem 20. If we define the Desargues graph to have vertices corresponding to 2- and 3-element subsets of a 5-element set, with an edge between vertices when one subset is contained in another, why does it look like the picture above? To draw a picture we know is correct, it’s actually easier to start with a big graph that has vertices for all the subsets of our 5-element set. If we draw an edge whenever an n-element subset is contained in an (n + 1)-element subset, the Desargues graph will be sitting inside this big graph. Here’s what the big graph looks like: This graph has 25 vertices. It’s actually a picture of a 5-dimensional hypercube! Unfortunately, to make the picture pretty, some vertices are hidden behind others: each orange vertex has another behind it. The vertices are arranged in rows. There’s • one 0-element subset, • five 1-element subsets, • ten 2-element subsets, • ten 3-element subsets, • five 4-element subsets, • one 5-element subset. So, the numbers of vertices in each row should go like this: 1 5 10 10 126 5 1 13 13.6 CHEMISTRY AND THE DESARGUES GRAPH Desargues’ theorem which is a row in Pascal’s triangle. We get the Desargues graph if we keep only the vertices corresponding to 2- and 3-element subsets, like this: However, each orange vertex has another hiding behind it, so we only see 7 vertices in each row here, not 10. Laurens Gunnarsen fixed this, and got the following picture of the Desargues graph: It’s less pretty than our earlier picture of the Desargues graph, but at least there’s no mystery to it. Also, it shows that the Desargues graph can be generalized in various ways. For example, there’s a theory of bipartite Kneser graphs H(n, k). The Desargues graph is H(5, 2). 13.6 Desargues’ theorem Finally, we can’t resist answering this question: why is this graph called the ‘Desargues graph’ ? This name comes from Desargues’ theorem, a famous result in projective geometry. Suppose you have two triangles ABC and abc, like this: 127 13 13.7 CHEMISTRY AND THE DESARGUES GRAPH Answers Suppose the lines Aa, Bb, and Cc all meet at a single point, the ‘center of perspectivity’. Then the point of intersection of ab and AB, the point of intersection of ac and AC, and the point of intersection of bc and BC all lie on a single line, the ‘axis of perspectivity’. The converse is true too. Quite amazing! Definition 24. The Desargues configuration consists of all the actors in this drama: • 10 points: A, B, C, a, b, c, the center of perspectivity, and the three points on the axis of perspectivity and • 10 lines: Aa, Bb, Cc, AB, AC, BC, ab, ac, bc and the axis of perspectivity Given any configuration of points and lines, we can form a graph called its Levi graph by drawing a vertex for each point or line, and drawing edges to indicate which points lie on which lines. And now for the punchline: the Levi graph of the Desargues configuration is the ‘Desargues–Levi graph’ !—or Desargues graph, for short. Alas, we don’t know how this is relevant to anything we’ve discussed. For now it’s just a tantalizing curiosity. 13.7 Answers We didn’t get any solutions of Problem 19 the first time we posted this on the Azimuth blog, so we reposed it on Azimuth and the n-Category Café. That brought out lots of great solutions, which we’ve used to compose the answer here. Special thanks go especially to J. M. Allen, Twan van Laarhoven, Peter Morgan and Johan Swanljung. Problem 19. Show that the graph with states of a trigonal bipyramidal molecule as vertices and pseudorotations as edges is indeed the Desargues graph. Answer. To be specific, let’s use iron pentacarbonyl as our example of a trigonal bipyramidal molecule: 128 13 13.7 CHEMISTRY AND THE DESARGUES GRAPH Answers It suffices to construct a 1-1 correspondence between the states of this molecule and those of the ethyl cation, such that two states of this molecule are connected by a transition if and only if the same holds for the corresponding states of the ethyl cation. Here’s the key idea: the ethyl cation has 5 hydrogens, with 2 attached to one carbon and 3 attached to the other. Similarly, the trigonal bipyramidal molecule has 5 carbonyl grops, with 2 axial and 3 equatorial. We’ll use this resemblance to set up our correspondence. There are various ways to describe states of the ethyl cation, but this is the best for us. Number the hydrogens 1,2,3,4,5. Then a state of the ethyl cation consists of a partition of the set {1, 2, 3, 4, 5} into a 2-element set and a 3-element set, together with one extra bit of information, saying which carbon has 2 hydrogens attached to it. This extra bit is the color here: What do transitions look like in this description? When a transition occurs, 129 13 13.7 CHEMISTRY AND THE DESARGUES GRAPH Answers two hydrogens that belonged to the 3-element set now become part of the 2element set. Meanwhile, both hydrogens that belonged to the 2-element set now become part of the 3-element set. (Ironically, the one hydrogen that hops is the one that stays in the 3-element set.) Moreover, the extra bit of information changes. That’s why every edge goes from a red dot to a blue one, or vice versa. So, to solve the problem, we need to show that the same description also works for the states and transitions of iron pentacarbonyl! In other words, we need to describe its states as ways of partitioning the set 1,2,3,4,5 into a 2-element set and a 3-element set, together with one extra bit of information. And we need its transitions to switch two elements of the 2-element set with two of the 3-element set, while changing that extra bit. To do this, number the carbonyl groups 1,2,3,4,5. The 2-element set consists of the axial ones, while the 3-element set consists of the equatorial ones. When a transition occurs, two of axial ones trade places with two of the equatorial ones, like this: So, now we just need to figure out what that extra bit of information is, and why it always changes when a transition occurs. Here’s how we calculate that extra bit. Hold the iron pentacarbonyl molecule vertically with one of the equatorial carbonyl groups pointing to your left. Remember, the carbonyl groups are numbered. So, write a list of these numbers, say (a,b,c,d,e), where a is the top axial one, b,c,d are the equatorial ones listed in counterclockwise order starting from the one pointing left, and e is the bottom axial one. This list is some permutation of the list (1,2,3,4,5). Take the sign of this permutation to be our bit! Let’s do an example: Here we get the list (2,5,3,1,4) since 2 is on top, 4 is on bottom, and 5,3,1 are the equatorial guys listed counterclockwise starting from the one at left. The 130 13 13.7 CHEMISTRY AND THE DESARGUES GRAPH Answers list (2,5,3,1,4) is an odd permutation of (1,2,3,4,5), so our bit of information is odd . Of course we must check that this bit is well-defined: namely, that it doesn’t change if we rotate the molecule. Rotating it a third of a turn gives an even permutation of the equatorial guys and leaves the axial ones alone, so this is an even permutation. Flipping it over gives an odd permutation of the equatorial guys, but it also gives an odd permutation of the axial ones, so this too is an even permutation. So, rotating the molecule doesn’t change the sign of the permutation we compute from it. The sign is thus a well-defined function of the state of the molecule. Next we must to check that this sign changes whenever our molecule undergoes a transition. For this we need to check that any transition changes our list of numbers by an odd permutation. Since all transitions are conjugate in the permutation group, it suffices to consider one example: Here we started with a state giving the list (2,5,3,1,4). The transition ttakes us to a state that gives the list (3,5,4,2,1) if we hold the molecule so that 3 is pointing up and 5 to the left. The reader can check that going from one list to another requires an odd permutation. So we’re done. Tracy Hall’s discussion of the above problem was very interesting and erudite. He also addressed an interesting subsidiary problem, namely: what graph do we get if we discard that extra bit of information that says which carbon in the ethyl cation has 3 hydrogens attached to it and which has 2? The answer is the Petersen graph: Hall wrote: As some comments have pointed out over on Azimuth, in both cases there are ten underlying states which simply pick out two of the five 131 13 13.7 CHEMISTRY AND THE DESARGUES GRAPH Answers pendant atoms as special, together with an extra parity bit (which can take either value for any of the ten states), giving twenty states in total. The correspondence of the ten states is clear: an edge exists between state A and state B, in either case, if and only if the two special atoms of state A are disjoint from the two special atoms of state B. This is precisely one definition of the Petersen graph (a famous 3-valent graph on 10 vertices that shows up as a small counterexample to lots of naove conjectures). Thus the graph in either case is a double cover of the Petersen graph—but that does not uniquely specify it, since, for example, both the Desargues graph and the dodecahedron graph are double covers of the Petersen graph. For a labeled graph, each double cover corresponds uniquely to an element of the Z/2Z cohomology of the graph (for an unlabeled graph, some of the double covers defined in this way may turn out to be isomorphic). Cohomology over Z/2Z takes any cycle as input and returns either 0 or 1, in a consistent way (the output of a Z/2Z sum of cycles is the sum of the outputs on each cycle). The double cover has two copies of everything in the base (Petersen) graph, and as you follow all the way around a cycle in the base, the element of cohomology tells you whether you come back to the same copy (for 0) or the other copy (for 1) in the double cover, compared to where you started. One well-defined double cover for any graph is the one which simply switches copies for every single edge (this corresponding to the element of cohomology which is 1 on all odd cycles and 0 on all even cycles). This always gives a double cover which is a bipartite graph, and which is connected if and only if the base graph is connected and not bipartite. So if we can show that in both cases (the fictitious ethyl cation and phosphorus pentachloride) the extra parity bit can be defined in such a way that it switches on every transition, that will show that we get the Desargues graph in both cases. The fictitious ethyl cation is easy: the parity bit records which carbon is which, so we can define it as saying which carbon has three neighbors. This switches on every transition, so we are done. Phosphorus pentachloride is a bit trickier; the parity bit distinguishes a labeled molecule from its mirror image, or enantiomer. As has already been pointed out on both sites, we can use the parity of a permutation to distinguish this, since it happens that the orientationpreserving rotations of the molecule, generated by a three-fold rotation acting as a three-cycle and by a two-fold rotation acting as a pair of two-cycles, are all even permutations, while the mirror image that switches only the two special atoms is an odd permutation. The pseudorotation can be followed by a quarter turn to return the five chlorine atoms to the five places previously occupied by chlorine atoms, which makes it act as a four-cycle, an odd permutation. 132 13 13.7 CHEMISTRY AND THE DESARGUES GRAPH Answers Since the parity bit in this case also can be defined in such a way that it switches on every transition, the particular double cover in each case is the Desargues graph—a graph one might be surprised to come across here. The five chlorine atoms in phosphorus pentachloride lie in six triangles which give a triangulation of the 2-sphere, and another way of thinking of the pseudorotation is that it corresponds to a Pachner move or bistellar flip on this triangulation—in particular, any bistellar flip on this triangulation that preserves the number of triangles and the property that all vertices in the triangulation have degree at least three corresponds to a pseudorotation as described. The next problem was both posed and answered by Greg Egan on Azimuth. Problem 20. If we define the Desargues graph to have vertices corresponding to 2- and 3-element subsets of a 5-element set, with an edge between vertices when one subset is contained in another, why does it look like this picture? Answer. For i = 0, ..., 4, and with all the additions below modulo 5, define five pairs of red dots as: {i, i + 1}, {i + 1, i + 4} and five pairs of blue dots as: {i, i + 1, i + 4}, {i + 1, i + 2, i + 4}. 133 13 13.7 CHEMISTRY AND THE DESARGUES GRAPH Answers The union of the ith pair of red dots is the first of the ith pair of blue dots, and the union of the second of the ith pair of red dots and the first of the (i + 1)th pair of red dots is the second of the ith pair of blue dots. So if we form a 20-sided polygon whose vertices are alternating red and blue dots generated in this order, all the edges of the polygon will join red dots to blue dots of which they are subsets: The first red dot of the ith pair is also a subset of the first blue dot of the (i + 1)th pair: {i + 1, i + 2, i} which gives the five short chords in the picture, while the second red dot of the ith pair is a subset of the second blue dot of the (i + 2) pair: {i + 3, i + 4, i + 1} which gives the five long chords in the picture. 134 14 14 GRAPH LAPLACIANS Graph Laplacians In Section 13 we saw how to get a graph whose vertices are states of a molecule and whose edges are transitions between states. We focused on two beautiful but not completely realistic examples that both give rise to the same highly symmetrical graph: the ‘Desargues graph’. In this section we’ll consider how a molecule can carry out a random walk on this graph. Then we’ll get to work showing how any graph gives: • A Markov process, namely a random walk on the graph. • A quantum process, where instead of having a probability to hop from vertex to vertex as time passes, we have an amplitude. The trick is to use an operator called the ‘graph Laplacian’, a discretized version of the Laplacian which happens to be both infinitesimal stochastic and self-adjoint. By our results in in Section 11, such an operator will give rise both to a Markov process and a quantum process (that is, a one-parameter unitary group). The most famous operator that’s both infinitesimal stochastic and selfadjoint is the Laplacian, ∇2 . Because it’s both, the Laplacian shows up in two important equations: one in stochastic mechanics, the other in quantum mechanics. • The heat equation: d ψ = ∇2 ψ dt describes how the probability ψ(x) of a particle being at the point x smears out as the particle randomly walks around. The corresponding Markov process is called ‘Brownian motion’. 135 14 14.1 GRAPH LAPLACIANS A random walk on the Desargues graph • The Schrödinger equation: d ψ = −i∇2 ψ dt describes how the amplitude ψ(x) of a particle being at the point x wiggles about as the particle ‘quantumly’ walks around. Both these equations have analogues where we replace space by a graph, and now we’ll describe them. 14.1 A random walk on the Desargues graph Back to business! We’ve been telling you about the analogy between quantum mechanics and stochastic mechanics. This analogy becomes especially interesting in chemistry, which lies on the uneasy borderline between quantum and stochastic mechanics. Fundamentally, of course, atoms and molecules are described by quantum mechanics. But sometimes chemists describe chemical reactions using stochastic mechanics instead. When can they get away with this? Apparently whenever the molecules involved are big enough and interacting with their environment enough for ‘decoherence’ to kick in. We won’t attempt to explain this now. Let’s imagine we have a molecule of iron pentacarbonyl with—here’s the unrealistic part, but it’s not really too bad—distinguishable carbonyl groups: 136 14 14.1 GRAPH LAPLACIANS A random walk on the Desargues graph Iron pentacarbonyl is liquid at room temperatures, so as time passes, each molecule will bounce around and occasionally do a maneuver called a ‘pseudorotation’: We can approximately describe this process as a random walk on a graph whose vertices are states of our molecule, and whose edges are transitions between states—namely, pseudorotations. And as we saw in Section 13, this graph is the Desargues graph: 137 14 14.1 GRAPH LAPLACIANS A random walk on the Desargues graph Note: all the transitions are reversible here. And thanks to the enormous amount of symmetry, the rates of all these transitions must be equal. Let’s write V for the set of vertices of the Desargues graph. A probability distribution of states of our molecule is a function ψ : V → [0, ∞) with X ψ(x) = 1 x∈V We can think of these probability distributions as living in this vector space: L1 (V ) = {ψ : V → R} We’re calling this space L1 because of the general abstract nonsense explained in Section 11: probability distributions on any measure space live in a vector space called L1 . Right now that notation is overkill, since every function on V lies in L1 . But please humor us. The point is that we’ve got a general setup that applies here. There’s a Hamiltonian: H : L1 (V ) → L1 (V ) describing the rate at which the molecule randomly hops from one state to another... and the probability distribution ψ ∈ L1 (X) evolves in time according to the equation: d ψ(t) = Hψ(t) dt 138 14 14.2 GRAPH LAPLACIANS Graph Laplacians But what’s the Hamiltonian H? It’s very simple, because it’s equally likely for the state to hop from any vertex to any other vertex that’s connected to that one by an edge. Why? Because the problem has so much symmetry that nothing else makes sense. So, let’s write E for the set of edges of the Desargues graph. We can think of this as a subset of V × V by saying (x, y) ∈ E when x is connected to y by an edge. Then X (Hψ)(x) = ψ(y) − 3ψ(x) y:(x,y)∈E We’re subtracting 3ψ(x) because there are 3 edges coming out of each vertex x, so this is the rate at which the probability of staying at x decreases. We could multiply this Hamiltonian by a constant if we wanted the random walk to happen faster or slower... but let’s not. The next step is to solve this discretized version of the heat equation: d ψ(t) = Hψ(t) dt Abstractly, the solution is easy: ψ(t) = exp(tH)ψ(0) But to actually compute exp(tH), we might want to diagonalize the operator H. In this particular example, we can do this by taking advantage of the enormous symmetry of the Desargues graph. But let’s not do this just yet. First let’s draw some general lessons from this example. 14.2 Graph Laplacians The Hamiltonian we just saw is an example of a ‘graph Laplacian’. We can write down such a Hamiltonian for any graph, but it gets a tiny bit more complicated when different vertices have different numbers of edges coming out of them. The word ‘graph’ means lots of things, but right now we’re talking about simple graphs. Such a graph has a set of vertices V and a set of edges E ⊆ V × V, such that (x, y) ∈ E ⇒ (y, x) ∈ E which says the edges are undirected, and (x, x) ∈ /E which says there are no loops. Let d(x) be the degree of the vertex x, meaning the number of edges coming out of it. Then the graph Laplacian is this operator on L1 (V ): X (Hψ)(x) = ψ(y) − d(x)ψ(x) y:(x,y)∈E There is a huge amount to say about graph Laplacians! If you want, you can get started here: 139 14 14.2 GRAPH LAPLACIANS Graph Laplacians [New] Michael William Newman, The Laplacian Spectrum of Graphs, Masters Thesis, Department of Mathematics, University of Manitoba, 2000. But for now, let’s just say that exp(tH) is a Markov process describing a random walk on the graph, where hopping from one vertex to any neighboring vertex has unit probability per unit time. We can make the hopping faster or slower by multiplying H by a constant. And here is a good time to admit that most people use a graph Laplacian that’s the negative of ours, and write time evolution as exp(−tH). The advantage is that then the eigenvalues of the Laplacian are ≥ 0. But what matters most is this. We can write the operator H as a matrix whose entry Hxy is 1 when there’s an edge from x to y and 0 otherwise, except when x = y, in which case the entry is −d(x). And then: Problem 21. Show that for any finite graph, the graph Laplacian H is infinitesimal stochastic, meaning that: X Hxy = 0 x∈V and x 6= y ⇒ Hxy ≥ 0 This fact implies that for any t ≥ 0, the operator exp(tH) is stochastic—just what we need for a Markov process. But we could also use H as a Hamiltonian for a quantum system, if we wanted. Now we think of ψ(x) as the amplitude for being in the state x ∈ V . But now ψ is a function ψ: V → C with X |ψ(x)|2 = 1 x∈V We can think of this function as living in the Hilbert space L2 (V ) = {ψ : V → C} where the inner product is hφ, ψi = X φ(x)ψ(x) x∈V Problem 22. Show that for any finite graph, the graph Laplacian H : L2 (V ) → L2 (V ) is self-adjoint, meaning that: Hxy = H yx 140 14 14.3 GRAPH LAPLACIANS The Laplacian of the Desargues graph This implies that for any t ∈ R, the operator exp(−itH) is unitary—just what we need for one-parameter unitary group. So, we can take this version of Schrödinger’s equation: d ψ = −iHψ dt and solve it: ψ(t) = exp(−itH)ψ(0) and we’ll know that time evolution is unitary! So, we’re in a dream world where we can do stochastic mechanics and quantum mechanics with the same Hamiltonian. We’d like to exploit this somehow, but we’re not quite sure how. Of course physicists like to use a trick called Wick rotation where they turn quantum mechanics into stochastic mechanics by replacing time by imaginary time. We can do that here. But we’d like to do something new, special to this context. You might want to learn more about the relation between chemistry and graph theory. Of course, graphs show up in at least two ways: first for drawing molecules, and second for drawing states and transitions, as we’ve been doing. These books seem to be good: [BR91] Danail Bonchev and D. H. Rouvray, eds., Chemical Graph Theory: Introduction and Fundamentals, Taylor and Francis, 1991. [Tri92] Nenad Trinajstic, Chemical Graph Theory, CRC Press, 1992. [Kin93] R. Bruce King, Applications of Graph Theory and Topology in Inorganic Cluster Coordination Chemistry, CRC Press, 1993. 14.3 The Laplacian of the Desargues graph Greg Egan figured out a nice way to explicitly describe the eigenvectors of the Laplacian H for the Desargues graph. This lets us explicitly solve the heat equation d ψ = Hψ dt and also the Schrödinger equation on this graph. First there is an obvious eigenvector: any constant function. Indeed, for any finite graph, any constant function ψ is an eigenvector of the graph Laplacian with eigenvalue zero: Hψ = 0 so that exp(tH)ψ = ψ This reflects the fact that the evenly smeared probability distribution is an equilibrium state for the random walk described by heat equation. For a connected graph this will be the only equilibrium state. For a graph with several connected 141 14 14.3 GRAPH LAPLACIANS The Laplacian of the Desargues graph components, there will be an equilibrium state for each connected component, which equals some positive constant on that component and zero elsewhere. For the Desargues graph, or indeed any connected graph, all the other eigenvectors will be functions that undergo exponential decay at various rates: Hψ = λψ with λ < 0, so that exp(tH)ψ = exp(λt)ψ decays as t increases. But since probability is conserved, any vector that undergoes exponential decay must have terms that sum to zero. So a vector like this won’t be a stochastic state, but rather a deviation from equilibrium. We can write any stochastic state as the equilibrium state plus a sum of terms that decay with different rate constants. If you put a value of 1 on every red dot and -1 on every blue dot, you get an eigenvector of the Hamiltonian with eigenvalue -6: -3 for the diagonal entry for each vertex, and -3 for the sum over the neighbors. By trial and error, it’s not too hard to find examples of variations on this where some vertices have a value of zero. Every vertex with zero either has its neighbors all zero, or two neighbors of opposite signs: • Eigenvalue -5: every non-zero vertex has two neighbors with the opposite value to it, and one neighbor of zero. • Eigenvalue -4: every non-zero vertex has one neighbor with the opposite value to it, and two neighbors of zero. • Eigenvalue -2: every non-zero vertex has one neighbor with the same value, and two neighbors of zero. • Eigenvalue -1: every non-zero vertex has two neighbors with the same value, and one neighbor of zero. So the general pattern is what you’d expect: the more neighbors of equal value each non-zero vertex has, the more slowly that term will decay. In more detail, the eigenvectors are given by the the following functions, together with functions obtained by rotating these pictures. Unlabelled dots are labelled by zero: 142 14 14.3 The Laplacian of the Desargues graph 143 GRAPH LAPLACIANS 14 14.3 GRAPH LAPLACIANS The Laplacian of the Desargues graph In fact you can span the eigenspaces by rotating the particular eigenvectors we drew by successive multiples of 1/5 of a rotation, i.e. 4 dots. If you do this for the examples with eigenvalues -2 and -4, it’s not too hard to see that all five rotated diagrams are linearly independent. If you do it for the examples with eigenvalues -1 and -5, four rotated diagrams are linearly independent. So along with the two 1-dimensional spaces, that’s enough to prove that you’ve exhausted the whole 20-dimensional space. 144 15 15 DIRICHLET OPERATORS Dirichlet operators and electrical circuits We’ve been comparing two theories: stochastic mechanics and quantum mechanics. Last time we saw that any graph gives us an example of both theories! It’s a bit peculiar, but now we’ll explore the intersection of these theories a little further, and see that it has another interpretation. It’s also the theory of electrical circuits made of resistors! That’s nice, because we’re supposed to be talking about ‘network theory’, and electrical circuits are perhaps the most practical networks of all: 15.1 Dirichlet operators and electrical circuits Last time we saw that any graph gives us an operator called the ‘graph Laplacian’ that’s both infinitesimal stochastic and self-adjoint. That means we get both: • a Markov process describing the random walk of a classical particle on the graph. and • a 1-parameter unitary group describing the motion of a quantum particle on the graph. That’s sort of neat, so it’s natural to wonder what are all the operators that are both infinitesimal stochastic and self-adjoint. They’re called ‘Dirichlet operators’, and at least in the finite-dimensional case we’re considering, they’re easy to completely understand. Even better, it turns out they describe electrical circuits made of resistors! Let’s take a lowbrow attitude and think of a linear operator H : Cn → Cn as an n × n matrix with entries Hij . Then: • H is self-adjoint if it equals the conjugate of its transpose: Hij = H ji 145 15 15.2 DIRICHLET OPERATORS Circuits made of resistors • H is infinitesimal stochastic if its columns sum to zero and its offdiagonal entries are nonnegative: X Hij = 0 i i 6= j ⇒ Hij ≥ 0 • H is a Dirichlet operator if it’s both self-adjoint and infinitesimal stochastic. What are Dirichlet operators like? Suppose H is a Dirichlet operator. Then its off-diagonal entries are ≥ 0, and since X Hij = 0 i its diagonal entries obey Hii = − X Hij ≤ 0 i6=j So all the entries of the matrix H are real, which in turn implies it’s symmetric: Hij = H ji = Hji So, we can build any Dirichlet operator H as follows: • Choose the entries above the diagonal, Hij with i < j, to be arbitrary nonnegative real numbers. • The entries below the diagonal, Hij with i > j, are then forced on us by the requirement that H be symmetric: Hij = Hji . • The diagonal entries are then P forced on us by the requirement that the columns sum to zero: Hii = − i6=j Hij . Note that because the entries are real, we can think of a Dirichlet operator as a linear operator H : Rn → Rn . We’ll do that for the rest of this section. 15.2 Circuits made of resistors Now for the fun part. We can easily draw any Dirichlet operator! To this we draw n dots, connect each pair of distinct dots with an edge, and label the edge connecting the ith dot to the jth with any number Hij ≥ 0. 146 15 15.2 DIRICHLET OPERATORS Circuits made of resistors This contains all the information we need to build our Dirichlet operator. To make the picture prettier, we can leave out the edges labelled by 0: Like last time, the graphs we’re talking about are simple: undirected, with no edges from a vertex to itself, and at most one edge from one vertex to another. So: Theorem 25. Any finite simple graph with edges labelled by positive numbers gives a Dirichlet operator, and conversely. We already talked about a special case last time: if we label all the edges by the number 1, our operator H is called the graph Laplacian. So, now we’re generalizing that idea by letting the edges have more interesting labels. What’s the meaning of this trick? Well, we can think of our graph as an electrical circuit where the edges are wires. What do the numbers labelling these wires mean? One obvious possibility is to put a resistor on each wire, and let that number be its resistance. But that doesn’t make sense, since we’re leaving out wires labelled by 0. If we leave out a wire, that’s not like having a wire of zero resistance: it’s like having a wire of infinite resistance! No current can go through when there’s no wire. So the number labelling an edge should be the conductance of the resistor on that wire. Conductance is the reciprocal of resistance. So, our Dirichlet operator above gives a circuit like this: 147 15 15.2 DIRICHLET OPERATORS Circuits made of resistors Here Ω is the symbol for an ‘ohm’, a unit of resistance... but the upside-down version, namely f, is the symbol for a ‘mho’, a unit of conductance that’s the reciprocal of an ohm. (Actually the term ‘siemens’ is officially preferred over ‘mho’, but we like the word ‘mho’.) Let’s see if this cute idea leads anywhere. Think of a Dirichlet operator H : Rn → Rn as a circuit made of resistors. What could a vector ψ ∈ Rn mean? It assigns a real number to each vertex of our graph. The only sensible option is for this number to be the electric potential at that point in our circuit. So let’s try that. Now, what’s hψ, Hψi? In quantum mechanics this would be a very sensible thing to look at: it would be gives us the expected value of the Hamiltonian H in a state ψ. But what does it mean in the land of electrical circuits? Up to a constant factor, it turns out to be the power consumed by the electrical circuit! Let’s see why. First, remember that when a current flows along a wire, power gets consumed. In other words, electrostatic potential energy gets turned into heat. The power consumed is P =VI where V is the voltage across the wire and I is the current flowing along the wire. If we assume our wire has resistance R we also have Ohm’s law: I = V /R so V2 R If we write this using the conductance C = 1/R instead of the resistance R, we get P = CV 2 P = 148 15 15.2 DIRICHLET OPERATORS Circuits made of resistors But our electrical circuit has lots of wires, so the power it consumes will be a sum of terms like this. We’re assuming Hij is the conductance of the wire from the ith vertex to the jth, or zero if there’s no wire connecting them. And by definition, the voltage across this wire is the difference in electrostatic potentials at the two ends: ψi − ψj . So, the total power consumed is X P = Hij (ψi − ψj )2 i6=j This is nice, but what does it have to do with hψ, Hψi? The answer is here: Theorem 26. If H : Rn → Rn is any Dirichlet operator, and ψ ∈ Rn is any vector, then 1X Hij (ψi − ψj )2 hψ, Hψi = − 2 i6=j Proof. Let’s start with the formula for power: X P = Hij (ψi − ψj )2 i6=j Note that this sum includes the condition i 6= j, since we only have wires going between distinct vertices. But the summand is zero if i = j, so we also have X P = Hij (ψi − ψj )2 i,j Expanding the square, we get X P = Hij ψi2 − 2Hij ψi ψj + Hij ψj2 i,j The middle term looks promisingly similar to hψ, Hψi, but what about the other two terms? Because Hij = Hji , they’re equal: X P = −2Hij ψi ψj + 2Hij ψj2 i,j And in fact they’re zero! Since H is infinitesimal stochastic, we have X Hij = 0 i so X Hij ψj2 = 0 i and it’s still zero when we sum over j. We thus have X P = −2 Hij ψi ψj i,j 149 15 15.3 DIRICHLET OPERATORS The big picture But since ψi is real, this is -2 times hψ, Hψi = X Hij ψ i ψj i,j So, we’re done. An instant consequence of this theorem is that a Dirichlet operator has hψ, Hψi ≤ 0 for all ψ. Actually most people use the opposite sign convention in defining infinitesimal stochastic operators. This makes Hij ≤ 0, which is mildly annoying, but it gives hψ, Hψi ≥ 0 which is nice. When H is a Dirichlet operator, defined with this opposite sign convention, hψ, Hψi is called a Dirichlet form. 15.3 The big picture Maybe it’s a good time to step back and see where we are. So far we’ve been exploring the analogy between stochastic mechanics and quantum mechanics. Where do networks come in? Well, they’ve actually come in twice so far: 1. First we saw that Petri nets can be used to describe stochastic or quantum processes where things of different kinds randomly react and turn into other things. A Petri net is a kind of network like this: The different kinds of things are the yellow circles; we called them species. The reactions where things turn into other things are the blue squares: we called them transitions. We label the transitions by numbers to say the rates at which they occur. 150 15 15.3 DIRICHLET OPERATORS The big picture 2. Then we looked at stochastic or quantum processes where in each transition a single thing turns into a single thing. We can draw these as Petri nets where each transition has just one species as input and one species as output. But we can also draw them as directed graphs with edges labelled by numbers: Now the dark blue boxes are species and the edges are transitions! Furthermore, in this situation we often call the species states. In this section we looked at a special case of the second kind of network: the Dirichlet operators. For these the ‘forward’ transition rate Hij equals the ‘reverse’ rate Hji , so our graph can be undirected: no arrows on the edges. And for these the rates Hii are determined by the rest, so we can omit the edges from vertices to themselves. We also omit edges labelled by zero: The result can be seen as an electrical circuit made of resistors! So we’re building up a little dictionary: • Stochastic mechanics: ψi is a probability and Hij is a transition rate (probability per time). • Quantum mechanics: ψi is an amplitude and Hij is a transition rate (amplitude per time). • Circuits made of resistors: ψi is a voltage and Hij is a conductance. This dictionary may seem rather odd—especially the third item, which looks completely different than the first two! But that’s good: when things aren’t odd, 151 15 15.3 DIRICHLET OPERATORS The big picture we don’t get many new ideas. The whole point of this ‘network theory’ business is to think about networks from many different viewpoints and let the sparks fly! Actually, this particular oddity is well-known in certain circles. We’ve been looking at the discrete version, where we have a finite set of states. But in the continuum, the classic example of a Dirichlet operator is the Laplacian H = ∇2 . And then: • The heat equation: d ψ = ∇2 ψ dt is fundamental to stochastic mechanics. • The Schrödinger equation: d ψ = −i∇2 ψ dt is fundamental to quantum mechanics. • The Poisson equation: ∇2 ψ = −ρ is fundamental to electrostatics. Briefly speaking, electrostatics is the study of how the electric potential ψ depends on the charge density ρ. The theory of electrical circuits made of resistors can be seen as a special case, at least when the current isn’t changing with time. If you want to learn more, this is a great place to start: [DSle] P. G. Doyle and J. L. Snell, Random Walks and Electrical Circuits, Mathematical Association of America, Washington DC, 1984. This free online book explains, in a really fun informal way, how random walks on graphs, are related to electrical circuits made of resistors. To dig deeper into the continuum case, try: [Fuk80] M. Fukushima, Dirichlet Forms and Markov Processes, North-Holland, Amsterdam, 1980. 152 16 16 PERRON–FROBENIUS THEORY Perron–Frobenius theory We’re in the middle of a battle: in addition to our typical man vs. equation scenario, it’s a battle between two theories. You should know the two opposing forces well by now. It’s our old friends, at it again: Stochastic Mechanics vs Quantum Mechanics! Today we’re reporting live from a crossroads, and we’re facing a skirmish that gives rise to what some might consider a paradox. Let’s sketch the main thesis before we get our hands dirty with the gory details. First we need to tell you that the battle takes place at the intersection of stochastic and quantum mechanics. We recall from Section 15 that there is a class of operators called ‘Dirichlet operators’ that are valid Hamiltonians for both stochastic and quantum mechanics. In other words, you can use them to generate time evolution both for old-fashioned random processes and for quantum processes. Staying inside this class allows the theories to fight it out on the same turf. We will be considering a special subclass of Dirichlet operators, which we call ‘irreducible’ Dirichlet operators. These are the ones that generate Markov processes where starting at any vertex, we have a nonzero chance of winding up at any other vertex. When considering this subclass, we see something interesting: Thesis. Let H be an irreducible Dirichlet operator with n eigenstates. In stochastic mechanics, there is only one valid state that is an eigenvector of H: the unique so-called ‘Perron–Frobenius state’. The other n − 1 eigenvectors are forbidden states of a stochastic system: the stochastic system is either in the Perron–Frobenius state, or in a superposition of at least two eigensvectors. In quantum mechanics, all n eigenstates of H are valid states. This might sound like a riddle, but now we’ll prove, riddle or not, that it’s a fact. 16.1 At the intersection of two theories The typical view of how quantum mechanics and probability theory come into contact looks like this: 153 16 16.1 PERRON–FROBENIUS THEORY At the intersection of two theories The idea is that quantum theory generalizes classical probability theory by considering observables that don’t commute. That’s perfectly valid, but we’ve been exploring an alternative view in this series. Here quantum theory doesn’t subsume probability theory, but they intersect: What goes in the middle, you might ask? We saw in Section 15.2 that electrical circuits made of resistors constitute the intersection! For example, a circuit like this: gives rise to a Hamiltonian H that’s good both for quantum mechanics and stochastic mechanics. Indeed, we saw that the power dissipated by a circuit made of resistors is related to the familiar quantum theory concept known as the expectation value of the Hamiltonian! power = −2hψ, Hψi Oh—and you might think we made a mistake and wrote our Ω (ohm) symbols upside down. We didn’t. It happens that f is the symbol for a ‘mho’—a unit of conductance that’s the reciprocal of an ohm. Reread Section 15 for the details. 154 16 16.2 PERRON–FROBENIUS THEORY Stochastic mechanics versus quantum mechanics 16.2 Stochastic mechanics versus quantum mechanics Before we plunge in, let’s review how states, time evolution, symmetries and observables work in the two theories. In this section we’ll take a lowbrow approach, fixing a basis for our vector space of states, and assuming it’s finite-dimensional so that all vectors have n components which are either complex or real numbers. In other words, we’ll treat our space as either Cn or Rn . In this fashion, linear operators that map such spaces to themselves will be represented as square matrices. Vectors will be written as ψi where the index i runs from 1 to n, and we think of each choice of the index as a state of our system—but since we’ll be using that word in other ways too, let’s call it a configuration. It’s just a basic way our system can be. States Besides the configurations i = 1, . . . , n, we have more general states that tell us the probability or amplitude of finding our system in one of these configurations: • Stochastic states are n-tuples of nonnegative real numbers: ψi ∈ R+ The probability of finding the system in the ith configuration is defined to be ψi . For these probabilities to sum to one, ψi needs to be normalized like this: X ψi = 1 i or in the notation we’re using: hψi = 1 where we define hψi = X ψi i • Quantum states are n-tuples of complex numbers: ψi ∈ C The probability of finding a state in the ith configuration is defined to be |ψ(x)|2 . For these probabilities to sum to one, ψ needs to be normalized like this: X |ψi |2 = 1 i or in other words hψ, ψi = 1 where the inner product of two vectors ψ and φ is defined by X ψ i φi hψ, φi = i 155 16 16.2 PERRON–FROBENIUS THEORY Stochastic mechanics versus quantum mechanics Now, the usual way to turn a quantum state ψ into a stochastic state is to take the absolute value of each number ψi and then square it. However, if the numbers ψi happen to be nonnegative, we can also turn ψ into a stochastic state simply by multiplying it by a number to ensure hψi = 1. This is very unorthodox, but it lets us evolve the same vector ψ either stochastically or quantum-mechanically, using the recipes we’ll describe next. In physics jargon these correspond to evolution in ‘real time’ and ‘imaginary time’. But don’t ask which is which: from a quantum viewpoint stochastic mechanics uses imaginary time, but from a stochastic viewpoint it’s the other way around! Time evolution Time evolution works similarly in stochastic and quantum mechanics, but with a few big differences: • In stochastic mechanics the state changes in time according to the master equation: d ψ(t) = Hψ(t) dt which has the solution ψ(t) = exp(tH)ψ(0) • In quantum mechanics the state changes in time according to Schrödinger’s equation: d ψ(t) = −iHψ(t) dt which has the solution ψ(t) = exp(−itH)ψ(0) The operator H is called the Hamiltonian. The properties it must have depend on whether we’re doing stochastic mechanics or quantum mechanics: • We need H to be infinitesimal stochastic for time evolution given by exp(tH) to send stochastic states to stochastic states. In other words, we need that (i) its columns sum to zero and (ii) its off-diagonal entries are real and nonnegative: X Hij = 0 i i 6= j ⇒ Hij > 0 • We need H to be self-adjoint for time evolution given by exp(−itH) to send quantum states to quantum states. So, we need 156 16 16.2 PERRON–FROBENIUS THEORY Stochastic mechanics versus quantum mechanics H = H† where we recall that the adjoint of a matrix is the conjugate of its transpose: (H † )ij := H ji We are concerned with the case where the operator H generates both a valid quantum evolution and also a valid stochastic one: • H is a Dirichlet operator if it’s both self-adjoint and infinitesimal stochastic. We will soon go further and zoom in on this intersection! But first let’s finish our review. Symmetries As explained in Section 11.3, besides states and observables we need symmetries, which are transformations that map states to states. These include the evolution operators which we only briefly discussed in the preceding subsection. • A linear map U that sends quantum states to quantum states is called an isometry, and isometries are characterized by this property: U †U = 1 • A linear map U that sends stochastic states to stochastic states is called a stochastic operator, and stochastic operators are characterized by these properties: X Uij = 1 i and Uij > 0 A notable difference here is that in our finite-dimensional situation, isometries are always invertible, but stochastic operators may not be! If U is an n × n matrix that’s an isometry, U † is its inverse. So, we also have UU† = 1 and we say U is unitary. But if U is stochastic, it may not have an inverse—and even if it does, its inverse is rarely stochastic. This explains why in stochastic mechanics time evolution is often not reversible, while in quantum mechanics it always is. Problem 23. Suppose U is a stochastic n×n matrix whose inverse is stochastic. What are the possibilities for U ? It is quite hard for an operator to be a symmetry in both stochastic and quantum mechanics, especially in our finite-dimensional situation: Problem 24. Suppose U is an n×n matrix that is both stochastic and unitary. What are the possibilities for U ? 157 16 16.3 PERRON–FROBENIUS THEORY From graphs to matrices Observables ‘Observables’ are real-valued quantities that can be measured, or predicted, given a specific theory. • In quantum mechanics, an observable is given by a self-adjoint matrix O, and the expected value of the observable O in the quantum state ψ is X hψ, Oψi = ψ i Oij ψj i,j • In stochastic mechanics, an observable O has a value Oi in each configuration i, and the expected value of the observable O in the stochastic state ψ is X hOψi = Oi ψi i We can turn an observable in stochastic mechanics into an observable in quantum mechanics by making a diagonal matrix whose diagonal entries are the numbers Oi . 16.3 From graphs to matrices In Section 15 we explained how a graph with positive numbers on its edges gives rise to a Hamiltonian in both quantum and stochastic mechanics—in other words, a Dirichlet operator. Recall how it works. We’ll consider simple graphs: graphs without arrows on their edges, with at most one edge from one vertex to another, and with no edge from a vertex to itself. And we’ll only look at graphs with finitely many vertices and edges. We’ll assume each edge is labelled by a positive number, like this: If our graph has n vertices, we can create an n × n matrix A where Aij is the number labelling the edge from i to j, if there is such an edge, and 0 if there’s not. This matrix is symmetric, with real entries, so it’s self-adjoint. So A is a valid Hamiltonian in quantum mechanics. 158 16 16.4 PERRON–FROBENIUS THEORY Perron’s theorem How about stochastic mechanics? Remember that a Hamiltonian in stochastic mechanics needs to be ‘infinitesimal stochastic’. So, its off-diagonal entries must be nonnegative, which is indeed true for our A, but also the sums of its columns must be zero, which is not true when our A is nonzero. But now comes the best news you’ve heard all day: we can improve A to a stochastic operator in a way that is completely determined by A itself! This is done by subtracting a diagonal matrix L whose entries are the sums of the columns of A: X Lii = Aij i i 6= j ⇒ Lij = 0 It’s easy to check that H =A−L is still self-adjoint, but now also infinitesimal stochastic. So, it’s a Dirichlet operator : a good Hamiltonian for both stochastic and quantum mechanics! In Section 15, we saw a bit more: every Dirichlet operator arises this way. It’s easy to see. You just take your Dirichlet operator and make a graph with one edge for each nonzero off-diagonal entry. Then you label the edge with this entry. So, Dirichlet operators are essentially the same as finite simple graphs with edges labelled by positive numbers. Now, a simple graph can consist of many separate ‘pieces’, called components. Then there’s no way for a particle hopping along the edges to get from one component to another, either in stochastic or quantum mechanics. So we might as well focus our attention on graphs with just one component. These graphs are called ‘connected’. In other words: Definition 27. A simple graph is connected if it is nonempty and there is a path of edges connecting any vertex to any other. Our goal now is to understand more about Dirichlet operators coming from connected graphs. For this we need to learn the Perron–Frobenius theorem. But let’s start with something easier. 16.4 Perron’s theorem In quantum mechanics it’s good to think about observables that have positive expected values: hψ, Oψi ≥ 0 for every quantum state ψ ∈ Cn . These are called positive definite. But in stochastic mechanics it’s good to think about matrices that are positive in a more naive sense: Definition 28. An n×n real matrix T is positive if all its entries are positive: Tij ≥ 0 for all 1 ≤ i, j ≤ n. 159 16 16.5 PERRON–FROBENIUS THEORY From matrices to graphs Similarly: Definition 29. A vector ψ ∈ Rn is positive if all its components are positive: ψi ≥ 0 for all 1 ≤ i ≤ n. We’ll also define nonnegative matrices and vectors in the same way, replacing > 0 by ≥ 0. A good example of a nonnegative vector is a stochastic state. In 1907, Perron proved the following fundamental result about positive matrices: Theorem 30 (Perron’s Theorem). Given a positive square matrix T , there is a positive real number r, called the Perron–Frobenius eigenvalue of T , such that r is an eigenvalue of T and any other eigenvalue λ of T has |λ| < r. Moreover, there is a positive vector ψ ∈ Rn with T ψ = rψ. Any other vector with this property is a scalar multiple of ψ. Furthermore, any nonnegative vector that is an eigenvector of T must be a scalar multiple of ψ. In other words, if T is positive, it has a unique eigenvalue with the largest absolute value. This eigenvalue is positive. Up to a constant factor, it has an unique eigenvector. We can choose this eigenvector to be positive. And then, up to a constant factor, it’s the only nonnegative eigenvector of T . 16.5 From matrices to graphs The conclusions of Perron’s theorem don’t hold for matrices that are merely nonnegative. For example, these matrices 1 0 0 1 , 0 1 0 0 are nonnegative, but they violate lots of the conclusions of Perron’s theorem. Nonetheless, in 1912 Frobenius published an impressive generalization of Perron’s result. In its strongest form, it doesn’t apply to all nonnegative matrices; only to those that are ‘irreducible’. So, let us define those. We’ve seen how to build a matrix from a graph. Now we need to build a graph from a matrix! Suppose we have an n × n matrix T . Then we can build a graph GT with n vertices where there is an edge from the ith vertex to the jth vertex if and only if Tij 6= 0. But watch out: this is a different kind of graph! It’s a directed graph, meaning the edges have directions, there’s at most one edge going from any vertex to any vertex, and we do allow an edge going from a vertex to itself. There’s a stronger concept of ‘connectivity’ for these graphs: Definition 31. A directed graph is strongly connected if there is a directed path of edges going from any vertex to any other vertex. 160 16 16.6 PERRON–FROBENIUS THEORY The Perron–Frobenius theorem So, you have to be able to walk along edges from any vertex to any other vertex, but always following the direction of the edges! Using this idea we define irreducible matrices: Definition 32. A nonnegative square matrix T is irreducible if its graph GT is strongly connected. 16.6 The Perron–Frobenius theorem Now we are ready to state: Theorem 33 (The Perron–Frobenius Theorem). Given an irreducible nonnegative square matrix T , there is a positive real number r, called the Perron– Frobenius eigenvalue of T , such that r is an eigenvalue of T and any other eigenvalue λ of T has |λ| ≤ r. Moreover, there is a positive vector ψ ∈ Rn with T ψ = rψ. Any other vector with this property is a scalar multiple of ψ. Furthermore, any nonnegative vector that is an eigenvector of T must be a scalar multiple of ψ. The only conclusion of this theorem that’s weaker than those of Perron’s theorem is that there may be other eigenvalues with |λ| = r. For example, this matrix is irreducible and nonnegative: 0 1 1 0 Its Perron–Frobenius eigenvalue is 1, but it also has -1 as an eigenvalue. In general, Perron–Frobenius theory says quite a lot about the other eigenvalues on the circle |λ| = r, but we won’t need that fancy stuff here. Perron–Frobenius theory is useful in many ways, from highbrow math to ranking football teams. We’ll need it not just now but also in Section 22. There are many books and other sources of information for those that want to take a closer look at this subject. If you’re interested, you can search online or take a look at these: • Dimitrious Noutsos, Perron Frobenius theory and some extensions, 2008. (Includes proofs of the basic theorems.) • V. S. Sunder, Perron Frobenius theory, 18 December 2009. (Includes applications to graph theory, Markov chains and von Neumann algebras.) • Stephen Boyd, Lecture 17: Perron Frobenius theory, Winter 2008-2009. (Includes a max-min characterization of the Perron–Frobenius eigenvalue and applications to Markov chains, economics, population growth and power control.) If you are interested and can read German, the original work appears here: [Per07] Oskar Perron, Zur Theorie der Matrizen, Math. Ann. 64 (1907), 248–63. 161 16 16.7 PERRON–FROBENIUS THEORY Irreducible Dirichlet operators [Fro12] Georg Frobenius, Über Matrizen aus nicht negativen Elemente, S.-B. Preuss Acad. Wiss. Berlin (1912), 456–477. And, of course, there’s this: • Wikipedia, Perron–Frobenius theorem. It’s quite good. 16.7 Irreducible Dirichlet operators Now comes the payoff. We saw how to get a Dirichlet operator H from any finite simple graph with edges labelled by positive numbers. Now let’s apply Perron–Frobenius theory to prove our thesis. Unfortunately, the matrix H is rarely nonnegative. If you remember how we built it, you’ll see its off-diagonal entries will always be nonnegative... but its diagonal entries can be negative. Luckily, we can fix this just by adding a big enough multiple of the identity matrix to H! The result is a nonnegative matrix T = H + cI where c > 0 is some large number. This matrix T has the same eigenvectors as H. The off-diagonal matrix entries of T are the same as those of H, so Tij is nonzero for i 6= j exactly when the graph we started with has an edge from i to j. So, for i 6= j, the graph GT will have an directed edge going from i to j precisely when our original graph had an edge from i to j. And that means that if our original graph was connected, GT will be strongly connected. Thus, by definition, the matrix T is irreducible! Since T is nonnegative and irreducible, the Perron–Frobenius theorem swings into action and we conclude: Lemma 34. Suppose H is the Dirichlet operator coming from a connected finite simple graph with edges labelled by positive numbers. Then the eigenvalues of H are real. Let λ be the largest eigenvalue. Then there is a positive vector ψ ∈ Rn with Hψ = λψ. Any other vector with this property is a scalar multiple of ψ. Furthermore, any nonnegative vector that is an eigenvector of H must be a scalar multiple of ψ. Theorem 35. The eigenvalues of H are real since H is self-adjoint. Notice that if r is the Perron–Frobenius eigenvalue of T = H + cI and T ψ = rψ then Hψ = (r − c)ψ By the Perron–Frobenius theorem the number r is positive, and it has the largest absolute value of any eigenvalue of T . Thanks to the subtraction, the eigenvalue 162 16 16.8 PERRON–FROBENIUS THEORY An example r − c may not have the largest absolute value of any eigenvalue of H. It is, however, the largest eigenvalue of H, so we take this as our λ. The rest follows from the Perron–Frobenius theorem. But in fact we can improve this result, since the largest eigenvalue λ is just zero. Let’s also make up a definition, to make our result sound more slick: Definition 36. A Dirichlet operator is irreducible if it comes from a connected finite simple graph with edges labelled by positive numbers. This meshes nicely with our earlier definition of irreducibility for nonnegative matrices. Now: Theorem 37. Suppose H is an irreducible Dirichlet operator. Then H has zero as its largest real eigenvalue. There is a positive vector ψ ∈ Rn with Hψ = 0. Any other vector with this property is a scalar multiple of ψ. Furthermore, any nonnegative vector that is an eigenvector of H must be a scalar multiple of ψ. Proof. Choose λ as in the Lemma, so that Hψ = λψ. Since ψ is positive we can normalize it to be a stochastic state: X ψi = 1 i Since H is a Dirichlet operator, exp(tH) sends stochastic states to stochastic states, so X (exp(tH)ψ)i = 1 i for all t ≥ 0. On the other hand, X X (exp(tH)ψ)i = etλ ψi = etλ i i so we must have λ = 0. What’s the point of all this? One point is that there’s a unique stochastic state ψ that’s an equilibrium, meaning that Hψ = 0, so doesn’t change with time. It’s also globally stable: since all the other eigenvalues of H are negative, all other stochastic states converge to this one as time goes forward. 16.8 An example There are many examples of irreducible Dirichlet operators. For instance, in Section 14 we talked about graph Laplacians. The Laplacian of a connected simple graph is always irreducible. But let us try a different sort of example, coming from the picture of the resistors we saw earlier: 163 16 16.8 PERRON–FROBENIUS THEORY An example Let’s create a matrix A whose entry Aij is the number labelling the edge from i to j if there is such an edge, and zero otherwise: 0 2 1 0 1 2 0 0 1 1 A= 1 0 0 2 1 0 1 2 0 1 1 1 1 1 0 Remember how the game works. The matrix A is already a valid Hamiltonian for quantum mechanics, since it’s self adjoint. However, to get a valid Hamiltonian for both stochastic and quantum mechanics—in other words, a Dirichlet operator—we subtract the diagonal matrix L whose entries are the sums of the columns of A. In this example it just so happens that the column sums are all 4, so L = 4I, and our Dirichlet operator is −4 2 1 0 1 2 −4 0 1 1 0 −4 2 1 H = A − 4I = 1 0 1 2 −4 1 1 1 1 1 −4 We’ve set up this example so it’s easy to see that the vector ψ = (1, 1, 1, 1, 1) has Hψ = 0 So, this is the unique eigenvector for the eigenvalue 0. We can use Mathematica to calculate the remaining eigenvalues of H. The set of eigenvalues is {0, −7, −8, −8, −3} As we expect from our theorem, the largest real eigenvalue is 0. By design, the eigenstate associated to this eigenvalue is |v0 i = (1, 1, 1, 1, 1) 164 16 16.9 PERRON–FROBENIUS THEORY Problems (This funny notation for vectors is common in quantum mechanics, so don’t worry about it.) All the other eigenvectors fail to be nonnegative, as predicted by the theorem. They are: |v1 i |v2 i |v3 i |v4 i = = = = (1, −1, −1, 1, 0), (−1, 0, −1, 0, 2), (−1, 1, −1, 1, 0), (−1, −1, 1, 1, 0). To compare the quantum and stochastic states, consider first |v0 i. This is the only eigenvector that can be normalized to a stochastic state. Remember, a stochastic state must have nonnegative components. This rules out |v1 i through to |v4 i as valid stochastic states, no matter how we normalize them! However, these are allowed as states in quantum mechanics, once we normalize them correctly. For a stochastic system to be in a state other than the Perron– Frobenius state, it must be a linear combination of least two eigenstates. For instance, ψa = (1 − a)|v0 i + a|v1 i can be normalized to give stochastic state only if 0 ≤ a ≤ 21 . And, it’s easy to see that it works this way for any irreducible Dirichlet operator, thanks to our theorem. So, our thesis has been proved true! 16.9 Problems Let us conclude with a few more problems. There are lots of ways to characterize irreducible nonnegative matrices; we don’t need to mention graphs. Here’s one: Problem 25. Let T be a nonnegative n × n matrix. Show that T is irreducible if and only if for all i, j ≥ 0, (T m )ij > 0 for some natural number m. You may be confused because we explained the usual concept of irreducibility for nonnegative matrices, but we also defined a concept of irreducibility for Dirichlet operators. Luckily there’s no conflict: Dirichlet operators aren’t nonnegative matrices, but if we add a big multiple of the identity to a Dirichlet operator it becomes a nonnegative matrix, and then: Problem 26. Show that a Dirichlet operator H is irreducible if and only if the nonnegative operator H + cI (where c is any sufficiently large constant) is irreducible. Irreducibility is also related to the nonexistence of interesting conserved quantities. In Section 10 we saw a version of Noether’s Theorem for stochastic mechanics. Remember that an observable O in stochastic mechanics assigns a number Oi to each configuration i = 1, . . . , n. We can make a diagonal matrix with Oi as its diagonal entries, and by abuse of language we call this O as well. Then we say O is a conserved quantity for the Hamiltonian H if the commutator [O, H] = OH − HO vanishes. 165 16 16.10 PERRON–FROBENIUS THEORY Answers Problem 27. Let H be a Dirichlet operator. Show that H is irreducible if and only if every conserved quantity O for H is a constant, meaning that for some c ∈ R we have Oi = c for all i. (Hint: examine the proof of Noether’s theorem.) In fact this works more generally: Problem 28. Let H be an infinitesimal stochastic matrix. Show that H + cI is an irreducible nonnegative matrix for all sufficiently large c if and only if every conserved quantity O for H is a constant. 16.10 Answers Here are the answers to the problems in this section, thanks to Arjun Jain. Problem 23. Suppose U is a stochastic n×n matrix whose inverse is stochastic. What are the possibilities for U ? Answer. The matrix U must be a permutation matrix: that is, all its entries must be 0 except for a single 1 in each row and each column. This follows from the answer to Problem 16, which appears in Section 11.7. In fact, Problem 23 is just Problem 16 stated in a less fancy way—but this prompted Arjun Jain to find the following less fancy solution. To see this, suppose that U is a stochastic n × n matrix whose inverse, V , is also stochastic. In particular, we have X Uij , Vij ≥ 0, Uij = 1 i and X Uik Vkj = δij k Fix i and j. First, suppose i 6= j. We have X Uik Vkj = 0 k Since Uik , Vkj ≥ 0, for every choice of k either Uik or Vkj must be zero. It follows that if some entry Uik is nonzero, then Vkj must be zero for all j except for j = i. Further if Ui0 k is also nonzero where i0 6= i, then Vkj must be zero except for j = i0 . Since i0 6= i, Vki is also zero, and so is Vki0 . Thus, if any two entries of a column in U were nonzero, some row of V would be zero, implying that V is noninvertible. As this is not the case, at most one entry in each column of U can be nonzero. Furthermore, since the sum of the entries in each column is 1: X Uij = 1 i the one nonzero entry in each column of U must equal 1. 166 16 16.10 PERRON–FROBENIUS THEORY Answers Next suppose i = j. Now we have X Uik Vki = 1 k Since each column of U has just one nonzero entry, for each k we have one choice of index I such that UIj 6= 0. Suppose that i is not equal to this choice of I for any k. Then X Uik Vki = 0 k which is a contradiction. Thus, for every i we have a unique k such that Uik is nonzero. In other words, there is exactly one nonzero entry in each row of U . In short, a stochastic matrix with a stochastic inverse must be a permutation matrix. Conversely, it is easy to check that any permutation matrix is stochastic, with its inverse being another permutation matrix, and thus also stochastic. Problem 24. Suppose U is an n×n matrix that is both stochastic and unitary. What are the possibilities for U ? Answer. In the solution to Problem 23 we only used the fact that U is a stochastic matrix whose inverse has nonnegative entries. If U is stochastic and unitary this is true, because its inverse is its transpose. So, the argument in Problem 23 shows that a stochastic unitary matrix must be a permutation matrix. Conversely, it is easy to check that any permutation matrix is stochastic and unitary. Problem 25. Let T be a nonnegative n × n matrix. Show that T is irreducible if and only if for all i, j ≥ 0, (T m )ij > 0 for some natural number m. Answer. As in Section 16.5, we build a directed graph GT with n vertices, with an edge from the ith vertex to the jth if and only if Tij 6= 0. We need to show this graph is strongly connected if and only if for all i, j ≥ 0, (T m )ij > 0 for some natural number m. Recall that a directed graph is strongly connected when there is a directed edge path from any vertex to any other. Notice that X Tij2 = Tik Tkj k If Tik Tkj is nonzero, it means that neither Tik nor Tkj is zero: that is, the graph GT has an edge from i to k and an edge from k to j. Thus Tik Tkj is nonzero if and only if there is a directed 2-edge path from i to j. Since all entries of T are nonnegative, Tij2 is zero only if there is no directed 2-edge path from i to j. If it is nonzero, there is at least one directed 2-edge path from i to j. Next consider X 2 Tij3 = Tik Tkj k 167 16 16.10 PERRON–FROBENIUS THEORY Answers This is nonzero if and only if there is at least one directed 2-edge path from i to k and a directed edge from k to j—and thus a directed 3-edge path from i to j. Using induction, we can show that Tijm is nonzero if and only if the graph GT has a directed m-edge path from i to j. Thus, this graph is strongly connected if and only if i, j ≥ 0, (T m )ij > 0 for some natural number m. Problem 26. Show that a Dirichlet operator H is irreducible if and only if the nonnegative operator H + cI (where c is any sufficiently large constant) is irreducible. Answer. Recall that by definition, an n × n Dirichlet operator H is irreducible if and only if the simple graph ΓH with an edge from i to j when Hij > 0 is connected. Similarly, the nonnegative operator T = H + cI is irreducible if and only if the directed graph GT with an edge from i to j when Tij > 0 is strongly connected. Since the matrix H is symmetric, so is T , so there is no difference between connectedness and strong connectedness for GT . Only the off-diagonal entries of H matter for determining the connectedness of ΓH , and similarly, only the off-diagonal entries of T matter for determining the connectedness of GT . Since H and T have the same off-diagonal entries, ΓH is connected if and only if GT is. Problem 27. Let H be a Dirichlet operator. Show that H is irreducible if and only if every conserved quantity O for H is a constant, meaning that for some c ∈ R we have Oi = c for all i. Answer. This is a special case of the next problem. Problem 28. Let H be an infinitesimal stochastic matrix. Show that H +cI is an irreducible nonnegative matrix for all sufficiently large c if and only if every conserved quantity O for H is a constant. Answer. Let T = H + cI where c is chosen large enough to make T be nonnegative. If O is a conserved quantity for H, then OH = HO, implying that OH m = H m O for all m ∈ N. Since a scalar multiply of the identity matrix commutes with everything, it follows that OT m = T m O for all m. As O is a diagonal matrix we have (Oii − Ojj )Tijm = 0 for all i, j. If T is irreducible, then from Problem 25 there exists a natural number m such that Tijm > 0 for every i, j with i 6= j. By the above equation this implies (Oii − Ojj ) = 0. Thus, O is a scalar multiple of the identity, or in other words, a constant. Conversely, assume that T is not irreducible. Let GT be the directed graph associated to T . Then there is a diagonal matrix O that is not a scalar multiple 168 16 16.10 PERRON–FROBENIUS THEORY Answers of the identity such that Oii is constant on each strongly connected component of GT . This implies that (Oii − Ojj )Tijm = 0 for all i, j and all m ∈ N. This in turn implies that OT m = T m O for all m, and thus OH = HO. 169 17 17 THE DEFICIENCY ZERO THEOREM The deficiency zero theorem We’ve seen how Petri nets can be used to describe chemical reactions. Indeed our very first example way back in Section 2 came from chemistry: However, chemists rarely use the formalism of Petri nets. They use a different but entirely equivalent formalism, called ‘reaction networks’. So now we’d like to tell you about those. You may wonder: why bother with another formalism, if it’s equivalent to the one we’ve already seen? Well, one goal of this network theory program is to get people from different subjects to talk to each other—or at least be able to. This requires setting up some dictionaries to translate between formalisms. Furthermore, lots of deep results on stochastic Petri nets are being proved by chemists—but phrased in terms of reaction networks. So you need to learn this other formalism to read their papers. Finally, this other formalism is actually better in some ways! 17.1 Reaction networks Here’s a reaction network: This network involves five species: that is, different kinds of things. They could be atoms, molecules, ions or whatever: chemists call all of these species, but there’s no need to limit the applications to chemistry: in population biology, they could even be biological species! We’re calling them A, B, C, D, and E, but in applications, we’d call them by specific names like CO2 and HCO3 , or ‘rabbit’ and ‘wolf’, or whatever. 170 17 17.1 THE DEFICIENCY ZERO THEOREM Reaction networks This network also involves five reactions, which are shown as arrows. Each reaction turns one bunch of species into another. So, written out more longwindedly, we’ve got these reactions: A→B B→A A+C →D B+E →A+C B+E →D If you remember how Petri nets work, you can see how to translate any reaction network into a Petri net, or vice versa. For example, the reaction network we’ve just seen gives this Petri net: Each species corresponds to a yellow circle in this Petri net. And each reaction corresponds to transition of this Petri net, drawn as a blue square. The arrows say how many things of each species appear as input or output to each transition. There’s less explicit emphasis on complexes in the Petri net notation, but you can read them off if you want them. 171 17 17.1 THE DEFICIENCY ZERO THEOREM Reaction networks In chemistry, a bunch of species is called a ‘complex’. But what do we mean by ‘bunch’, exactly? Well, we mean that in a given complex, each species can show up 0,1,2,3... or any natural number of times. So, we can formalize things like this: Definition 38. Given a set S of species, a complex of those species is a function C : S → N. Roughly speaking, a reaction network is a graph whose vertices are labelled by complexes. Unfortunately, the word ‘graph’ means different things in mathematics—appallingly many things! Everyone agrees that a graph has vertices and edges, but there are lots of choices about the details. Most notably: • We can either put arrows on the edges, or not. • We can either allow more than one edge between vertices, or not. • We can either allow edges from a vertex to itself, or not. If we say ‘no’ in every case we get the concept of ‘simple graph’, which we discussed in Sections 15 and 16. At the other extreme, if we say ‘yes’ in every case we get the concept of ‘directed multigraph’, which is what we want now. A bit more formally: Definition 39. A directed multigraph consists of a set V of vertices, a set E of edges, and functions s, t : E → V saying the source and target of each edge. Given this, we can say: Definition 40. A reaction network is a set of species together with a directed multigraph whose vertices are labelled by complexes of those species. You can now prove that reaction networks are equivalent to Petri nets: Problem 29. Show that any reaction network gives rise to a Petri net, and vice versa. In a stochastic Petri net each transition is labelled by a rate constant: that is, a number in (0, ∞). This lets us write down some differential equations saying how species turn into each other. So, let’s make this definition (which is not standard, but will clarify things for us): Definition 41. A stochastic reaction network is a reaction network where each reaction is labelled by a rate constant. Now you can do this: Problem 30. Show that any stochastic reaction network gives rise to a stochastic Petri net, and vice versa. 172 17 17.2 THE DEFICIENCY ZERO THEOREM The deficiency zero theorem For extra credit, show that in each of these problems we actually get an equivalence of categories! For this you need to define morphisms between Petri nets, morphisms between reaction networks, and similarly for stochastic Petric nets and stochastic reaction networks. If you get stuck, ask Eugene Lerman for advice. There are different ways to define morphisms, but he knows a cool one. We’ve been downplaying category theory so far, but it’s been lurking beneath everything we do, and someday it may rise to the surface. 17.2 The deficiency zero theorem You may already have noticed one advantage of reaction networks over Petri nets: they’re quicker to draw. This is true even for tiny examples. For instance, this reaction network: 2X1 + X2 ↔ 2X3 corresponds to this Petri net: But there’s also a deeper advantage. As we know, any stochastic Petri net gives two equations: • The master equation, which says how the probability that we have a given number of things of each species changes with time. • The rate equation, which says how the expected number of things of each species changes with time. 173 17 17.2 THE DEFICIENCY ZERO THEOREM The deficiency zero theorem The simplest solutions of these equations are the equilibrium solutions, where nothing depends on time. Back in Section 8, we explained when an equilibrium solution of the rate equation gives an equilibrium solution of the master equation. But when is there an equilibrium solution of the rate equation in the first place? The ‘deficiency zero theorem’ gives a handy sufficient condition. And this condition is best stated using reaction networks! But to understand it, we need to understand the ‘deficiency’ of a reaction network. So let’s define that, and then say what all the words in the definition mean: Definition 42. The deficiency of a reaction network is: • the number of vertices minus • the number of connected components minus • the dimension of the stoichiometric subspace. The first two concepts here are easy. A reaction network is a graph (okay, a directed multigraph). So, it has some number of vertices, and also some number of connected components. Two vertices lie in the same connected component iff you can get from one to the other by a path where you don’t care which way the arrows point. For example, this reaction network: has 5 vertices and 2 connected components. So, what’s the ‘stoichiometric subspace’ ? ‘Stoichiometry’ is a scary-sounding word. It comes from the Greek words stoicheion, meaning element, and metron, meaning measure. In chemistry, it’s the study of the relative quantities of reactants and products in chemical reactions. But for us, stoichiometry is just the art of counting species. To do this, we can form a vector space RS where S is the set of species. A vector in RS is a function from species to real numbers, saying how much of each species is present. Any complex gives a vector in RS , because it’s actually a function from species to natural numbers. Definition 43. The stoichiometric subspace of a reaction network is the subspace Stoch ⊆ RS spanned by vectors of the form x − y where x and y are complexes connected by a reaction. ‘Complexes connected by a reaction’ makes sense because vertices in the reaction network are complexes, and edges are reactions. Let’s see how it works in our example: 174 17 17.2 THE DEFICIENCY ZERO THEOREM The deficiency zero theorem Each complex here can be seen as a vector in RS , which is a vector space whose basis we can call A, B, C, D, E. Each reaction gives a difference of two vectors coming from complexes: • The reaction A → B gives the vector B − A. • The reaction B → A gives the vector A − B. • The reaction A + C → D gives the vector D − A − C. • The reaction B + E → A + C gives the vector A + C − B − E. • The reaction B + E → D gives the vector D − B − E. The pattern is obvious, we hope. These 5 vectors span the stoichiometric subspace. But this subspace isn’t 5-dimensional, because these vectors are linearly dependent! The first vector is the negative of the second one. The last is the sum of the previous two. And those are all the linear dependencies, so the stoichiometric subspace is 3 dimensional. For example, it’s spanned by these 3 linearly independent vectors: A − B, D − A − C, and D − B − E. We hope you see the moral of this example: the stoichiometric subspace is the space of ways to move in RS that are allowed by the reactions in our reaction network! And this is important because the rate equation describes how the amount of each species changes as time passes. So, it describes a point moving around in RS . Thus, if Stoch ⊆ RS is the stoichiometric subspace, and x(t) ∈ RS is a solution of the rate equation, then x(t) always stays within the set x(0) + Stoch = {x(0) + y : y ∈ Stoch} Mathematicians would call this set the coset of x(0), but chemists call it the stoichiometric compatibility class of x(0). Anyway: what’s the deficiency of the reaction network in our example? It’s 5−2−3=0 since there are 5 complexes, 2 connected components and the dimension of the stoichiometric subspace is 3. But what’s the deficiency zero theorem? You’re almost ready for it. You just need to know one more piece of jargon! A reaction network is weakly reversible if whenever there’s a reaction going from a complex x to a complex 175 17 17.2 THE DEFICIENCY ZERO THEOREM The deficiency zero theorem y, there’s a path of reactions going back from y to x. Here the paths need to follow the arrows. So, this reaction network is not weakly reversible: since we can get from A + C to D but not back from D to A + C, and from B + E to D but not back, and so on. However, the network becomes weakly reversible if we add a reaction going back from D to B + E: If a reaction network isn’t weakly reversible, one complex can turn into another, but not vice versa. In this situation, what typically happens is that as time goes on we have less and less of one species. We could have an equilibrium where there’s none of this species. But we have little right to expect an equilibrium solution of the rate equation that’s positive, meaning that it sits at a point x ∈ (0, ∞)S , where there’s a nonzero amount of every species. The argument here is not watertight: you’ll note that we fudged the difference between species and complexes. But it can be made so when our reaction network has deficiency zero: Theorem 44 (Deficiency Zero Theorem). Suppose we are given a reaction network with a finite set of species S, and suppose its deficiency is zero. Then: (i) If the network is not weakly reversible and the rate constants are positive, the rate equation does not have a positive equilibrium solution. (ii) If the network is not weakly reversible and the rate constants are positive, the rate equation does not have a positive periodic solution, that is, a periodic solution lying in (0, ∞)S . (iii) If the network is weakly reversible and the rate constants are positive, the rate equation has exactly one equilibrium solution in each positive stoichiometric compatibility class. This equilibrium solution is complex balanced. 176 17 17.3 THE DEFICIENCY ZERO THEOREM References and remarks Any sufficiently nearby solution that starts in the same stoichiometric compatibility class will approach this equilibrium as t → +∞. Furthermore, there are no other positive periodic solutions. This is quite an impressive result. Even better, the ‘complex balanced’ condition means we can instantly turn the equilibrium solutions of the rate equation we get from this theorem into equilibrium solutions of the master equation, thanks to the Anderson–Craciun–Kurtz theorem! If you don’t remember what we’re talking about here, go back to Section 8. We’ll look at an easy example of this theorem in Section 18. 17.3 References and remarks The deficiency zero theorem appears here: [Fei87] Martin Feinberg, Chemical reaction network structure and the stability of complex isothermal reactors: I. The deficiency zero and deficiency one theorems, Chemical Engineering Science 42 (1987), 2229–2268. [HJ72] F. Horn and Roy Jackson, General mass action kinetics, Archive for Rational Mechanics and Analysis, 47, 81–116, 1972. You can learn more about chemical reaction networks and the deficiency zero theorem here: [Fei79] Martin Feinberg, Lectures on Reaction Networks, 1979. [Gun03] Jeremy Gunawardena, Chemical reaction network theory for in-silico biologists, 2003. At first glance the deficiency zero theorem might seem to settle all the basic questions about the dynamics of reaction networks, or stochastic Petri nets... but actually, it just means that deficiency zero reaction networks don’t display very interesting dynamics in the limit as t → +∞. So, to get more interesting behavior, we need to look at reaction networks that don’t have deficiency zero. For example, in biology it’s interesting to think about ‘bistable’ chemical reactions: reactions that have two stable equilibria. An electrical switch of the usual sort is a bistable system: it has stable ‘on’ and ‘off’ positions. A bistable chemical reaction can serve as a kind of biological switch: [CTF06] Gheorghe Craciun, Yangzhong Tang and Martin Feinberg, Understanding bistability in complex enzyme-driven reaction networks, PNAS 103 (2006), 8697–8702. It’s also interesting to think about chemical reactions with stable periodic solutions. Such a reaction can serve as a biological clock: [For11] Daniel B. Forger, Signal processing in cellular clocks, PNAS 108 (2011), 4281–4285. 177 18 18 EXAMPLE OF THE DEFICIENCY ZERO THEOREM Example of the deficiency zero theorem In the last section we explained how ‘reaction networks’, as used in chemistry, are just another way of talking about Petri nets. We stated an amazing result on reaction networks: the deficiency zero theorem. This settles quite a number of questions about chemical reactions. Now let’s illustrate it with an example. Our example won’t show how powerful this theorem is: it’s too simple. But it’ll help explain the ideas involved. 18.1 Diatomic molecules A diatomic molecule consists of two atoms of the same kind, stuck together: At room temperature there are 5 elements that are diatomic gases: hydrogen, nitrogen, oxygen, fluorine, chlorine. Bromine is a diatomic liquid, but easily evaporates into a diatomic gas: Iodine is a crystal at room temperatures: 178 18 18.2 EXAMPLE OF THE DEFICIENCY ZERO THEOREM A reaction network but if you heat it a bit, it becomes a diatomic liquid and then a gas: so people often list it as a seventh member of the diatomic club. When you heat any diatomic gas enough, it starts becoming a ‘monatomic’ gas as molecules break down into individual atoms. However, just as a diatomic molecule can break apart into two atoms: A2 → A + A two atoms can recombine to form a diatomic molecule: A + A → A2 So in equilibrium, the gas will be a mixture of diatomic and monatomic forms. The exact amount of each will depend on the temperature and pressure, since these affect the likelihood that two colliding atoms stick together, or a diatomic molecule splits apart. The detailed nature of our gas also matters, of course. But we don’t need to get into these details here! Instead, we can just write down the ‘rate equation’ for the reactions we’re talking about. All the details we’re ignoring will be hiding in some constants called ‘rate constants’. We won’t try to compute these; we’ll leave that to our chemist friends. 18.2 A reaction network To write down our rate equation, we start by drawing a ‘reaction network’. For this, we can be a bit abstract and call the diatomic molecule B instead of A2 . Then it looks like this: We could write down the same information using a Petri net: 179 18 18.2 EXAMPLE OF THE DEFICIENCY ZERO THEOREM A reaction network But now let’s focus on the reaction network! Staring at this picture, we can read off various things: • Species. The species are the different kinds of atoms, molecules, etc. In our example the set of species is S = {A, B}. • Complexes. A complex is a finite sum of species, like A, or A + A, or for a fancier example using more efficient notation, 2A+3B. So, we can think of a complex as a vector v ∈ RS . The complexes that actually show up in our reaction network form a set C ⊆ RS . In our example, C = {A+A, B}. • Reactions. A reaction is an arrow going from one complex to another. In our example we have two reactions: A + A → B and B → A + A. Chemists define a reaction network to be a triple (S, C, T ) where S is a set of species, C is the set of complexes that appear in the reactions, and T is the set of reactions v → w where v, w ∈ C. (Stochastic Petri net people call reactions transitions, hence the letter T .) So, in our example we have: • Species: S = {A, B}. • Complexes: C = {A + A, B}. • Reactions: T = {A + A → B, B → A + A}. To get the rate equation, we also need one more piece of information: a rate constant r(τ ) for each reaction τ ∈ T . This is a nonnegative real number that affects how fast the reaction goes. All the details of how our particular diatomic gas behaves at a given temperature and pressure are packed into these constants! 180 18 18.3 18.3 EXAMPLE OF THE DEFICIENCY ZERO THEOREM The rate equation The rate equation The rate equation says how the expected numbers of the various species, atoms, molecules and the like changes with time. This equation is deterministic. It’s a good approximation when the numbers are large and any fluctuations in these numbers are negligible by comparison. Here’s the general form of the rate equation: X dxi = r(τ ) (ni (τ ) − mi (τ )) xm(τ ) dt τ ∈T Let’s take a closer look. The quantity xi is the expected population of the ith species. So, this equation tells us how that changes. But what about the right hand side? As you might expect, it’s a sum over reactions. And: • The term for the reaction τ is proportional to the rate constant r(τ ). • Each reaction τ goes between two complexes, so we can write it as m(τ ) → n(τ ). Among chemists the input m(τ ) is called the reactant complex, and the output is called the product complex. The difference ni (τ ) − mi (τ ) tells us how many items of species i get created, minus how many get destroyed. So, it’s the net amount of this species that gets produced by the reaction τ . The term for the reaction τ is proportional to this, too. • Finally, the law of mass action says that the rate of a reaction is proportional to the product of the concentrations of the species that enter as inputs. More precisely, if we have a reaction τ where the input is the comm (τ ) m (τ ) plex m(τ ), we define xm(τ ) = x1 1 · · · xk k . The law of mass action says the term for the reaction τ is proportional to this, too! Let’s see what this says for the reaction network we’re studying: Let’s write x1 (t) for the number of A atoms and x2 (t) for the number of B molecules. Let the rate constant for the reaction B → A + A be α, and let the rate constant for A + A → B be β. Then the rate equation is this: d x1 = 2αx2 − 2βx21 dt d x2 = −αx2 + βx21 dt This is a bit intimidating. However, we can solve it in closed form thanks to something very precious: a conserved quantity. 181 18 18.3 EXAMPLE OF THE DEFICIENCY ZERO THEOREM The rate equation We’ve got two species, A and B. But remember, B is just an abbreviation for a molecule made of two A atoms. So, the total number of A atoms is conserved by the reactions in our network. This is the number of A’s plus twice the number of B’s: x1 + 2x2 . So, this should be a conserved quantity: it should not change with time. Indeed, by adding the first equation above to twice the second, we see: d (x1 + 2x2 ) = 0 dt As a consequence, any solution will stay on a line x1 + 2x2 = c for some constant c. We can use this fact to rewrite the rate equation just in terms of x1 : d x1 = α(2c − x1 ) − 2βx21 dt This is a separable differential equation, so we can solve it if we can figure out how to do this integral Z dx1 t= α(2c − x1 ) − 2βx21 and then solve for x1 . This sort of trick won’t work for more complicated examples. But the idea remains important: the numbers of atoms of various kinds—hydrogen, helium, lithium, and so on—are conserved by chemical reactions, so a solution of the rate equation can’t roam freely in RS . It will be trapped on some hypersurface, which is called a ‘stoichiometric compatibility class’. And this is very important. We don’t feel like doing the integral required to solve our rate equation in closed form, because this idea doesn’t generalize too much. On the other hand, we can always solve the rate equation numerically. So let’s try that! For example, suppose we set α = β = 1. We can plot the solutions for three different choices of initial conditions, say (x1 , x2 ) = (0, 3), (4, 0), and (3, 3). We get these graphs: It looks like the solution always approaches an equilibrium. We seem to be getting different equilibria for different initial conditions, and the pattern is a bit mysterious. However, something nice happens when we plot the ratio x21 /x2 : 182 18 18.4 EXAMPLE OF THE DEFICIENCY ZERO THEOREM Deficiency zero theorem Apparently it always converges to 1. Why should that be? It’s not terribly surprising. With both rate constants equal to 1, the reaction A + A → B proceeds at a rate equal to the square of the number of A’s, namely x21 . The reverse reaction proceeds at a rate equal to the number of B’s, namely x2 . So in equilibrium, we should have x21 = x2 . But why is the equilibrium stable? In this example we could see that using the closed-form solution, or maybe just common sense. But it also follows from a powerful theorem that handles a lot of reaction networks. 18.4 The deficiency zero theorem It’s called the deficiency zero theorem, and we saw it in Section 17.2. Very roughly, it says that if our reaction network is ‘weakly reversible’ and has ‘deficiency zero’, the rate equation will have equilibrium solutions that behave about as nicely as you could want. Let’s see how this works. We need to remember some jargon: • Weakly reversible. A reaction network is weakly reversible if for every reaction v → w in the network, there exists a path of reactions in the network starting at w and leading back to v. • Reversible. A reaction network is reversible if for every reaction v → w in the network, w → v is also a reaction in the network. Any reversible reaction network is weakly reversible. Our example is reversible, since it consists of reactions A + A → B, B → A + A. But what about ‘deficiency zero’ ? We defined that concept in Section 18, but let’s review: • Connected component. A reaction network gives a kind of graph with complexes as vertices and reactions as edges. Two complexes lie in the same connected component if we can get from one to the other by a path of reactions, where at each step we’re allowed to go either forward or backward along a reaction. Chemists call a connected component a linkage class. In our example there’s just one: 183 18 18.4 EXAMPLE OF THE DEFICIENCY ZERO THEOREM Deficiency zero theorem • Stoichiometric subspace. The stoichiometric subspace is the subspace Stoch ⊆ RS spanned by the vectors of the form w−v for all reactions v → w in our reaction network. This subspace describes the directions in which a solution of the rate equation can move. In our example, it’s spanned by B − 2A and 2A − B, or if you prefer, (−2, 1) and (2, −1). These vectors are linearly dependent, so the stoichiometric subspace has dimension 1. • Deficiency. The deficiency of a reaction network is the number of complexes, minus the number of connected components, minus the dimension of the stoichiometric subspace. In our example there are 2 complexes, 1 connected component, and the dimension of the stoichiometric subspace is 1. So, our reaction network has deficiency 2 - 1 - 1 = 0. So, the deficiency zero theorem applies! What does it say? To understand it, we need a bit more jargon. First of all, a vector x ∈ RS tells us how much we’ve got of each species: the amount of species i ∈ S is the number xi . And then: • Stoichiometric compatibility class. Given a vector v ∈ RS , its stoichiometric compatibility class is the subset of all vectors that we could reach using the reactions in our reaction network: {v + w : w ∈ Stoch} In our example, where the stoichiometric subspace is spanned by (2, −1), the stoichiometric compatibility class of the vector (a, b) is the line consisting of points (x1 , x2 ) = (a, b) + s(2, −1) where the parameter s ranges over all real numbers. Notice that this line can also be written as x1 + 2x2 = c We’ve already seen that if we start with initial conditions on such a line, the solution will stay on this line. And that’s how it always works: as time passes, any solution of the rate equation stays in the same stoichiometric compatibility class! In other words: the stoichiometric subspace is defined by a bunch of linear equations, one for each linear conservation law that all the reactions in our network obey. Here a linear conservation law is a law saying that some linear combination of the numbers of species does not change. Next: • Positivity. A vector in RS is positive if all its components are positive; this describes a container of chemicals where all the species are actually present. The positive stoichiometric compatibility class of x ∈ RS consists of all positive vectors in its stoichiometric compatibility class. 184 18 18.4 EXAMPLE OF THE DEFICIENCY ZERO THEOREM Deficiency zero theorem We finally have enough jargon in our arsenal to state the deficiency zero theorem. We’ll only state the part we need now: Theorem 45 (Deficiency Zero Theorem). If a reaction network is weakly reversible and the rate constants are positive, the rate equation has exactly one equilibrium solution in each positive stoichiometric compatibility class. Any sufficiently nearby solution that starts in the same class will approach this equilibrium as t → +∞. In our example, this theorem says there’s just one positive equilibrium (x1 , x2 ) in each line x1 + 2x2 = c We can find it by setting the time derivatives to zero: d x1 = 2αx2 − 2βx21 = 0 dt d x2 = −αx2 + βx21 = 0 dt Solving these, we get x21 α = x2 β So, these are our equilibrium solutions. It’s easy to verify that indeed, there’s one of these in each stoichiometric compatibility class x1 + 2x2 = c. And the deficiency zero theorem also tells us that any sufficiently nearby solution that starts in the same class will approach this equilibrium as t → ∞. This partially explains what we saw before in our graphs. It shows that in the case α = β = 1, any solution that starts by nearly having x21 =1 x2 will actually have x21 =1 t→+∞ x2 lim But in fact, in this example we don’t even need to start near the equilibrium for our solution to approach the equilibrium! What about in general? For many years this was an open question: Conjecture 46 (Global Attractor Conjecture). If a reaction network is weakly reversible and the rate constants are positive, the rate equation has exactly one equilibrium solution in each positive stoichiometric compatibility class, and any positive solution that starts in the same class will approach this equilibrium as t → +∞. In their groundbreaking 1972 paper [HJ72], Horn and Jackson thought they had proved this, but in 1974 Horn realized they had not: 185 18 18.4 EXAMPLE OF THE DEFICIENCY ZERO THEOREM Deficiency zero theorem [Hor74] Fritz Horn, The dynamics of open reaction systems, in Mathematical Aspects of Chemical and Biochemical Problems and Quantum Chemistry, ed. Donald S. Cohen, SIAM–AMS Proceedings 8, American Mathematical Society, Providence, R.I., 1974, pp. 125–137. It was dubbed the ‘global attractor conjecture’ in the following paper, which proved it in a special case: [CDSS] Gheorghe Craciun, Alicia Dickenstein, Anne Shiu and Bernd Sturmfels, Toric dynamical systems. Available as arXiv:0708.3431. In 2011 Anderson proved the conjecture for reaction networks with a single connected component, or ‘linkage class’: [And] David F. Anderson, A proof of the Global Attractor Conjecture in the single linkage class case. Available as arXiv:1101.0761. In 2015, Craciun proved the conjecture in general: [Cra15] Gheorghe Craciun, Toric differential inclusions and a proof of the Global Attractor Conjecture. Available as arXiv:1501.02860. 186 19 19 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Example of the Anderson–Craciun–Kurtz theorem In Section 18 we started looking at a simple example: a diatomic gas. A diatomic molecule of this gas can break apart into two atoms: A2 → A + A and conversely, two atoms can combine to form a diatomic molecule: A + A → A2 We can draw both these reactions using a chemical reaction network: where we’re writing B instead of A2 to abstract away some detail that’s just distracting here. In Section 18 we looked at the rate equation for this chemical reaction network, and found equilibrium solutions of that equation. Now let’s look at the master equation, and find equilibrium solutions of that. This will illustrate the Anderson–Craciun–Kurtz theorem. We’ll also see how the conservation law we noticed in the last section is related to Noether’s theorem for Markov processes. 19.1 The master equation We’ll start from scratch. The master equation is all about how atoms or molecules or rabbits or wolves or other things interact randomly and turn into other things. So, let’s write ψm,n for the probability that we have m atoms of A and n molecules of B in our container. These probabilities are functions of time, and the master equation will say how they change. First we need to pick a rate constant for each reaction. Let’s say the rate constant for the reaction that produces As is some number α > 0: B →A+A 187 19 19.1 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM The master equation while the rate constant for the reaction that produces Bs is some number β > 0: A+A→B Before we make it pretty using the ideas we’ve been explaining all along, the master equation says: d ψm,n (t) = α(n+1) ψm−2,n+1 − αn ψm,n + β(m+2)(m+1) ψm+2,n−1 − βm(m−1) ψm,n dt Yuck! Normally we don’t show you such nasty equations. Indeed the whole point of our work has been to demonstrate that by packaging the equations in a better way, we can understand them using high-level concepts instead of mucking around with millions of scribbled symbols. But we thought we’d show you what’s secretly lying behind our beautiful abstract formalism, just once. Each term has a meaning. For example, the third one: β(m + 2)(m + 1)ψm+2,n−1 (t) means that the reaction A + A → B will tend to increase the probability of there being m atoms of A and n molecules of B if we start with m + 2 atoms of A and n − 1 molecules of B. This reaction can happen in (m + 2)(m + 1) ways. And it happens at a probabilistic rate proportional to the rate constant for this reaction, β. We won’t go through the rest of the terms. It’s a good exercise to do so, but there could easily be a typo in the formula, since it’s so long and messy. So let us know if you find one! To simplify this mess, the key trick is to introduce a generating function that summarizes all the probabilities in a single power series: X Ψ= ψm,n y m z n m,n≥0 It’s a power series in two variables, y and z, since we have two chemical species: As and Bs. Using this trick, the master equation looks like d Ψ(t) = HΨ(t) dt where the Hamiltonian H is a sum of terms, one for each reaction. This Hamiltonian is built from operators that annihilate and create As and Bs. The annihilation and creation operators for A atoms are: a= ∂ , ∂y a† = y The annihilation operator differentiates our power series with respect to the variable y. The creation operator multiplies it by that variable. Similarly, the annihilation and creation operators for B molecules are: b= ∂ , ∂z 188 b† = z 19 19.2 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Equilibrium solutions In Section 7.3 we explained a recipe that lets us stare at our chemical reaction network and write down this Hamiltonian: 2 2 H = α(a† b − b† b) + β(b† a2 − a† a2 ) As promised, there’s one term for each reaction. But each term is itself a sum of two: one that increases the probability that our container of chemicals will be in a new state, and another that decreases the probability that it’s in its original state. We get a total of four terms, which correspond to the four terms in our previous way of writing the master equation. Problem 31. Show that this way of writing the master equation is equivalent to the previous one. 19.2 Equilibrium solutions Now we will look for all equilibrium solutions of the master equation: in other words, solutions that don’t change with time. So, we’re trying to solve HΨ = 0 Given the rather complicated form of the Hamiltonian, this seems tough. The challenge looks more concrete but even more scary if we go back to our original formulation. We’re looking for probabilities ψm,n , nonnegative numbers that sum to one, such that α(n+1) ψm−2,n+1 − αn ψm,n + β(m+2)(m+1) ψm+2,n−1 − βm(m−1) ψm,n = 0 This equation is horrid! But the good news is that it’s linear, so a linear combination of solutions is again a solution. This lets us simplify the problem using a conserved quantity. Clearly, there’s a quantity that the reactions here don’t change: What’s that? It’s the number of As plus twice the number of Bs. After all, a B can turn into two As, or vice versa. So a B counts twice. Of course the secret reason is that B is a diatomic molecule made of two As. But you’d be able to follow the logic here even if you didn’t know that, just by looking at the chemical reaction network... and sometimes this more abstract approach is handy! Indeed, the way chemists first discovered that certain molecules are made of certain atoms is by seeing which reactions were possible and which weren’t. Suppose we start in a situation where we know for sure that the number of As plus twice the number of Bs equals some number k: ψm,n = 0 unless m + 2n = k 189 19 19.2 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Equilibrium solutions Then we know Ψ is initially of the form X Ψ= ψm,n y m z n m+2n=k But since the number of As plus twice the number of Bs is conserved, if Ψ obeys the master equation it will continue to be of this form! Put a fancier way, we know that if a solution of the master equation starts in this subspace: X Lk = {Ψ : Ψ = ψm,n y m z n for some ψm,n } m+2n=k it will stay in this subspace. So, because the master equation is linear, we can take any solution Ψ and write it as a linear combination of solutions Ψk , one in each subspace Lk . In particular, we can do this for an equilibrium solution Ψ. And then all the solutions Ψk are also equilibrium solutions: they’re linearly independent, so if one of them changed with time, Ψ would too. This means we can just look for equilibrium solutions in the subspaces Lk . If we find these, we can get all equilibrium solutions by taking linear combinations. Once we’ve noticed that, our horrid equation makes a bit more sense: α(n+1) ψm−2,n+1 − αn ψm,n + β(m+2)(m+1) ψm+2,n−1 − βm(m−1) ψm,n = 0 Note that if the pair of subscripts m, n obey m + 2n = k, the same is true for the other pairs of subscripts here! So our equation relates the values of ψm,n for all the points (m, n) with integer coordinates lying on this line segment: m + 2n = k, m, n ≥ 0 You should be visualizing something like this: If you think about it a minute, you’ll see that if we know ψm,n at two points on such a line, we can keep using our equation to recursively work out all the rest. So, there are at most two linearly independent equilibrium solutions of the master equation in each subspace Lk . Why at most two? Why not two? Well, we have to be a bit careful about what happens at the ends of the line segment: remember that ψm,n is defined to 190 19 19.2 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Equilibrium solutions be zero when m or n becomes negative. If we think very hard about this, we’ll see there’s just one linearly independent equilibrium solution of the master equation in each subspace Lk . But this is the sort of nitty-gritty calculation that’s not fun to watch someone else do, so we won’t bore you with that. Soon we’ll move on to a more high-level approach to this problem. But first, one remark. Our horrid equation is like a fancy version of the usual discretized form of the equation d2 ψ =0 dx2 namely: ψn−1 − 2ψn + ψn+1 = 0 And this makes sense, since we get d2 ψ =0 dx2 by taking the heat equation: ∂2ψ ∂ψ = ∂t ∂x2 and assuming ψ doesn’t depend on time. So what we’re doing is a lot like looking for equilibrium solutions of the heat equation. The heat equation describes how heat smears out as little particles of heat randomly move around. True, there don’t really exist ‘little particles of heat’, but this equation also describes the diffusion of any other kind of particles as they randomly move around undergoing Brownian motion. Similarly, our master equation describes a random walk on this line segment: m + 2n = k, m, n ≥ 0 or more precisely, the points on this segment with integer coordinates. The equilibrium solutions arise when the probabilities ψm,n have diffused as much as possible. If you think about it this way, it should be physically obvious that there’s just one linearly independent equilibrium solution of the master equation for each value of k. There’s a general moral here, too, which we’re seeing in a special case: the master equation for a chemical reaction network really describes a bunch of random walks, one for each allowed value of the conserved quantities that can be built as linear combinations of number operators. In our case we have one such conserved quantity, but in general there may be more (or none). Furthermore, these ‘random walks’ are what we’ve been calling Markov processes in Section 10. 191 19 19.3 19.3 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Noether’s theorem Noether’s theorem We simplified our task of finding equilibrium solutions of the master equation by finding a conserved quantity. The idea of simplifying problems using conserved quantities is fundamental to physics: this is why physicists are so enamored with quantities like energy, momentum, angular momentum and so on. Nowadays physicists often use ‘Noether’s theorem’ to get conserved quantities from symmetries. There’s a very simple version of Noether’s theorem for quantum mechanics, but in Section 10 we saw a version for stochastic mechanics, and it’s that version that is relevant now. We don’t really need Noether’s theorem now, since we found the conserved quantity and exploited it without even noticing the symmetry. Nonetheless it’s interesting to see how it relates to what we’re doing. For the reaction we’re looking at now, the idea is that the subspaces Lk are eigenspaces of an operator that commutes with the Hamiltonian H. It follows from standard math that a solution of the master equation that starts in one of these subspaces, stays in that subspace. What is this operator? It’s built from ‘number operators’. The number operator for As is NA = a† a and the number operator for Bs is NB = b† b A little calculation shows NA y m z n = m y m z n , NB y m z n = n y m z n so the eigenvalue of NA is the number of As, while the eigenvalue of NB is the number of Bs. This is why they’re called number operators. As a consequence, the eigenvalue of the operator NA + 2NB is the number of As plus twice the number of Bs: (NA + 2NB ) y m z n = (m + 2n) y m z n Let’s call this operator O, since it’s so important: O = NA + 2NB If you think about it, the spaces Lk we saw a minute ago are precisely the eigenspaces of this operator: Lk = {Ψ : OΨ = kΨ} As we’ve seen, solutions of the master equation that start in one of these eigenspaces will stay there. This lets us take some techniques that are very familiar in quantum mechanics, and apply them to this stochastic situation. 192 19 19.3 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Noether’s theorem First of all, time evolution as described by the master equation is given by the operators exp(tH). In other words, d Ψ(t) = HΨ(t) dt and Ψ(0) = Φ ⇒ Ψ(t) = exp(tH)Φ But if you start in some eigenspace of O, you stay there. Thus if Φ is an eigenvector of O, so is exp(tH)Φ, with the same eigenvalue. In other words, OΦ = kΦ implies O exp(tH)Φ = k exp(tH)Φ = exp(tH)OΦ But since we can choose a basis consisting of eigenvectors of O, we must have O exp(tH) = exp(tH)O or, throwing caution to the winds and differentiating: OH = HO So, as we’d expect from Noether’s theorem, our conserved quantity commutes with the Hamiltonian! This in turn implies that H commutes with any polynomial in O, which in turn suggests that exp(sO)H = H exp(sO) and also exp(sO) exp(tH) = exp(tH) exp(sO) The last equation says that O generates a 1-parameter family of ‘symmetries’: operators exp(sO) that commute with time evolution. But what do these symmetries actually do? Since Oy m z n = (m + 2n)y m z n we have exp(sO)y m z n = es(m+2n) y m z n So, this symmetry takes any probability distribution ψm,n and multiplies it by es(m+2n) . In other words, our symmetry multiplies the relative probability of finding our container of gas in a given state by a factor of es for each A atom, and by a factor of e2s for each B molecule. It might not seem obvious that this operation commutes with time evolution! However, experts on chemical reaction theory are familiar with this fact. Finally, a couple of technical points. Starting where we said “throwing caution to the winds”, our treatment has not been rigorous, since O and H are unbounded operators, and these must be handled with caution. Nonetheless, all the commutation relations we wrote down are true. 193 19 19.4 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM The Anderson–Craciun–Kurtz theorem The operators exp(sO) are unbounded for positive s. They’re bounded for negative s, so they give a 1-parameter semi group of bounded operators. But they’re not stochastic operators: even for s negative, they don’t map probability distributions to probability distributions. However, they do map any nonzero vector Ψ with ψm,n ≥ 0 to a vector exp(sO)Ψ with the same properties. So, we can just normalize this vector and get a probability distribution. The need for this normalization is why we spoke of relative probabilities. 19.4 The Anderson–Craciun–Kurtz theorem Now we’ll actually find all equilibrium solutions of the master equation in closed form. To understand this final section, you really do need to remember some things we’ve discussed earlier. In Section 17 we considered the same chemical reaction network we’re studying now, but we looked at its rate equation, which looks like this: d x1 = 2αx2 − 2βx21 dt d x2 = −αx2 + βx21 dt This describes how the number of As and Bs changes in the limit where there are lots of them and we can treat them as varying continuously, in a deterministic way. The number of As is x1 , and the number of Bs is x2 . We saw that the quantity x1 + 2x2 is conserved, just as now we’ve seen that NA + 2NB is conserved. We saw that the rate equation has one equilibrium solution for each choice of x1 + 2x2 . And we saw that these equilibrium solutions obey x21 α = x2 β The Anderson–Craciun–Kurtz theorem, discussed in Section 8, is a powerful result that gets equilibrium solution of the master equation from equilibrium solutions of the rate equation. It only applies to equilibrium solutions that are ‘complex balanced’, but that’s okay: Problem 32. Show that the equilibrium solutions of the rate equation for the chemical reaction network are complex balanced. 194 19 19.4 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM The Anderson–Craciun–Kurtz theorem So, given any equilibrium solution (x1 , x2 ) of our rate equation, we can hit it with the Anderson–Craciun–Kurtz theorem and get an equilibrium solution of the master equation! And it looks like this: Ψ = e−(x1 +x2 ) X xm xn 1 2 m n y z m!n! m,n≥0 In this solution, the probability distribution ψm,n = e−(x1 +x2 ) n xm 1 x2 m!n! is a product of Poisson distributions. The factor in front is there to make the numbers ψm,n add up to one. And remember, x1 , x2 are any nonnegative numbers with x21 α = x2 β So from all we’ve said, the above formula gives an explicit closed-form solution of the horrid equation α(m+2)(m+1) ψm+2,n−1 − αm(m−1) ψm,n + β(n+1) ψm−2,n+1 − βn ψm,n = 0 That’s pretty nice. We found some solutions without ever doing any nasty calculations. But we’ve really done better than getting some equilibrium solutions of the master equation. By restricting attention to n, m with m + 2n = k, our formula for ψm,n gives an equilibrium solution that lives in the eigenspace Lk : Ψk = e−(x1 +x2 ) X m+2n=k n xm 1 x2 m n y z m!n! And by what we’ve said, linear combinations of these give all equilibrium solutions of the master equation. And we got them with very little work! Despite all our fancy talk, we essentially just took the equilibrium solutions of the rate equation and plugged them into a straightforward formula to get equilibrium solutions of the master equation. This is why the Anderson–Craciun–Kurtz theorem is so nice. And of course we’re looking at a very simple reaction network: for more complicated ones it becomes even better to use this theorem to avoid painful calculations. We could go further. For example, we could study nonequilibrium solutions using Feynman diagrams like this: 195 19 19.4 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM The Anderson–Craciun–Kurtz theorem But instead, we will leave off with another puzzle. We introduced some symmetries, but we haven’t really explored them yet: Problem 33. What do the symmetries associated to the conserved quantity O do to the equilibrium solutions of the master equation given by Ψ = e−(x1 +x2 ) X xm xn 1 2 m n y z m!n! m,n≥0 where (x1 , x2 ) is an equilibrium solution of the rate equation? In other words, what is the significance of the 1-parameter family of solutions exp(sO)Ψ? Also, we used a conceptual argument to check that H commutes with O, but it’s good to know that we can check this sort of thing directly: Problem 34. Compute the commutator [H, O] = HO − OH and show it vanishes. 196 19 19.5 19.5 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Answers Answers The answers to Problems 33 and 34 were written with help from Greg Egan and Arjun Jain, respectively. Problem 33. What do the symmetries associated to the conserved quantity O do to the equilibrium solutions of the master equation given by Ψ = e−(x1 +x2 ) X xm xn 1 2 m n y z m!n! m,n≥0 where (x1 , x2 ) is an equilibrium solution of the rate equation? In other words, what is the significance of the 1-parameter family of solutions exp(sO)Ψ? Answer. The symmetry exp(sO) maps the equilibrium solution of the master equation associated with the solution (x1 , x2 ) of the rate equation to that associated with (es x1 , e2s x2 ). Clearly the equation x21 α = x2 β is still satisfied by the new concentrations x01 = es x1 and x02 = e2s x2 . Indeed, the symmetries exp(sO) are related to a 1-parameter group of symmetries of the rate equation (x1 , x2 ) 7→ (x01 , x02 ) = (es x1 , e2s x2 ) These symmetries map the parabola of equilibrium solutions x21 α = , x2 β x1 , x2 ≥ 0 to itself. For example, if we have an equilibrium solution of the rate equation, we can multiply the number of lone atoms by 1.5 and multiply the number of molecules by 2.25, and get a new solution. What’s surprising is that this symmetry exists even when we consider small numbers of atoms and molecules, where we treat these numbers as integers instead of real numbers. If we have 3 atoms, we can’t multiply the number of atoms by 1.5. So this is a bit shocking at first! The trick is to treat the gas stochastically using the master equation rather than deterministically using the rate equation. What our symmetry does is multiply the relative probability of finding our container of gas in a given state by a factor of es for each lone atom, and by a factor of e2s for each molecule. This symmetry commutes with time evolution as given by the master equation. And for probability distributions that are products of Poisson distributions, this symmetry has the effect of multiplying the mean number of lone atoms by es , and the mean number of molecules by e2s . On the other hand, the symmetry exp(sO) maps each subspace Lk to itself. So this symmetry has the property that if we start in a state with a definite 197 19 19.5 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Answers total number of atoms (that is, lone atoms plus twice the number of molecules), it will map us to another state with the same total number of molecules! And if we start in a state with a definite number of lone atoms and a definite number of molecules, the symmetry will leave this state completely unchanged! These facts sound paradoxical at first, but of course they’re not. They’re just a bit weird. They’re closely related to another weird fact, which however is well-known. If we take a quantum system and start it off in an eigenstate of energy, it will never change, except for an unobservable phase. Every state is a superposition of energy eigenstates. So you might think that nothing can ever change in quantum mechanics. But that’s wrong: the phases that are unobservable in a single energy eigenstate become observable relative phases in a superposition. Indeed the math is exactly the same, except now we’re multiplying relative probabilities by positive real numbers, instead of multiplying amplitudes by complex numbers! Problem 34. Compute the commutator [H, O] = HO − OH and show it vanishes. Answer. One approach is to directly compute the commutator. Recall that 2 2 H = α(a† b − b† b) + β(b† a2 − a† a2 ) and O = a† a + 2b† b. Let us show that the term in H proportional to α commutes with O. For this we need to show that 2 2 (a† b − b† b)(a† a + 2b† b) − (a† a + 2b† b)(a† b − b† b) vanishes. As a and a† commute with b and b† , we can put the a and a† terms in front, obtaining 3 2 2 2 (a† ab − a† ab† b + 2a† bb† b − 2b† bb† b) − (a† a† b − a† ab† b + 2a† b† b2 − 2b† bb† b). The second term cancels the sixth, and the fourth cancels the eighth, leaving 2 2 a† (a† a + 2a† bb† − aa† − 2a† b† b)b The second and fourth terms in the parentheses sum to 2a† , while the first and third terms give −2a† , so we get 0. A similar argument shows that the term in H proportional to β commutes with O. A more generally applicable strategy begins by showing that every formal power series is a (typically infinite) linear combination of vectors in the eigenspaces Lk = {Ψ : OΨ = kΨ} 198 19 19.5 EXAMPLE OF THE ANDERSON–CRACIUN–KURTZ THEOREM Answers and noting that an operator commutes with O if it maps each eigenspace to itself. We can check that [NA , a] = −a, [NA , a† ] = a, [NA , b] = 0, [NA , b† ] = 0, [NB , b] = −b, [NB , b† ] = b, [NB , a] = 0, [NB , a† ] = 0, and use these to show a : Lk → Lk−1 , a† : Lk → Lk+1 , b : Lk → Lk−2 , b† : Lk → Lk+2 . For example, if Ψ ∈ Lk then OΨ = kΨ so ObΨ = (NA + 2NB )bΨ = (bNA + 2(bNB − 1))Ψ = (k − 2)bΨ so bΨ ∈ Lk−2 . Using this, we can see that each term in 2 2 H = α(a† b − b† b) + β(b† a2 − a† a2 ) commutes with O, so [H, O] = 0. For example, consider the first term, involving 2 2 a† b. If Ψ ∈ Lk then bΨ ∈ Lk−2 , so a† bΨ ∈ Lk−1 and a† b ∈ Lk again. Thus 2 αa† b maps each eigenspace of O to itself. Each of the other terms does this as well. 199 20 20 THE DEFICIENCY OF A REACTION NETWORK The deficiency of a reaction network In the last few sections we’ve been talking about ‘reaction networks’, like this: Here A, B, C, D, and E are names of chemicals, and the arrows are chemical reactions. If we know how fast each reaction goes, we can write down a ‘rate equation’ describing how the amount of each chemical changes with time. In Section 17 we met the deficiency zero theorem, a powerful tool for finding equilibrium solutions of the rate equation: in other words, solutions where the amounts of the various chemicals don’t change at all. To apply this theorem, two conditions must hold. Both are quite easy to check: • Your reaction network needs to be ‘weakly reversible’: if you have a reaction that takes one bunch of chemicals to another, there’s a series of reactions that takes that other bunch back to the one you started with. • A number called the ‘deficiency’ that you can compute from your reaction network needs to be zero. The first condition makes a lot of sense, intuitively: you won’t get an equilibrium with all the different chemicals present if some chemicals can turn into others but not the other way around. But the second condition, and indeed the concept of ‘deficiency’, seems mysterious. Luckily, when you work through the proof of the deficiency zero theorem, the mystery evaporates. It turns out that there are two equivalent ways to define the deficiency of a reaction network. One makes it easy to compute, and that’s the definition people usually give. But the other explains why it’s important. In fact the whole proof of the deficiency zero theorem is full of great insights, so we want to show it to you. This will be the climax of the book: we’ll see that Markov processes, Noether’s theorem and Perron–Frobenius theory play crucial roles, even though the deficiency zero theorem concerns not the master equation but the rate equation, which is deterministic and nonlinear! In this section, we’ll just unfold the concept of ‘deficiency’ so we see what it means. In the next, we’ll show you a slick way to write the rate equation, which is crucial to proving the deficiency zero theorem. Then we’ll start the actual proof. 200 20 20.1 20.1 THE DEFICIENCY OF A REACTION NETWORK Reaction networks revisited Reaction networks revisited Let’s recall what a reaction network is, and set up our notation. In chemistry we consider a finite set of ‘species’ like C, O2 , H2 O and so on... and then consider reactions like CH4 + 3O2 −→ CO2 + 2H2 O On each side of this reaction we have a finite linear combination of species, with natural numbers as coefficients. Chemists call such a thing a complex. So, given any finite collection of species, say S, let’s write NS to mean the set of finite linear combinations of elements of S, with natural numbers as coefficients. The complexes appearing in our reactions will form a subset of this, say K. We’ll also consider a finite collection of reactions—or as we’ve been calling them, ‘transitions’. Let’s call this T . Each transition goes from some complex to some other complex: if we want a reaction to be reversible we’ll explicitly include another reaction going the other way. So, given a transition τ ∈ T it will always go from some complex called its source s(τ ) to some complex called its target t(τ ). All this data, put together, is a reaction network: Definition 47. A reaction network (S, s, t : T → K) consists of: • a finite set S of species, • a finite set T of transitions, • a finite set K ⊂ NS of complexes, • source and target maps s, t : T → K. We can draw a reaction network as a graph with complexes as vertices and transitions as edges: The set of species here is S = {A, B, C, D, E}, and the set of complexes is K = {A, B, D, A + C, B + E}. But to write down the ‘rate equation’ describing our chemical reactions, we need a bit more information: constants r(τ ) saying the rate at which each transition τ occurs. So, we define: 201 20 20.1 THE DEFICIENCY OF A REACTION NETWORK Reaction networks revisited Definition 48. A stochastic reaction network is a reaction network (S, s, t : T → K) together with a map r : T → (0, ∞) assigning a rate constant to each transition. Let us remind you how the rate equation works. At any time we have some amount xi ∈ [0, ∞) of each species i. These numbers are the components of a vector x ∈ RS , which of course depends on time. The rate equation says how this vector changes: X dxi = r(τ ) (ti (τ ) − si (τ )) xs(τ ) dt τ ∈T Here we are writing si (τ ) for the ith component of the vector s(τ ), and similarly for ti (τ ). We should remind you what xs(τ ) means, since here we are raising a vector to a vector power, which is a bit unusual. So, suppose we have any vector x = (x1 , . . . , xk ) and we raise it to the power of s = (s1 , . . . , sk ). Then by definition we get xs = xs11 · · · xskk Given this, we hope the rate equation makes intuitive sense! There’s one term for each transition τ . The factor of ti (τ ) − si (τ ) shows up because our transition destroys si (τ ) things of the ith species and creates ti (τ ) of them. The big product s(τ ) s(τ ) xs(τ ) = x1 1 · · · xk k shows up because our transition occurs at a rate proportional to the product of the numbers of things it takes as inputs. The constant of proportionality is the reaction rate r(τ ). The deficiency zero theorem says lots of things, but in the next few episodes we’ll prove a weak version, like this: Theorem 49 (Deficiency Zero Theorem—Baby Version). Suppose we have a weakly reversible reaction network with deficiency zero. Then for any choice of rate constants there exists an equilibrium solution x ∈ (0, ∞)S of the rate equation. In other words: X r(τ ) (ti (τ ) − si (τ )) xs(τ ) = 0 τ ∈T An important feature of this result is that all the components of the vector x are positive. In other words, there’s actually some chemical of each species present! But what do the hypotheses in this theorem mean? A reaction network is weakly reversible if for any transition τ ∈ T going from a complex κ to a complex κ0 , there is a sequence of transitions going from κ0 back to κ. But what about ‘deficiency zero’ ? As we mentioned, this requires more work to understand. So, let’s dive in! 202 20 20.2 20.2 THE DEFICIENCY OF A REACTION NETWORK The concept of deficiency The concept of deficiency In modern math, we like to take all the stuff we’re thinking about and compress it into a diagram with a few objects and maps going between these objects. So, to understand the deficiency zero theorem, we wanted to isolate the crucial maps. For starters, there’s an obvious map Y : K → NS sending each complex to the linear combination of species that it is. We can think of K as an abstract set equipped with this map saying how each complex is made of species, if we like. Then all the information in a stochastic reaction network sits in this diagram: (0, ∞) o r s T t // K Y / NS This is fundamental to everything we’ll do from now on, so take a minute to lock it into your brain. We’ll do lots of different things with this diagram. For example, we often want to use ideas from linear algebra, and then we want our maps to be linear. For example, Y extends uniquely to a linear map Y : RK → RS sending real linear combinations of complexes to real linear combinations of species. Reusing the name Y here won’t cause confusion. We can also extend r, s, and t to linear maps in a unique way, getting a little diagram like this: Ro r RT s t // RK Y / RS Linear algebra lets us talk about differences of complexes. Each transition τ gives a vector ∂τ = t(τ ) − s(τ ) ∈ RK saying the change in complexes that it causes. And we can extend ∂ uniquely to a linear map ∂ : RT → RK defined on linear combinations of transitions. Mathematicians call ∂ a boundary operator. So, we have a little sequence of linear maps: RT ∂ / RK Y / RS This turns a transition into a change in complexes, and then a change in species. If you know fancy math you’ll be wondering if this sequence is a ‘chain complex’, which is a fancy way of saying that Y ∂ = 0. The answer is no. This 203 20 20.2 THE DEFICIENCY OF A REACTION NETWORK The concept of deficiency equation means that every linear combination of reactions leaves the amount of all species unchanged. Or equivalently: every reaction leaves the amount of all species unchanged. This only happens in very silly examples. Nonetheless, it’s possible for a linear combination of reactions to leave the amount of all species unchanged. For example, this will happen if we have a linear combination of reactions that leaves the amount of all complexes unchanged. But this sufficient condition is not necessary. And this leads us to the concept of ‘deficiency zero’: Definition 50. A reaction network has deficiency zero if any linear combination of reactions that leaves the amount of every species unchanged also leaves the amount of every complex unchanged. In short, a reaction network has deficiency zero iff Y (∂ρ) = 0 ⇒ ∂ρ = 0 for every ρ ∈ RT . In other words—using some basic concepts from linear algebra—a reaction network has deficiency zero iff Y is one-to-one when restricted to the image of ∂. Remember, the image of ∂ is im∂ = {∂ρ : ρ ∈ RT } Roughly speaking, this consists of all changes in complexes that can occur due to reactions. In still other words, a reaction network has deficiency zero if 0 is the only vector in both the image of ∂ and the kernel of Y : im∂ ∩ kerY = {0} Remember, the kernel of Y is kerY = {φ ∈ RK : Y φ = 0} Roughly speaking, this consists of all changes in complexes that don’t cause changes in species. So, ‘deficiency zero’ roughly says that if a reaction causes a change in complexes, it causes a change in species. (All this ‘roughly speaking’ stuff is because in reality we should be talking about linear combinations of transitions, complexes and species. But it’s a bit distracting to keep saying that when we’re trying to get across the basic ideas!) Now we’re ready to understand deficiencies other than zero, at least a little. They’re defined like this: Definition 51. The deficiency of a reaction network is the dimension of im∂ ∩ kerY . 204 20 20.3 20.3 THE DEFICIENCY OF A REACTION NETWORK How to compute the deficiency How to compute the deficiency You can compute the deficiency of a reaction network just by looking at it. However, it takes a little training. First, remember that a reaction network can be drawn as a graph with complexes as vertices and transitions as edges, like this: There are three important numbers we can extract from this graph: • We can count the number of vertices in this graph; let’s call that |K|, since it’s just the number of complexes. • We can count the number of pieces or ‘components’ of this graph; let’s call that #components for now. • We can also count the dimension of the image of Y ∂. This space, imY ∂, is called the stoichiometric subspace: vectors in here are changes in species that can be accomplished by transitions in our reaction network, or linear combinations of transitions. These three numbers, all rather easy to compute, let us calculate the deficiency: Theorem 52. The deficiency of a reaction network equals |K| − #components − dim(imY ∂) Proof. By definition we have deficiency = dim(im∂ ∩ kerY ) but another way to say this is deficiency = dim(kerY |im∂ ) where we are restricting Y to the subspace im∂, and taking the dimension of the kernel of that. The rank-nullity theorem says that whenever you have a linear map T : V → W between finite-dimensional vector spaces, dim(kerT ) = dim(dom) − dim(imT ) 205 20 20.4 THE DEFICIENCY OF A REACTION NETWORK Examples where domT is the domain of T , namely the vector space V . It follows that dim(kerY |im∂ ) = dim(domY |im∂ ) − dim(imY |im∂ ) The domain of Y |im∂ is just im∂, while its image equals im(Y ∂), so deficiency = dim(im∂) − dim(imY ∂) The theorem then follows from the following lemma. Lemma 53. dim(im∂) = |K| − #components. Proof. In fact this holds whenever we have a finite set of complexes and a finite set of transitions going between them. We get a diagram T s t // K We can extend the source and target functions to linear maps as usual: RT s t // RK and then we can define ∂ = t − s. We claim that dim(im∂) = |K| − #components where #components is the number of connected components of the graph with K as vertices and T as edges. This is easiest to see using an inductive argument where we start by throwing out all the edges of our graph and then put them back in one at a time. When our graph has no edges, ∂ = 0 and the number of components is |K|, so our claim holds: both sides of the equation above are zero. Then, each time we put in an edge, there are two choices: either it goes between two different components of the graph we have built so far, or it doesn’t. If it goes between two different components, it increases dim(im∂) by 1 and decreases the number of components by 1, so our equation continues to hold. If it doesn’t, neither of these quantities change, so our equation again continues to hold. 20.4 Examples 206 20 20.4 THE DEFICIENCY OF A REACTION NETWORK Examples This reaction network is not weakly reversible, since we can get from B + E and A + C to D but not back. It becomes weakly reversible if we throw in another transition: Taking a reaction network and throwing in the reverse of an existing transition never changes the number of complexes, the number of components, or the dimension of the stoichiometric subspace. So, the deficiency of the reaction network remains unchanged. We computed it back in Section 17.2, but let’s do it again. For either reaction network above: • the number of complexes is 5: |K| = |{A, B, D, A + C, B + E}| = 5 • the number of components is 2: #components = 2 • the dimension of the stoichometric subspace is 3. For this we go through each transition, see what change in species it causes, and take the vector space spanned by these changes. Then we find a basis for this space and count it: imY ∂ = hB − A, A − B, D − A − C, D − (B + E), (B + E) − (A + C)i = hB − A, D − A − C, D − A − Ei so dim (imY ∂) = 3 As a result, the deficiency is zero: deficiency = |K| − #components − dim (imY ∂) = 5−2−3 = 0 Now let’s throw in another complex and some more transitions: 207 20 20.4 THE DEFICIENCY OF A REACTION NETWORK Examples Now: • the number of complexes increases by 1: |K| = |{A, B, D, E, A + C, B + E}| = 6 • the number of components is unchanged: #components = 2 • the dimension of the stoichometric subspace increases by 1. We never need to include reverses of transitions when computing this: imY ∂ = hB − A, E − B, D − (A + C), D − (B + E), (B + E) − (A + C)i = hB − A, E − B, D − A − C, D − B − Ei so dim (imY ∂) = 4 As a result, the deficiency is still zero: deficiency = |K| − #components − dim (imY ∂) = 6−2−4 = 0 Do all reaction networks have deficiency zero? That would be nice. Let’s try one more example: Now: • The number of complexes is the same as in our last example: |K| = |{A, B, D, E, A + C, B + E}| = 6 • the number of components is also the same: #components = 2 • the dimension of the stoichometric subspace is also the same: imY ∂ = hB − A, D − (A + C), D − (B + E), (B + E) − (A + C), (B + E) − Bi = hB − A, D − A − C, D − B, Ei so dim (imY ∂) = 4 208 20 20.5 THE DEFICIENCY OF A REACTION NETWORK Different kinds of graphs So the deficiency is still zero: deficiency = |K| − #components − dim (imY ∂) = 6−2−4 = 0 It’s sure easy to find examples with deficiency zero! Problem 35. Can you find a reaction network where the deficiency is not zero? Problem 36. If you can’t find an example, prove the deficiency is always zero. If you can, find 1) the smallest example and 2) the smallest example that actually arises in chemistry. Note that not all reaction networks can actually arise in chemistry. For example, the transition A → A+A would violate conservation of mass. Nonetheless a reaction network like this might be useful in a very simple model of amoeba reproduction, one that doesn’t take limited resources into account. As you might recall, we talked about amoebas in Section 6. 20.5 Different kinds of graphs We’ll conclude with some technical remarks that only a mathematician could love; feel free to skip them if you’re not in the mood. As you’ve already seen, it’s important that a reaction network can be drawn as a graph: But there are many kinds of graph. What kind of graph is this, exactly? As mentioned in Section 17.1, it’s a directed multigraph, meaning that the edges have arrows on them, more than one edge can go from one vertex to another, and we can have any number of edges going from a vertex to itself. Not all those features are present in this example, but they’re certainly allowed by our definition of reaction network! After all, we’ve seen that a stochastic reaction network amounts to a little diagram like this: (0, ∞) o r s T t // K Y / NS If we throw out the rate constants, we’re left with a reaction network. So, a reaction network is just a diagram like this: T s t // K 209 Y / NS 20 20.5 THE DEFICIENCY OF A REACTION NETWORK Different kinds of graphs If we throw out the information about how complexes are made of species, we’re left with an even smaller diagram: T s t // K And this precisely matches the slick definition of a directed multigraph: it’s a set E of edges, a set V of vertices, and functions s, t : E → V saying where each edge starts and where it ends: its source and target. Why don’t we just call this a ‘graph’ ? Why the fancy term ‘directed multigraph’ ? The problem is that there are many kinds of graphs. While their diversity is a bit annoying at first, we must learn to love it, at least if we’re mathematicians and want everything to be completely precise. Indeed, there are at least 23 = 8 kinds of graphs, depending on our answers to three questions: • Do the edges have arrows on them? • Can more than one edge go between a specific pair of vertices? and • Can an edge go from a vertex to itself? We get directed multigraphs if we answer YES, YES and YES to these questions. Since they allow all three features, directed multigraphs are very important. For example, a ‘category’ is a directed multigraph equipped with some extra structure. Also, what mathematicians call a ‘quiver’ is just another name for a directed multigraph. We’ve met two other kinds of graph so far: • In Section 14.2 and Section Section 15 we studied graph Laplacians and circuits made of resistors—or in other words, Dirichlet operators—using ‘simple graphs’. We get simple graphs when we answer NO, NO and NO to the three questions above. The slick definition of a simple graph is that it’s a set V of vertices together with a subset E of the collection of 2-element subsets of V . • In Section 16.5 we studied Markov processes on finite sets—or, in other words, infinitesimal stochastic operators—using ‘directed graphs’. We get directed graphs when we answer YES, NO and YES to the three questions. The slick definition of a directed graph is that it’s a set V of vertices together with a subset E of the ordered pairs of V : E ⊆V ×V There is a lot to say about this business, but for now we’ll just note that you can use directed multigraphs with edges labelled by positive numbers to describe 210 20 20.6 THE DEFICIENCY OF A REACTION NETWORK Answers Markov processes, just as we used directed graphs. You don’t get anything more general, though! After all, if we have multiple edges from one vertex to another, we can replace them by a single edge as long as we add up their rate constants. And an edge from a vertex to itself has no effect at all. In short, both for Markov processes and reaction networks, we can take ‘graph’ to mean either ‘directed graph’ or ‘directed multigraph’, as long as we make some minor adjustments. In what follows, we’ll use directed multigraphs for both Markov processes and reaction networks. One tiny advantage is that you can take a directed multigraph and add a new edge from one vertex to another without first checking to see if there’s already one there. Combining our work with category theory will also go easier if we use directed multigraphs. 20.6 Answers Problem 35. Can you find a reaction network where the deficiency is not zero? Answer. The reaction network A→A+A→A+A+A has deficiency 1. If we go on this way adding multiples of A we can keep on increasing the deficiency, as only the number of complexes increases. Problem 36. Find 1) the smallest example of a reaction network where the deficiency is nonzero and 2) the smallest example that actually arises in chemistry. Answer. The answers here depend on the notion of ‘smallest’. However, the reaction network in the answer above has the smallest possible number of components, namely one—and among those, it has the smallest possible number of complexes, namely 3, since a reaction network with just one component and one or two complexes has deficiency zero. The reaction network above cannot occur in chemistry, except as a simplified model that omits some of the species involved in the reactions, because it violates conservation of mass. Here is a rather small reaction network that can occur in chemistry: A→B A+C →B+C This arises when a molecule A can turn into a molecule B either with or without the assistance of a catalyst C. The number of complexes is 4, the number of components is 2, and the dimension of the stochiometric subspace is 1, so the deficienty is 4 − 2 − 1 = 1. 211 21 21 REWRITING THE RATE EQUATION Rewriting the rate equation Okay, now let’s dig deeper into the proof of the deficiency zero theorem. We’re only going to prove a baby version, at first. Later we can enhance it: Theorem 54 (Deficiency Zero Theorem—Baby Version). Suppose we have a weakly reversible reaction network with deficiency zero. Then for any choice of rate constants there exists an equilibrium solution of the rate equation where all species are present in nonzero amounts. The first step is to write down the rate equation in a new, more conceptual way. It’s incredibly cool. We’ve mentioned Schrödinger’s equation, which describes the motion of a quantum particle: dψ = −iHψ dt We’ve spent a lot of time on Markov processes, which describe the motion of a ‘stochastic’ particle: dψ = Hψ dt A ‘stochastic’ particle is one that’s carrying out a random walk, and now ψ describes its probability to be somewhere, instead of its amplitude. But now comes a surprise: soon we’ll see that the rate equation for a reaction network looks somewhat similar: dx = Y HxY dt where Y is some matrix, and xY is defined using a new thing called ‘matrix exponentiation’, which makes the equation nonlinear! If you’re reading this you probably know how to multiply a vector by a matrix. But if you’re like us, you’ve never seen anyone take a vector and raise it to the power of some matrix! We’ll explain it, don’t worry... right now we’re just trying to get you intrigued. It’s not complicated, but it’s exciting how this unusual operation shows up naturally in chemistry. Since we’re looking for an equilibrium solution of the rate equation, we actually want to solve dx =0 dt or in other words Y HxY = 0 In fact we will do better: we will find a solution of HxY = 0 And we’ll do this in two stages: 212 21 21.1 REWRITING THE RATE EQUATION The rate equation and matrix exponentiation • First we’ll find all solutions of Hψ = 0 This equation is linear, so it’s easy to understand. • Then, among these solutions ψ, we’ll find one that also obeys ψ = xY This is a nonlinear problem involving matrix exponentiation, but still, we can do it, using a clever trick called ‘logarithms’. Putting the pieces together, we get our solution of HxY = 0 and thus our equilibrium solution of the rate equation: dx = Y HxY = 0 dt That’s a rough outline of the plan. But now let’s get started, because the details are actually fascinating. In this section we’ll just show you how to rewrite the rate equation in this new way. 21.1 The rate equation and matrix exponentiation Remember how the rate equation goes. We start with a stochastic reaction network, meaning a little diagram like this: (0, ∞) o r s T t // K Y / NS This contains quite a bit of information: • a finite set T of transitions, • a finite set K of complexes, • a finite set S of species, • a map r : T → (0, ∞) giving a rate constant for each transition, • source and target maps s, t : T → K saying where each transition starts and ends, • a one-to-one map Y : K → NS saying how each complex is made of species. 213 21 21.1 REWRITING THE RATE EQUATION The rate equation and matrix exponentiation Given all this, the rate equation says how the amount of each species changes with time. We describe these amounts with a vector x ∈ [0, ∞)S . So, we want a differential equation filling in the question marks here: dx =??? dt Now in Section 20.1, we thought of K as a subset of NS , and thus of the vector space RS . Back then, we wrote the rate equation as follows: X dx = r(τ ) (t(τ ) − s(τ )) xs(τ ) dt τ ∈T where vector exponentiation is defined by xs = xs11 · · · xskk when x and s are vectors in RS . However, we’ve now switched to thinking of our set of complexes K as a set in its own right that is mapped into NS by Y . This is good for lots of reasons, like defining the concept of ‘deficiency’, which we did last time. But it means the rate equation above doesn’t quite parse anymore! Things like s(τ ) and t(τ ) live in K; we need to explicitly convert them into elements of RS using Y for our equation to make sense! So now we have to write the rate equation like this: X dx =Y r(τ ) (t(τ ) − s(τ )) xY s(τ ) dt τ ∈T This looks more ugly, but if you’ve got even one mathematical bone in your body, you can already see vague glimmerings of how we’ll rewrite this the way we want: dx = Y HxY dt Here’s how. First, we extend our maps s, t and Y to linear maps between vector spaces: RT s t // RK Y / RS Then, we put an inner product on the vector spaces RT , RK and RS . For RK we do this in the most obvious way, by letting the complexes be an orthonormal basis. So, given two complexes κ, κ0 , we define their inner product by hκ, κ0 i = δκ,κ0 We do the same for RS . But for RT we define the inner product in a more clever way involving the rate constants. If τ, τ 0 ∈ T are two transitions, we define their inner product by: 1 hτ, τ 0 i = δτ,τ 0 r(τ ) 214 21 21.1 REWRITING THE RATE EQUATION The rate equation and matrix exponentiation Having put inner products on these three adjoints of the linear maps between them, to other way: ∂ / Y RT o RK o † s Y † vector spaces, we can take the get linear maps going back the / RS These are defined in the usual way—though we’re using daggers here the way physicists do, where mathematicians would prefer to see stars. For example, s† : RK → RT is defined by the relation hs† φ, ψi = hφ, sψi and so on. Next, we set up a random walk on the set of complexes. Remember, our reaction network is a graph with complexes as vertices and transitions as edges, like this: Each transition τ has a number attached to it: the rate constant r(τ ). So, we can randomly hop from complex to complex along these transitions, with probabilities per unit time described by these numbers. The probability of being at some particular complex will then be described by a function ψ: K → R which also depends on time, and changes according to the equation dψ = Hψ dt for some Hamiltonian H : RK → RK This Hamiltonian can be described as a matrix with off-diagonal entries with X i 6= j ⇒ Hij = r(τ ) τ : j→i where we write τ : j → i when τ is a transition with source j and target i. Since H is infinitesimal stochastic, the diagonal entries are determined by the requirement that the columns sum to zero. 215 21 21.1 REWRITING THE RATE EQUATION The rate equation and matrix exponentiation In fact, there’s a very slick formula for this Hamiltonian: H = ∂s† We’ll prove this in Section 22.3. For now, the main point is that with this Hamiltonian, the rate equation is equivalent to this: dx = Y HxY dt The only thing not defined yet is the funny exponential xY . That’s what makes the equation nonlinear. We’re taking a vector to the power of a matrix and getting a vector. This sounds weird—but it actually makes sense! It only makes sense because we have chosen bases for our vector spaces. To understand it, let’s number our species 1, . . . , k as we’ve been doing all along, and number our complexes 1, . . . , `. Our linear map Y : RK → RS then becomes a k × ` matrix of natural numbers. Its entries say how many times each species shows up in each complex: Y11 Y12 · · · Y1` Y21 Y22 · · · Y2` Y = . .. .. . . . . . . . Yk1 Yk2 · · · Yk` The entry Yij says how many times the ith species shows up in the jth complex. Now, let’s be a bit daring and think of the vector x ∈ RS as a row vector with k entries: x = (x1 , x2 , . . . , xk ) Then we can multiply x on the right by the matrix Y and get a vector in RK : Y11 Y12 · · · Y1` Y21 Y22 · · · Y2` xY = (x1 , x2 , . . . , xk ) . .. .. .. .. . . . Yk1 Yk2 · · · Yk` = (x1 Y11 + · · · + xk Yk1 , . . . , x1 Y1` + · · · + xk Yk` ) So far, no big deal. But now you’re ready to see the definition of xY , which is very similar: xY Y11 Y21 .. . Y12 Y22 .. . ··· ··· .. . Y1` Y2` .. . Yk1 Yk2 ··· Yk` = (x1 , x2 , . . . , xk ) = (xY1 11 · · · xYk k1 , . . . , xY1 1` · · · xYk k` ) 216 21 21.1 REWRITING THE RATE EQUATION The rate equation and matrix exponentiation It’s exactly the same, but with multiplication replacing addition, and exponentiation replacing multiplication! Apparently our class on matrices stopped too soon: we learned about matrix multiplication, but matrix exponentiation is also worthwhile. What’s the point of it? Well, suppose you have a certain number of hydrogen molecules, a certain number of oxygen molecules, a certain number of water molecules, and so on—a certain number of things of each species. You can list these numbers and get a vector x ∈ RS . Then the components of xY describe how many ways you can build up each complex from the things you have. For example, xY1 11 xY2 21 · · · xYk k1 say roughly how many ways you can build complex 1 by picking Y11 things of species 1, Y21 things of species 2, and so on. Why ‘roughly’ ? Well, we’re pretending we can pick the same thing twice. So if we have 4 water molecules and we need to pick 3, this formula gives 43 . The right answer is 4 × 3 × 2. To get this answer we’d need to use the ‘falling power’ 43 = 4 × 3 × 2, as explained in Section 5.7. But the rate equation describes chemistry in the limit where we have lots of things of each species. In this limit, the ordinary power becomes a good approximation. Problem 37. In this section we’ve seen a vector raised to a matrix power, which is a vector, and also a vector raised to a vector power, which is a number. How are they related? There’s more to say about this. But let’s get to the punchline: Theorem 55. The rate equation: X dx =Y r(τ ) (t(τ ) − s(τ )) xY s(τ ) dt τ ∈T is equivalent to this equation: dx = Y ∂s† xY dt or in other words: dx = Y HxY dt where H = ∂s† . Proof. Since ∂ = t − s, it is enough to show X (t − s)s† xY = r(τ ) (t(τ ) − s(τ )) xY s(τ ) τ ∈T So, we’ll compute (t − s)s† xY , and think about the meaning of each quantity we get en route. 217 21 21.1 REWRITING THE RATE EQUATION The rate equation and matrix exponentiation We start with x ∈ RS . This is a list of numbers saying how many things of each species we have: our raw ingredients, as it were. Then we compute xY = (xY1 11 · · · xYk k1 , . . . , xY1 1` · · · xYk k` ) This is a vector in RK . It’s a list of numbers saying how many ways we can build each complex starting from our raw ingredients. Alternatively, we can write this vector xY as a sum over basis vectors: X xY = xY1 1κ · · · xYk kκ κ κ∈K Next let’s apply s† to this. We claim that X s† κ = r(τ ) τ τ : s(τ )=κ In other words, we claim s† κ is the sum of all the transitions having κ as their source, weighted by their rate constants! To prove this claim, it’s enough to take the inner product of each side with any transition τ 0 , and check that we get the same answer. For the left side we get hs† κ, τ 0 i = hκ, s(τ 0 )i = δκ,s(τ 0 ) To compute the right side, we need to use the cleverly chosen inner product on RT . Here we get * + X X 0 r(τ )τ, τ = δτ,τ 0 = δκ,s(τ 0 ) τ : s(τ )=κ τ : s(τ )=κ In the first step here, the factor of 1/r(τ ) in the cleverly chosen inner product canceled the visible factor of r(τ ). For the second step, you just need to think for half a minute—or ten, depending on how much coffee you’ve had. Either way, we conclude that indeed X s† κ = r(τ )τ τ : s(τ )=κ Next let’s combine this with our formula for xY : X xY = xY1 1κ · · · xYk kκ κ κ∈K We get this: s† xY = X r(τ ) xY1 1κ · · · xYk kκ τ κ,τ : s(τ )=κ In other words, s† xY is a linear combination of transitions, where each one is weighted both by the rate it happens and how many ways it can happen starting with our raw ingredients. 218 21 21.2 REWRITING THE RATE EQUATION References Our goal is to compute (t − s)s† xY . We’re almost there. Remember, s says which complex is the input of a given transition, and t says which complex is the output. So, (t − s)s† xY says the total rate at which complexes are created and/or destroyed starting with the species in x as our raw ingredients. That sounds good. But let’s just pedantically check that everything works. Applying t − s to both sides of our last equation, we get X (t − s)s† xY = r(τ ) xY1 1κ · · · xYk kκ (t(τ ) − s(τ )) κ,τ : s(τ )=κ Remember, our goal was to prove that this equals X r(τ ) (t(τ ) − s(τ )) xY s(τ ) τ ∈T But if you stare at these a while and think, you’ll see they’re equal. All this seems peculiar at first, but ultimately it all makes sense. The interesting subtlety is that we use the linear map called ‘multiplying by Y ’: RK ψ → 7 → RS Yψ to take a bunch of complexes and work out the species they contain, while we use the nonlinear map called ‘raising to the Y th power’: RS x → RK 7 → xY to take a bunch of species and work out how many ways we can build each complex from them. There is much more to say about this: for example, these maps arise from a pair of what category theorists call ‘adjoint functors’. But we’re worn out and you probably are too. 21.2 References We found this thesis to be the most helpful reference when trying to understand the proof of the deficiency zero theorem: [Gub03] Jonathan M. Guberman, Mass Action Reaction Networks and the Deficiency Zero Theorem, B.A. thesis, Department of Mathematics, Harvard University, 2003. We urge you to check it out. In particular, Section 2 and Appendix A discuss matrix exponentiation. Here’s another good modern treatment of the deficiency zero theorem: [Gun03] Jeremy Gunawardena, Chemical reaction network theory for in-silico biologists, 2003. 219 21 21.2 REWRITING THE RATE EQUATION References Some of the theorem was first proved here: [Fei87] Martin Feinberg, Chemical reaction network structure and the stability of complex isothermal reactors: I. The deficiency zero and deficiency one theorems, Chemical Engineering Science 42 (1987), 2229-2268. However, Feinberg’s treatment here relies heavily on this paper: [HJ72] F. Horn and R. Jackson, General mass action kinetics, Archive for Rational Mechanics and Analysis 47 (1972), 81-116. These lectures are also very helpful: [Fei79] Martin Feinberg, Lectures on Reaction Networks, 1979. 220 22 22 MARKOV PROCESSES The rate equation and Markov processes We’ve been looking at reaction networks, and we’re getting ready to find equilibrium solutions of the equations they give. To do this, we’ll need to connect them to another kind of network we’ve studied. A reaction network is something like this: It’s a bunch of complexes, which are sums of basic building-blocks called species, together with arrows called transitions going between the complexes. If we know a number for each transition describing the rate at which it occurs, we get an equation called the ‘rate equation’. This describes how the amount of each species changes with time. We’ve been talking about this equation ever since the start of this course! In Section 21, we wrote it down in a new very compact form: dx = Y HxY dt Here x is a vector whose components are the amounts of each species, while H and Y are certain matrices. But now suppose we forget how each complex is made of species! Suppose we just think of them as abstract things in their own right, like numbered boxes: 221 22 MARKOV PROCESSES We can use these boxes to describe states of some system. The arrows still describe transitions, but now we think of these as ways for the system to hop from one state to another. Say we know a number for each transition describing the probability per time at which it occurs: Then we get a ‘Markov process’—or in other words, a random walk where our system hops from one state to another. If ψ is the probability distribution saying how likely the system is to be in each state, this Markov process is described by this equation: dψ = Hψ dt This is simpler than the rate equation, because it’s linear. But the matrix H is the same—we’ll see that explicitly later on. 222 22 22.1 MARKOV PROCESSES The Markov process of a weighted graph What’s the point? Well, our ultimate goal is to prove the deficiency zero theorem, which gives equilibrium solutions of the rate equation. That means finding x with Y HxY = 0 Now we’ll find all equilibria for the Markov process, meaning all ψ with Hψ = 0 In the next section we’ll show some of these have the form ψ = xY So, we’ll get HxY = 0 and thus Y HxY = 0 as desired! So, let’s get to to work. 22.1 The Markov process of a weighted graph We’ve been looking at stochastic reaction networks, which are things like this: (0, ∞) o r T s t // K Y / NS However, we can build a Markov process starting from just part of this information: s / r /K (0, ∞) o T t Let’s call this thing a weighted graph. We’ve been calling the things in K ‘complexes’, but now we’ll think of them as ‘states’. So: Definition 56. A weighted graph consists of: • a finite set of states K, • a finite set of transitions T , • a map r : T → (0, ∞) giving a rate constant for each transition, • source and target maps s, t : T → K saying where each transition starts and ends. Graph theorists would call the states vertices, the transitions edges and the rate constants weights. However, graph theorists might let their weights be other kinds of numbers, not necessarily positive real numbers. 223 22 22.2 MARKOV PROCESSES Equilibrium solutions of the master equation Anyway, starting from a weighted graph, we can get a Markov process describing how a probability distribution ψ on our set of states will change with time. As usual, this Markov process is described by a master equation: dψ = Hψ dt for some Hamiltonian: H : RK → RK What is this Hamiltonian, exactly? Let’s think of it as a matrix where Hij is the probability per time for our system to hop from the state j to the state i. This looks backwards, but don’t blame me—blame the guys who invented the usual conventions for matrix algebra. Clearly if i 6= j this probability per time should be the sum of the rate constants of all transitions going from j to i: X i 6= j ⇒ Hij = r(τ ) τ : j→i where we write τ : j → i when τ is a transition with source j and target i. Now, we saw way back in Section 4.4 that for a probability distribution to remain a probability distribution as it evolves in time according to the master equation, we need H to be infinitesimal stochastic: its off-diagonal entries must be nonnegative, and the sum of the entries in each column must be zero. The first condition holds already, and the second one tells us what the diagonal entries must be. So, we’re basically done describing H. But we can summarize it this way: Problem 38. Think of RK as the vector space consisting of finite linear combinations of elements κ ∈ K. Then show X Hκ = r(τ )(t(τ ) − s(τ )) s(τ )=κ Check directly that H is infinitesimal stochastic. Not only can we get an infinitesimal stochastic operator and thus a Markov process from any weighted graph, the process can also be reversed! Problem 39. Suppose K is any finite set and suppose H is an infinitesimal stochastic operator on RK . Show that H comes from some weighted graph having K as its set of states. Furthermore, show that this graph is unique (up to isomorphism) if we require that it has most one transition from any state i ∈ K to any state j ∈ K. 22.2 Equilibrium solutions of the master equation Now we’ll classify equilibrium solutions of the master equation, meaning ψ ∈ RK with Hψ = 0 224 22 22.2 MARKOV PROCESSES Equilibrium solutions of the master equation We’ll do only do this when our weighted graph is ‘weakly reversible’. This concept doesn’t actually depend on the rate constants, so let’s be general and say: Definition 57. A graph is weakly reversible if for every edge τ : i → j, there is directed path going back from j to i, meaning that we have edges τ1 : j → j1 , τ2 : j1 → j2 , ..., τn : jn−1 → i This weighted graph is not weakly reversible: but this one is: The good thing about the weakly reversible case is that we get one equilibrium solution of the master equation for each component of our graph, and all equilibrium solutions are linear combinations of these. This is not true in general! For example, this guy is not weakly reversible: 225 22 22.2 MARKOV PROCESSES Equilibrium solutions of the master equation It has only one component, but the master equation has two linearly independent equilibrium solutions: one that vanishes except at the state 0, and one that vanishes except at the state 2. The idea of a ‘component’ is supposed to be fairly intuitive—our graph falls apart into pieces called components—but we should make it precise. As explained in Section 12, the graphs we’re using here are directed multigraphs, meaning things like s, t : E → V where E is the set of edges (our transitions) and V is the set of vertices (our states). There are actually two famous concepts of ‘component’ for graphs of this sort: ‘strongly connected’ components and ‘connected’ components. We only need connected components, but let us explain both concepts, in a doubtless futile attempt to slake your insatiable thirst for knowledge. Two vertices i and j of a graph lie in the same strongly connected component iff you can find a directed path of edges from i to j and also one from j back to i. Remember, a directed path from i to j looks like this: i→a→b→c→j Here’s a path from x to y that is not directed: i→a←b→c→j and we hope you can write down the obvious but tedious definition of an ‘undirected path’, meaning a path made of edges that don’t necessarily point in the correct direction. Given that, we say two vertices i and j lie in the same connected component iff you can find an undirected path going from i to j. In this case, there will automatically also be an undirected path going from j to i. For example, i and j lie in the same connected component here, but not the same strongly connected component: i→a←b→c→j Here’s a graph with one connected component and 3 strongly connected components, which are marked in blue: 226 22 22.2 MARKOV PROCESSES Equilibrium solutions of the master equation For the theory we’re looking at now, we only care about connected components, not strongly connected components! However: Problem 40. Show that for weakly reversible graphs, the connected components are the same as the strongly connected components. With these definitions out of the way, we can state this theorem on equilibria for Markov processes: Theorem 58. Suppose H is the Hamiltonian of a weakly reversible weighted graph: s / r /K (0, ∞) o T t Then for each connected component C ⊆ K, there exists a unique probability distribution ψC ∈ RK that is positive on that component, zero elsewhere, and is an equilibrium solution of the master equation: HψC = 0 Moreover, these probability distributions ψC form a basis for the space of equilibrium solutions of the master equation. So, the dimension of this space is the number of components of K. Proof. We start by assuming our graph has one connected component. We use the Perron–Frobenius theorem, as explained in Section 16.6. This applies to ‘nonnegative’ matrices, meaning those whose entries are all nonnegative. That is not true of H itself, but only its diagonal entries can be negative, so if we choose a large enough number c > 0, H + cI will be nonnegative. Since our graph is weakly reversible and has one connected component, it follows straight from the definitions that the operator H + cI will also be ‘irreducible’ in the sense of Section 16.5. The Perron–Frobenius theorem then swings into action, and we instantly conclude several things. First, H + cI has a positive real eigenvalue r such that any other eigenvalue, possibly complex, has absolute value ≤ r. Second, there is an eigenvector ψ with eigenvalue r and all positive components. Third, any other eigenvector with eigenvalue r is a scalar multiple of ψ. Subtracting c, it follows that λ = r −c is the eigenvalue of H with the largest real part. We have Hψ = λψ, and any other vector with this property is a scalar multiple of ψ. We can show that in fact λ = 0. To do this we copy the argument in Theorem 37. First, since ψ is positive we can normalize it to be a probability distribution: X ψi = 1 i∈K Since H is infinitesimal stochastic, exp(tH) sends probability distributions to probability distributions: X (exp(tH)ψ)i = 1 i∈K 227 22 22.2 MARKOV PROCESSES Equilibrium solutions of the master equation for all t ≥ 0. On the other hand, X X (exp(tH)ψ)i = etλ ψi = etλ i∈K i∈K so we must have λ = 0. We conclude that when our graph has one connected component, there is a probability distribution ψ ∈ RK that is positive everywhere and has Hψ = 0. Moreover, any φ ∈ RK with Hφ = 0 is a scalar multiple of ψ. When our graph is weakly reversible but has several components, the matrix H is block diagonal, with one block for each component. So, we can run the above argument on each component C ⊆ K and get a probability distribution ψC ∈ RK that is positive on C. We can then check that HψC = 0 and that every φ ∈ RK with Hφ = 0 can be expressed as a linear combination of these probability distributions ψC in a unique way. By the way: even in the fully general case, where our weighted graph may not be weakly reversible, we can completely understand all equilibrium solutions of the master equation. We do not need this to prove the deficiency zero theorem. So, only read the rest of this section if you want to become an expert on Markov processes. Suppose we have any weighted graph. Write its set K of vertices as a disjoint union of strongly connected components: K= n [ Ci . i=1 Say i ≤ j if there is a directed path from some vertex in Ci to some vertex in Cj . (This is equivalent to there being a directed path from any vertex in Ci to any vertex in Cj .) If i ≤ j, you can show that probability distributions that are nonzero only on vertices in Ci will become nonzero on Cj as you evolve them using the master equation. In other words, probability will flow from the ith strongly connected component to the jth. We say a strongly connected component Ci is terminal if i ≤ j implies i = j. The idea is that probability cannot flow out of a terminal component. Using this idea, you can classify equilibrium solutions of the master equation for any weighted graph: Problem 41. Suppose H is the Hamiltonian of a weighted graph: (0, ∞) o r T s t // K Prove that for each terminal strongly connected component of this graph, there is a unique probability distribution ψ ∈ RK that is positive on that component, zero elsewhere, and is an equilibrium solution of the master equation: Hψ = 0. 228 22 22.3 MARKOV PROCESSES The Hamiltonian, revisited Prove that probability distributions of this sort are a basis for the space of equilibrium solutions of the master equation. So, the dimension of this space is the number of terminal strongly connected components of K. 22.3 The Hamiltonian, revisited One last small piece of business: in Section 21 we mentioned very slick formula for the Hamiltonian of a weighted graph: H = ∂s† We’d like to prove it agrees with the formula we derived in Problem 38: X Hκ = r(τ )(t(τ ) − s(τ )) s(τ )=κ for any complex κ, regarded as one of th standard basis vectors in RK . Recall the setup. We start with any weighted graph: (0, ∞) o r s T t // K We extend s and t to linear maps between vector spaces: s RT t // RK We define the boundary operator by: ∂ =t−s Then we put an inner product on the vector spaces RT and RK . For RK we let the elements of K be an orthonormal basis, but for RT we define the inner product in a more clever way involving the rate constants: hτ, τ 0 i = 1 δτ,τ 0 r(τ ) where τ, τ 0 ∈ T . These inner products let us define adjoints of the maps s, t and ∂, via formulas like this: hs† φ, ψi = hφ, sψi Then we claim: Theorem 59. The operator H = ∂s† has Hκ = X r(τ )(t(τ ) − s(τ )) s(τ )=κ for any κ ∈ K. 229 22 22.3 MARKOV PROCESSES The Hamiltonian, revisited Proof. First we claim that X s† κ = r(τ ) τ τ : s(τ )=κ where we write τ ∈ T for the corresponding standard basis vector of RT . To prove this, it’s enough to check that taking the inner products of either side with any basis vector τ 0 , we get results that agree. On the one hand: hτ 0 , s† κi = hsτ 0 , κi = δs(τ 0 ),κ On the other hand, using our clever inner product on RT we see that: X X hτ 0 , r(τ ) τ i = r(τ ) hτ 0 , τ i τ : s(τ )=κ τ : s(τ )=κ X = δτ 0 ,τ τ : s(τ )=κ = δs(τ 0 ),κ where the factor of 1/r(τ ) in the inner product on RT cancels the visible factor of r(τ ). So indeed the results match. Using this formula for s† κ we now see that Hκ = ∂s† κ = ∂ X r(τ ) τ τ : s(τ )=κ X = r(τ ) (t(τ ) − s(τ )) τ : s(τ )=κ which is precisely what we want. We hope you see through the formulas to their intuitive meaning. As usual, the formulas are just a way of precisely saying something that makes plenty of sense. If κ is some state of our Markov process, s† κ is the sum of all transitions starting at this state, weighted by their rates. Applying ∂ to a transition tells us what change in state it causes. So ∂s† κ tells us the rate at which things change when we start in the state κ. That’s why ∂s† is the Hamiltonian for our Markov process. After all, the Hamiltonian tells us how things change: dψ = Hψ dt Okay, we’ve got all the machinery in place. Next we’ll prove the deficiency zero theorem! 230 23 23 PROOF OF THE DEFICIENCY ZERO THEOREM Proof of the deficiency zero theorem Now we’ve reached the climax of our story so far: we’re ready to prove the deficiency zero theorem. First let’s talk about it informally a bit. Then we’ll review the notation, and then—hang on to your seat!—we’ll give the proof. The crucial trick is to relate a bunch of chemical reactions, described by a ‘reaction network’ like this: to a simpler problem where a system randomly hops between states arranged in the same pattern: Tihs is sort of amazing, because we’ve thrown out lots of detail. It’s also amazing because this simpler problem is linear. In the original problem, the chance that a reaction turns a B +E into a D is proportional to the number of B’s times the number of E’s. That’s nonlinear! But in the simplified problem, the chance that your system hops from state 4 to state 3 is just proportional to the probability that it’s in state 4 to begin with. That’s linear. 231 23 PROOF OF THE DEFICIENCY ZERO THEOREM The wonderful thing is that, at least under some conditions, we can an find an equilibrium for our chemical reactions starting from an equilibrium solution of the simpler problem. Let’s roughly sketch how it works, and where we are so far. Our simplified problem is described by an equation like this: d ψ = Hψ dt where ψ is a function that the probability of being in each state, and H describes the probability per time of hopping from one state to another. We can easily understand quite a lot about the equilibrium solutions, where ψ doesn’t change at all: Hψ = 0 because this is a linear equation. We did this in Theorem 58. Of course, when we say ‘easily’, that’s a relative thing: we needed to use the Perron–Frobenius theorem, which we introduced in Section 16. But that’s a well-known theorem in linear algebra, and it’s easy to apply here. In Theorem 55, we saw that the original problem was described by an equation like this, called the ‘rate equation’: dx = Y HxY dt Here x is a vector whose entries describe the amount of each kind of chemical: the amount of A’s, the amount of B’s, and so on. The matrix H is the same as in the simplified problem, but Y is a matrix that says how many times each chemical shows up in each spot in our reaction network: The key thing to notice is xY , where we take a vector and raise it to the power of a matrix. We explained this operation back in Section 21.1. It’s this operation that says how many B + E pairs we have, for example, given the number of B’s and the number of E’s. It’s this that makes the rate equation nonlinear. Now, we’re looking for equilibrium solutions of the rate equation, where the rate of change is zero: Y HxY = 0 232 23 PROOF OF THE DEFICIENCY ZERO THEOREM But in fact we’ll do even better! We’ll find solutions of this: HxY = 0 And we’ll get these by taking our solutions of this: Hψ = 0 and adjusting them so that ψ = xY while ψ remains a solution of Hψ = 0. But: how do we do this ‘adjusting’ ? That’s the crux of the whole business! That’s what we’ll do now. Remember, ψ is a function that gives a probability for each ‘state’, or numbered box here: The picture here consists of two pieces, called ‘connected components’: the piece containing boxes 0 and 1, and the piece containing boxes 2, 3 and 4. It turns out that we can multiply ψ by a function that’s constant on each connected component, and if we had Hψ = 0 to begin with, that will still be true afterward. The reason is that there’s no way for ψ to ‘leak across’ from one component to another. It’s like having water in separate buckets. You can increase the amount of water in one bucket, and decrease it in another, and as long as the water’s surface remains flat in each bucket, the whole situation remains in equilibrium. That’s sort of obvious. What’s not obvious is that we can adjust ψ this way so as to ensure ψ = xY for some x. 233 23 23.1 PROOF OF THE DEFICIENCY ZERO THEOREM Review And indeed, it’s not always true! It’s only true if our reaction network obeys a special condition. It needs to have ‘deficiency zero’. We defined this concept back in Section 20, but now we’ll finally use it for something. It turns out to be precisely the right condition to guarantee we can tweak any function on our set of states, multiplying it by a function that’s constant on each connected component, and get a new function ψ with ψ = xY When all is said and done, that is the key to the deficiency zero theorem. 23.1 Review The battle is almost upon us—we’ve got one last chance to review our notation. We start with a stochastic reaction network: (0, ∞) o r T s t // K / NS Y This consists of: • finite sets of transitions T , complexes K and species S, • a map r : T → (0, ∞) giving a rate constant for each transition, • source and target maps s, t : T → K saying where each transition starts and ends, • a one-to-one map Y : K → NS saying how each complex is made of species. Then we extend s, t and Y to linear maps: RT s t // RK Y / RS Then we put inner products on these vector spaces as described in Section 21.1, which lets us ‘turn around’ the maps s and t by taking their adjoints: s† , t† : RK → RT More surprisingly, we can ‘turn around’ Y and get a nonlinear map using ‘matrix exponentiation’: RS → RK x 7→ xY This is most easily understood by thinking of x as a row vector and Y as a matrix: Y11 Y12 · · · Y1` Y21 Y22 · · · Y2` .. .. .. .. . . . . Yk1 Yk2 · · · Yk` x1 , x2 , . . . , xk xY = = xY1 11 · · · xYk k1 , . . . , xY1 1` · · · xYk k` 234 23 23.1 PROOF OF THE DEFICIENCY ZERO THEOREM Review Remember, complexes are made out of species. The matrix entry Yij says how many things of the jth species there are in a complex of the ith kind. If ψ ∈ RK says how many complexes there are of each kind, Y ψ ∈ RS says how many things there are of each species. Conversely, if x ∈ RS says how many things there are of each species, xY ∈ RK says how many ways we can build each kind of complex from them. So, we get these maps: RT s† u // R K s i t u (·)Y Y / RS t† Next, the boundary operator ∂ : R T → RK describes how each transition causes a change in complexes: ∂ =t−s As we saw in Theorem 59, there is a Hamiltonian H : RK → RK describing a Markov processes on the set of complexes, given by H = ∂s† But the star of the show is the rate equation. This describes how the number of things of each species changes with time. We write these numbers in a list and get a vector x ∈ RS with nonnegative components. The rate equation says: dx = Y HxY dt We can read this as follows: • x says how many things of each species we have now. • xY says how many complexes of each kind we can build from these species. • s† xY says how many transitions of each kind can originate starting from these complexes, with each transition weighted by its rate. • HxY = ∂s† xY is the rate of change of the number of complexes of each kind, due to these transitions. • Y HxY is the rate of change of the number of things of each species. 235 23 23.2 PROOF OF THE DEFICIENCY ZERO THEOREM The proof 23.2 The proof We are looking for equilibrium solutions of the rate equation, where the number of things of each species doesn’t change at all: Y HxY = 0 In fact we will find complex balanced equilibrium solutions, where even the number of complexes of each kind doesn’t change: HxY = 0 More precisely, we have: Theorem 60 (Deficiency Zero Theorem—Child’s Version). Suppose we have a reaction network obeying these two conditions: • It is weakly reversible, meaning that whenever there’s a transition from one complex κ to another κ0 , there’s a directed path of transitions going back from κ0 to κ. • It has deficiency zero, meaning im∂ ∩ kerY = {0}. Then for any choice of rate constants there exists a complex balanced equilibrium solution of the rate equation where all species are present in nonzero amounts. In other words, there exists x ∈ RS with all components positive and such that: HxY = 0 Proof. Because our reaction network is weakly reversible, Theorem 58 there exists ψ ∈ (0, ∞)K with Hψ = 0 This ψ may not be of the form xY , but we shall adjust ψ so that it becomes of this form, while still remaining a solution of Hψ = 0. To do this, we need a couple of lemmas: Lemma 61. ker∂ † + imY † = RK . Proof. We need to use a few facts from linear algebra. If V is a finite-dimensional vector space with inner product, the orthogonal complement L⊥ of a subspace L ⊆ V consists of vectors that are orthogonal to everything in L: L⊥ = {v ∈ V : ∀w ∈ L hv, wi = 0} We have (L ∩ M )⊥ = L⊥ + M ⊥ where L and M are subspaces of V and + denotes the sum of two subspaces: that is, the smallest subspace containing both. Also, if T : V → W is a linear map between finite-dimensional vector spaces with inner product, we have (kerT )⊥ = imT † 236 23 23.2 PROOF OF THE DEFICIENCY ZERO THEOREM The proof and (imT )⊥ = kerT † Now, because our reaction network has deficiency zero, we know that im∂ ∩ kerY = {0} Taking the orthogonal complement of both sides, we get (im∂ ∩ kerY )⊥ = RK and using the rules we mentioned, we obtain ker∂ † + imY † = RK as desired. Now, given a vector φ in RK or RS with all positive components, we can define the logarithm of such a vector, componentwise: (ln φ)i = ln(φi ) Similarly, for any vector φ in either of these spaces, we can define its exponential in a componentwise way: (exp φ)i = exp(φi ) These operations are inverse to each other. Moreover: Lemma 62. The nonlinear operator RS x → RK 7 → xY RS x → 7 → is related to the linear operator RK Y †x by the formula xY = exp(Y † ln x) which holds for all x ∈ (0, ∞)S . Proof. A straightforward calculation. By the way, this formula would look a bit nicer if we treated ln x as a row vector and multiplied it on the right by Y : then we would have xY = exp((ln x)Y ) The problem is that we are following the usual convention of multiplying vectors by matrices on the left, yet writing the matrix on the right in xY . Taking the transpose Y † of the matrix Y serves to compensate for this. 237 23 23.2 PROOF OF THE DEFICIENCY ZERO THEOREM The proof Now, given our vector ψ ∈ (0, ∞)K with Hψ = 0, we can take its logarithm and get ln ψ ∈ RK . Lemma 61 says that RK = ker∂ † + imY † so we can write ln ψ = α + Y † β where α ∈ ker∂ † and β ∈ RS . Moreover, we can write β = ln x for some x ∈ (0, ∞)S , so that ln ψ = α + Y † (ln x) Exponentiating both sides componentwise, we get ψ = exp(α) exp(Y † (ln x)) where at right we are taking the componentwise product of vectors. Thanks to Lemma 62, we conclude that ψ = exp(α)xY as desired. So, we have taken ψ and almost written it in the form xY —but not quite! We can adjust ψ to make it be of this form: exp(−α)ψ = xY Clearly all the components of exp(−α)ψ are positive, since the same is true for both ψ and exp(−α). So, the only remaining task is to check that H(exp(−α)ψ) = 0 We do this using two lemmas: Lemma 63. If Hψ = 0 and α ∈ ker∂ † , then H(exp(−α)ψ) = 0. Proof. It is enough to check that multiplication by exp(−α) commutes with the Hamiltonian H, since then H(exp(−α)ψ) = exp(−α)Hψ = 0 Recall from Section 22.1 that H is the Hamiltonian of a Markov process associated to this ‘weighted graph’: (0, ∞) o r T s t As noted here: 238 // K 23 23.2 PROOF OF THE DEFICIENCY ZERO THEOREM The proof [BF] John Baez and Brendan Fong, A Noether theorem for Markov processes, J. Math. Phys. 54, 013301, 2013. Also available as arXiv:1203.2035. multiplication by some function on K commutes with H if and only if that function is constant on each connected component of this graph. Such functions are called conserved quantities. So, it suffices to show that exp(−α) is constant on each connected component. For this, it is enough to show that α itself is constant on each connected component. But this will follow from the next lemma, since α ∈ ker∂ † . Lemma 64. A function α ∈ RK is a conserved quantity iff ∂ † α = 0. In other words, α is constant on each connected component of the graph s, t : T → K iff ∂ † α = 0. Proof. Suppose ∂ † α = 0, or in other words, α ∈ ker∂ † , or in still other words, α ∈ (im∂)⊥ . To show that α is constant on each connected component, it suffices to show that whenever we have two complexes connected by a transition, like this: τ : κ → κ0 then α takes the same value at both these complexes: ακ = ακ0 To see this, note ∂τ = t(τ ) − s(τ ) = κ0 − κ and since α ∈ (im∂)⊥ , we conclude hα, κ0 − κi = 0 But calculating this inner product, we see ακ0 − ακ = 0 as desired. For the converse, we simply turn the argument around: if α is constant on each connected component, we see hα, κ0 − κi = 0 whenever there is a transition τ : κ → κ0 . It follows that hα, ∂τ i = 0 for every transition τ , so α ∈ (im∂)⊥ . And thus concludes the proof of the lemma! And thus concludes the proof of the theorem! Of course there is much more to say, but we’ll stop here—except for a few final tidbits. 239 23 23.2 PROOF OF THE DEFICIENCY ZERO THEOREM The proof FURTHER DIRECTIONS There’s a lot we didn’t get off our chests yet. And we thought: perhaps you want to say something too? Because we’re so nice, we’re going to tell you about a few special tidbits we’ve been saving. We think they represent good starting points for further research. The entire subject of stochastic mechanics is really just getting started! So please let us know if you take it further in any direction. In Chapter 24 we present some further thoughts on Noether’s theorem relating symmetries and conserved quantities. In Chapter 10 we saw that Noether’s theorem looks a bit more complicated in stochastic mechanics than in quantum mechanics. But in Chapters 15 and 16 we saw that Dirichlet operators can serve as Hamiltonians for both quantum and stochastic mechanics. This suggests that the stochastic version of Noether’s theorem may become more beautiful when the Hamiltonian is a Dirichlet operator. And indeed this is true! Our presentation centers around a result of this sort, proved with the help of Ville Bergholm. We find it tantalizing that this result looks so much like the quantum version of Noether’s theorem, yet its proof is different. Hopefully this will shed more light on the differences and connections between stochastic and quantum mechanics. We conclude Chapter 24 with a series of problems for further study. Chapter 25 discusses various connections between Petri nets and computation. Because of their versatility and simplicity, Petri nets have long been used as models of computation. In Sections 25.1 and 25.2 we discuss the ‘reachability problem’: the question of which collections of molecules can turn into which other collections, given a fixed set of chemical reactions. In Section 25.2 we explain how the reachability problem for Petri nets is really a problem in category theory. More recently, molecular biologists have used stochastic Petri nets as a model of ‘chemical computers’. Section 25.3, based on a talk by David Soloveichik [Sol14], reviews a bit of this recent work. Finally, Section 25.5 lists some free software for working with Petri nets and reaction networks. We thank Jim Stuttard for making this possible. 240 24 24 NOETHER’S THEOREM FOR DIRICHLET OPERATORS Noether’s theorem for Dirichlet operators In 1915, Emmy Noether showed that the symmetries of a physical system give conserved quantities, and vice versa. Her result has become a staple of modern physics, so it is often known as Noether’s theorem, though in fact she proved more than one important theorem—even on this subject! Though she introduced it in classical mechanics, this theorem and its generalizations have found particularly wide use in quantum mechanics. In Chapters 10 and 12 we introduced a version for stochastic mechanics and compared it to the quantum version. The stochastic version was different than what would be expected from the quantum one, so it is interesting to try to figure out why. Together with Ville Bergholm, we’ve been trying to get to the bottom of these differences. This chapter is based on an article by Bergholm on the Azimuth blog. We thank him for writing this article and working with us on the result presented here: a Noether’s theorem applicable to stochastic mechanics when the Hamiltonian is a Dirichlet operator. As you may recall from Chapters 15 and 16, Dirichlet operators lie in the intersection of stochastic and quantum mechanics. That is, they can be used to generate both Markov semigroups and 1-parameter unitary groups. So, they’re a subclass of the infinitesimal stochastic operators considered in the stochastic version of Noether’s theorem. The extra structure of Dirichlet operators—compared with the wider class of infinitesimal stochastic operators—provides a handle for us to dig a little deeper into understanding the intersection of these two theories. We hope the problems at the end of this chapter hint at some directions for future work. Before we get into the details, let’s recall how conservation laws work in quantum mechanics, and then contrast this with what was shown for stochastic systems in Chapter 10. (For a more detailed comparison between the stochastic and quantum versions of Noether’s theorem, see Chapter 12.) 241 24 24.1 24.1 NOETHER’S THEOREM FOR DIRICHLET OPERATORS The quantum case The quantum case Let’s review the basics of quantum theory, to ensure that everything we need is right here in front of us. In a quantum system with n states, a state ψ is a unit vector in Cn . The unitary time evolution of a state ψ(t) ∈ Cn is generated by a self-adjoint n × n matrix H called the Hamiltonian. In other words, ψ(t) satisfies Schrödinger’s equation i~ d ψ(t) = Hψ(t). dt Thus, the state of a system starting off at time zero in the state ψ0 and evolving for a time t is given by ψ(t) = e−itH ψ0 . The observable properties of a quantum system are associated with self-adjoint operators. In the state ψ ∈ Cn , the expected value of the observable associated to a self-adjoint operator O is hψ, Oψi This expected value is constant in time for all states if and only if O commutes with the Hamiltonian H: [O, H] = 0 ⇐⇒ d hOiψ(t) = 0 ∀ψ0 , ∀t dt In this case we say O is a conserved quantity. The equivalence of the two conditions above is exactly the quantum version of Noether’s theorem, as you’ll recall from Theorem 21 in Chapter 12. 24.2 The stochastic case In stochastic mechanics, the story changes. Now a state ψ is a probability distribution: a vector with n nonnegative components that sum to 1. Schrödinger’s equation gets replaced by the master equation: d ψ(t) = Hψ(t) dt If we start with a probability distribution ψ0 at time zero and evolve it according to this equation, at any later time we have ψ(t) = etH ψ0 We want this always to be a probability distribution. To ensure this, the Hamiltonian H should be infinitesimal stochastic: that is, a real-valued n × n matrix where the off-diagonal entries are nonnegative and the entries of each column sum to zero. It no longer needs to be self-adjoint! When H is infinitesimal stochastic, the operators etH map the set of probability distributions to itself whenever t ≥ 0, and we call this family of operators a continuous-time Markov process, or more precisely a Markov semigroup. 242 24 24.2 NOETHER’S THEOREM FOR DIRICHLET OPERATORS The stochastic case In stochastic mechanics, we say an observable O is a diagonal real n × n matrix. This lets us make sense of the commutator [O, H], which is crucial in the stochastic version of Noether’s theorem. However, we can also create a b ∈ Rn whose components are the diagonal entries of O: vector O bi = Oii O This lets us write the expected value of the observable I in the state ψ as b ψi hO, We saw this formula long ago, at the very end of Section 11.2. Soon we will finally use it to do something interesting! Here is the stochastic version of Noether’s theorem. We will state it in a slightly different way than before, but it is really just Theorem 6: Theorem 65 (Noether’s Theorem for Markov Processes). Suppose H is an infinitesimal stochastic matrix and O is an observable (a real diagonal matrix). Then [O, H] = 0 if and only if d hÔ, ψ(t)i = 0 dt and d c2 hO , ψ(t)i = 0 dt for all t ≥ 0 and all ψ(t) obeying the master equation. c2 is the vector whose components are the diagonal entries of O2 . Here O So, just as in quantum mechanics, whenever [O, H] = 0 the expected value of O will be conserved: d hO, ψ(t)i = 0 dt for any ψ0 and all t ≥ 0. However, in Chapter 10 we already saw that—unlike in quantum mechanics—you need more than just the expectation value of the observable O to be constant to obtain the equation [O, H] = 0. You really need both d hÔ, ψ(t)i = 0 dt together with d c2 hO , ψ(t)i = 0 dt for all initial states ψ0 to be sure that [O, H] = 0. So, symmetries and conserved quantities have a rather different relationship than they do in quantum theory. 243 24 24.3 24.3 NOETHER’S THEOREM FOR DIRICHLET OPERATORS A Noether theorem for Dirichlet operators A Noether theorem for Dirichlet operators Remember in Chapter 15 where we studied when the infinitesimal generator of our Markov semigroup is also self-adjoint? We’re going to think about that scenario here too. In other words, what if H is both an infinitesimal stochastic matrix but also its own transpose. Then it’s called a Dirichlet operator... and in this case, we get a stochastic version of Noether’s theorem that more closely resembles the usual quantum one. Theorem 66 (Noether’s Theorem for Dirichlet Operators). Suppose H is a Dirichlet operator and O is an observable (a real diagonal matrix). Then d b hO, ψ(t)i = 0 dt for all t ≥ 0 and all ψ(t) obeying the master equation. [O, H] = 0 ⇐⇒ Proof. Reminiscent of the general proof appearing in Chapter 10, the ⇒ direction is straightforward, and it follows from Theorem 65. The point is to show the ⇐ direction. Since H is self-adjoint, we’re allowed to use a spectral decomposition: X H= Ek p k k where we choose an orthonormal basis of eigenvectors for H, say φk , with Hφk = Ek φk and let pk be the projection operator onto the eigenvector φk : pk ψ = hφk , ψi φk For any stochastic state ψ0 let ψ(t) = etH ψ0 be the result of evolving it according to the master equation. Then d hÔ, ψ(t)i = 0 dt ∀ψ0 , ∀t ≥ 0 ⇐⇒ hÔ, HetH ψ0 i = 0 ∀ψ0 , ∀t ≥ 0 ⇐⇒ hHetH Ô, ψ0 i = 0 ∀ψ0 , ∀t ≥ 0 ⇐⇒ HetH Ô = 0 ∀t ≥ 0 ⇐⇒ X ∀t ≥ 0 Ek etEk hφk , Ôiφk = 0 k ⇐⇒ Ek etEk hφk , Ôi = 0 ⇐⇒ Ô ∈ Span{φk : Ek = 0} = ker H 244 ∀k, ∀t ≥ 0 24 24.4 NOETHER’S THEOREM FOR DIRICHLET OPERATORS Problems where the next to last equivalence is due to the vectors φk being linearly independent. For any infinitesimal stochastic operator H the corresponding weighted graph consists of m connected components for some number m. Thus, we can reorder (permute) the vertices of this graph such that H becomes block-diagonal with m blocks. We saw in Theorem 58 that the kernel of H is spanned by m eigenvectors φk , one for each block. Since H is also symmetric, the components of each such φk can be chosen to be ones within the block and zeros outside it. Consequently Ô ∈ ker H implies that we can choose the basis of eigenvectors of O to be the vectors φk , which implies [O, H] = 0 In summary, by restricting ourselves to the intersection of quantum and stochastic generators, we have found a version of Noether’s theorem for stochastic mechanics that looks formally just like the quantum version! We conclude with a few problems about stochastic processes. 24.4 Problems In this section, we’ve considered a stochastic variant of Noether’s theorem for Dirchlet operators. Dirchlet operators turn out to be a subclass of a wider class of operators, which generate so-called ‘doubly stochastic processes’. We introduce these in the next four problems. We then leave the class of doubly stochastic processes and consider general stochastic process and the equations of motion for possibly time-dependent observables. From this we arrive at a stochastic variant of Ehrenfest’s theorem from quantum mechanics. Readers are asked to prove this stochastic variant by completing two final problems. First, recall from Section 22.1 how any weighted graph gives an infinitesimal stochastic operator called its Hamiltonian, H. In Problem 39 you saw that conversely, any infinitesimal stochastic operator comes from a weighted graph in this way. Problem 42. Fix a weighted graph and let H be its Hamiltonian. Define the indegree of a vertex to be the sum of the rate constants of edges having that vertex as their target. Define the outdegree to be the sum of the rate constants of edges having that vertex as their source. Show that if H is a Dirichlet operator, the indegree and outdegree of any vertex of our graph must be equal. Weighted graphs with this property are called balanced. Problem 43. An n × n matrix U is a doubly stochastic operator if it is stochastic and its transpose is also stochastic. In other words, each column and 245 24 24.4 NOETHER’S THEOREM FOR DIRICHLET OPERATORS Problems each row of U sums to 1: X Uij = i X Uij = 1 j and all its matrix entries are nonnegative. Prove that U is doubly stochastic if and only if is stochastic and U ψ = ψ where ψ is the ‘flat’ probability distribution, with ψi = n1 for all i. Problem 44. Suppose that U (t) = etH is a doubly stochastic semigroup, meaning that U (t) is doubly stochastic for all t ≥ 0. Prove that the Hamiltonian H is infinitesimal doubly stochastic, meaning that X X Hij = Hij = 0 i j and all the off-diagonal entries of H are nonnegative. Also prove the converse: if H is infinitesimal doubly stochastic, etH is a doubly stochastic semigroup. Problem 45. Prove that any infinitesimal doubly stochastic operator H is the Hamiltonian of a balanced graph. Problem 46. Show that every Dirichlet operator is doubly infinitesimal stochastic. Problem 47. Find a doubly infinitesimal stochastic operator that is not a Dirichlet operator. Problem 48. Find necessary and sufficient conditions for the Hamiltonian of a weighted graph to be a Dirichlet operator. The next two problems concern the Ehrenfest theorem for time-dependent observables. We ask the reader to first prove the usual quantum version, and then a stochastic analogue: Problem 49. Let O(t) be a possibly time-dependent quantum observable, and let ψ(t) be a quantum state evolving according to Schrödinger’s equation with some self-adjoint Hamiltonian H. Prove the Ehrenfest theorem: d ∂O(t) hO(t)iψ(t) = −i h[O(t), H]iψ(t) + dt ∂t ψ(t) Here we are using the abbreviation hAiψ = hψ, Aψi for the expected value of any operator A in the quantum state ψ. Problem 50. Let O(t) be a possibly time-dependent stochastic observable, and let ψ(t) be a stochastuc state evolving according to the master equation. Show that d ∂O(t) hO(t)iψ(t) = h[O(t), H]iψ(t) + dt ∂t ψ(t) 246 24 24.4 NOETHER’S THEOREM FOR DIRICHLET OPERATORS Problems Here we are using the abbreviation hAiψ = hÂ, ψi for the expected value of any operator A in the stochastic state ψ. 247 25 25 COMPUTATION AND PETRI NETS Computation and Petri nets As we’ve seen, a stochastic Petri net can be used to describe a bunch of chemical reactions with certain reaction rates. We could try to use these reactions to build a ‘chemical computer’. But how powerful can such a computer be? People know quite a lot about the answer to this question. But before people got interested in stochastic Petri nets, computer scientists spent quite some time studying plain old Petri nets, which don’t include the information about reaction rates. They used these as simple models of computation. And since computer scientists like to know which questions are decidable by means of an algorithm and which aren’t, they proved some interesting theorems about decidability for Petri nets. Let’s talk about ‘reachability’: the question of which collections of molecules can turn into which other collections, given a fixed set of chemical reactions. For example, suppose you have these chemical reactions: C + O2 −→ CO2 CO2 + NaOH −→ NaHCO3 NaHCO2 + HCl −→ H2 O + NaCl + CO2 Can you use these to turn C + O2 + NaOH + HCl into CO2 + H2 O + NaCl? It’s not hard to settle this particular question—we’ll do it soon. But settling all possible questions of this type turns out to be very hard. 25.1 The reachability problem Remember: Definition 67. A Petri net consists of a set S of species and a set T of transitions, together with a function i: S × T → N saying how many copies of each state shows up as input for each transition, and a function o : S×T →N saying how many times it shows up as output. Let’s assume both S and T are finite. We hope you remember that we draw the species as yellow circles and the transitions as aqua boxes, so the chemical reactions mentioned before give this Petri net: 248 25 25.1 COMPUTATION AND PETRI NETS The reachability problem A ‘complex’ amounts to a way of putting dots in the yellow circles. In chemistry this says how many molecules we have of each kind. Here’s an example: This complex happens to have just zero or one dot in each circle, but that’s not required: we could have any number of dots in each circle. So, mathematically, a complex is a finite linear combination of species, with natural numbers as coefficients: an element of NS . In this particular example it’s C + O2 + NaOH + HCl Given two complexes, we say the second is reachable from the first if, loosely speaking, we can get from the first from the second by a finite sequence of transitions. In chemistry, this means that we are using the reactions described by our Petri net to turn one collection of molecules into another, with no byproducts or leftover raw ingredients. For example, earlier we asked if you can get from the complex we just mentioned to this one: CO2 + H2 O + NaCl And the answer is yes! We can do it with this sequence of transitions: 249 25 25.1 COMPUTATION AND PETRI NETS The reachability problem Now, the reachability problem asks: given a Petri net and two complexes, is one reachable from the other? If the answer is ‘yes’, of course you can show that by an exhaustive search of all possibilities. But if the answer is ‘no’, how can you be sure? It is not obvious, in general. Back in the 1970’s, computer scientists felt this problem should be decidable by some algorithm... but they had a lot of trouble finding such an algorithm. In 1976, Roger J. Lipton showed that if such an algorithm existed, it would need to take at least an exponential amount of memory space and consequently an exponential amount of time to run: [Lip76] Roger J. Lipton, The reachability problem requires exponential space, Technical Report 62, Yale University, 1976. This means that most computer scientists would consider any algorithm to solve the reachability problem ‘infeasible’, since they prefer polynomial time algorithms. On the bright side, it means that Petri nets might be fairly powerful when viewed as computers themselves! After all, for a universal Turing machine, the analogue of the reachability problem is undecidable. So if the reachability problem for Petri nets were decidable, they couldn’t serve as universal computers. But if it were decidable but hard, Petri nets might be fairly powerful—though still not universal—computers. At an important computer science conference in 1977, two researchers presented a proof that the reachability problem was decidable: [ST77] S. Sacerdote and R. Tenney, The decidability of the reachability problem for vector addition systems, Conference Record of the Ninth Annual ACM Symposium on Theory of Computing, 2-4 May 1977, Boulder, Colorado, USA, ACM, 1977, pp. 61–76. However, their proof turned out to be flawed! You can read the story here: 250 25 25.1 COMPUTATION AND PETRI NETS The reachability problem [Pet81] James L. Peterson, Petri Net Theory and the Modeling of Systems, Prentice– Hall, New Jersey, 1981. This is a very nice introduction to early work on Petri nets and decidability. Peterson had an interesting idea, too: he thought there might be a connection between Petri nets and something called ‘Presburger arithmetic’. He gave some evidence, and suggested using this to settle the decidability of the reachability problem. Let’s explain why this is so interesting. Presburger arithmetic is a simple set of axioms for the arithmetic of natural numbers, much weaker than Peano arithmetic. Peano arithmetic is a set of axioms involving 0, 1, × and + and the principle of mathematical induction. Presburger arithmetic is similar, but it doesn’t mention multiplication. The only axioms are these: • x + 0 = x. • x + (y + 1) = (x + y) + 1. • There is no x such that 0 = x + 1. • If x + 1 = y + 1 then x = y. • The principle of mathematical induction. This is an ‘axiom schema’: for any formula P written in the language of Presburger arithmetic, we have an axiom saying that P (0) and ∀x(P (x) ⇒ P (x + 1)) imply ∀xP (x). Unlike Peano arithmetic, Presburger arithmetic is complete: for every statement, either it or its negation is provable. And unlike Peano arithmetic, Presburger arithmetic is decidable: you can write an algorithm that decides for any statement in Presburger arithmetic whether it or its negation is provable. However, any such algorithm must be very slow! In 1974, Fischer and Rabin showed that any decision algorithm for Presburger arithmetic has a worst-case runtime of at least pn 22 for some constant p, where n is the length of the statement. So we say the complexity is at least doubly exponential. That’s much worse than exponential! On the other hand, an algorithm with a worst-case runtime no more than triply exponential was found by Oppen in 1978: [FR74] M. J. Fischer and Michael O. Rabin, Super-exponential complexity of Presburger arithmetic, Proceedings of the SIAM-AMS Symposium in Applied Mathematics 7 (1974), 27–41. 2pn [Opp78] Derek C. Oppen, A 22 upper bound on the complexity of Presburger arithmetic, J. Comput. Syst. Sci. 16 (1978), 323–332. How is all this connected to the reachability problem for Petri nets? First, provability is a lot like reachability, since in a proof you’re trying to reach the conclusion starting from the assumptions using certain rules. Second, like 251 25 25.1 COMPUTATION AND PETRI NETS The reachability problem Presburger arithmetic, Petri nets are all about addition, since they consists of transitions going between linear combinations like this: 6CO2 + 6H2 −→ C6 H12 O6 + 6O2 That’s why the old literature calls Petri nets vector addition systems. And third, the difficulty of deciding provability in Presburger arithmetic smells a bit like the difficulty of deciding reachability in Petri nets. So, it is interesting to know what happened after Peterson wrote his book. For starters, in 1981, the very year Peterson’s book came out, Ernst Mayr showed that the reachability problem for Petri nets is decidable: [May81] Ernst Mayr, Persistence of vector replacement systems is decidable, Acta Informatica 15 (1981), 309–318. As you can see from the title, Mayr actually proved some other property was decidable. However, it follows that reachability is decidable, and Mayr pointed this out in his paper. In fact the decidability of reachability for Petri nets is equivalent to lots of other interesting questions. You can see a bunch here: [EN94] Javier Esparza and Mogens Nielsen, Decidability issues for Petri nets— a survey, Bulletin of the European Association for Theoretical Computer Science 52 (1994), 245–262. Though Mayr found an algorithm to decide the reachability problem, this algorithm is complicated. Worse, it seems to take a hugely long time to run! The best known upper bound on its runtime is a function that’s not even ‘primitive recursive’. Here is the basic idea. We can define multiplication by iterating addition: n × m = n + n + n + ··· + n where add n to itself m times. Then we can define exponentiation by iterating multiplication: n ↑ m = n × n × n × ··· × n where we multiply n by itself m times. Here we’re using Knuth’s up-arrow notation. Then we can define tetration by iterating exponentiation: n ↑2 m = n ↑ (n ↑ (n ↑ · · · ↑ n))) Then we can define an operation ↑3 by iterating tetration, and so on. All these functions are said to be ‘primitive recursive’: very roughly, this means that each one can be computed from a previously defined one using a program with an extra loop. But the nth Ackermann number is not a primitive function of n; it’s defined to be n ↑n n The difference is the ↑n : now the number of nested loops in our program must itself grow with n. 252 25 25.2 COMPUTATION AND PETRI NETS Symmetric monoidal categories The Ackermann numbers grow at an insanely rapid rate: they eventually surpass any primitive recursive function! Here are the first three: 1↑1 = 1 2 ↑2 2 = 4 3 ↑3 3 = . 3. . 33 3 where we have a stack of 33 threes—or in other words, 37625597484987 threes! When we get to 4 ↑4 4, our minds boggle. We wish it didn’t, but it does. In 1998 someone claimed to have a faster algorithm for deciding the reachability problem: [Bou98] Zakaria Bouziane, A primitive recursive algorithm for the general Petri net reachability problem, in 39th Annual Symposium on Foundations of Computer Science, IEEE, 1998, pp. 130–136. Unfortunately, it seems that Bouzaine made a mistake: [Jan08] Petr Jančar, Bouziane’s transformation of the Petri net reachability problem and incorrectness of the related algorithm, Information and Computation, 206 (2008), 1259–1263. So at present, if we list a bunch of chemicals and reactions involving these chemicals, you can decide when some combination of these chemicals can turn into another combination. But it might take a very long time to decide this. What about the connection to Presburger arithmetic? That is explained here: [Ler08] Jérôme Leroux, The general vector addition system reachability problem by Presburger inductive separators, 2008. If we have a Petri net and one complex is reachable from another, we can eventually discover this by a brute-force search. So, the hard part of Mayr’s algorithm concerns the case when one complex is not reachable from another. In this case, Mayr’s algorithm will eventually prove it is impossible—thus deciding the reachability problem. What Leroux does is find another kind of proof, using Presburger arithmetic. Suppose we have two complexes κ, κ0 ∈ NS such that κ0 is not reachable from κ. Then Leroux’s algorithm constructs sets X, X 0 ⊆ NS , definable using Presburger arithmetic, such that κ ∈ X, κ ∈ X 0 , and it is provable using Presburger arithmetic that no complex in X 0 is reachable from one in X. In short, problems about what chemical reactions can do lead naturally to quite profound issues in computation and even the foundations of arithmetic! 25.2 Symmetric monoidal categories Next, let us explain yet another reason why the reachability problem is so fascinating. This little section is only for people who know about symmetric monoidal categories. 253 25 25.2 COMPUTATION AND PETRI NETS Symmetric monoidal categories As we mentioned long ago, a Petri net is actually nothing but a presentation of a symmetric monoidal category that is freely generated by some set S of objects and some set T of morphisms going between tensor products of those objects. The objects in S are our species. The tensor products of these objects are our complexes. The morphisms in T are our transitions or reactions. The details are explained here: [MM90] José Meseguer and Ugo Montanari, Petri nets are monoids, Information and Computation 88 (1990), 105–155. [BMMS01] Roberto Bruni, José Meseguer, Ugo Montanari and Vladimiro Sassone, Functorial models for Petri nets, Information and Computation 170 (2001), 207–236. In chemistry we write the tensor product as +, but category theorists usually write it as ⊗. Then the reachability problem consists of questions like this: Problem 51. Suppose we have a symmetric monoidal category freely generated by objects A, B, C, D, E, F, G, H and morphisms f: g: h: A⊗B C ⊗D E⊗F → C → E → C ⊗G⊗H Is there a morphism from A ⊗ B ⊗ D ⊗ F to C ⊗ G ⊗ H? Problems of this sort are reminiscent of the word problem for groups and other problems where we are given a presentation of an algebraic structure and have to decide if two elements are equal... but now, instead of asking whether two elements are equal we are asking if there is a morphism from one object to another. So, it is fascinating that this problem is decidable—unlike the word problem for groups—but still very hard to decide, in general. Just in case you want to see a more formal statement, let’s finish off by giving you that: Reachability Problem. Given a symmetric monoidal category C freely generated by a finite set of objects and a finite set of morphisms between tensor products of these objects, and given two objects x, y ∈ C, is there a morphism f : x → y? Theorem 68. (Lipton, Mayr) There is an algorithm that decides the reachability problem. However, for any such algorithm, for any c > 0, the worst-case run-time exceeds 2cn where n is the size of the problem: the sum of the number of generating objects, the number of factors in the sources and targets of all the generating morphisms, and the number of factors in the objects x, y ∈ C for which the reachability problem is posed. 254 25 25.3 25.3 COMPUTATION AND PETRI NETS Computation with Petri nets Computation with Petri nets We’ve been talking about using computers to figure out what a Petri net can do. Now let’s turn the question around and talk about using a Petri net as a computer. We could do this for ordinary Petri nets, but in keeping with the main theme of this book let us consider stochastic Petri nets. Remember, these are Petri nets equipped with a rate constant for each reaction. This allows us to formulate both the rate equation, describing how the expected number of things of each species changes with time, and the master equation, describing how the probability that we have a given number of things of each species changes with time. We could try to use either the rate equation or the master equation for computation, and people have studied both. The rate equation lets us build analog computers: [CDS14] Ho-Lin Chen, David Doty and David Soloveichik, Rate-independent computation in continuous chemical reaction networks, in ITCS 2014: Proceedings of the 5th Innovations in Theoretical Computer Science Conference, ACM, New York, pp. 313–326. In this particular formalism, the authors want the concentration of a particular species to converge in the limit as t → ∞ to a desired function of the ‘input’ concentrations, independent of the choice of rate constants. They show that this is possible precisely for continuous piecewise-linear functions. The master equation lets us build digital computers, as discussed here: [CSWB09] Matt Cook, David Soloveichik, Erik Winfree and Jehoshua Bruck, Programmability of chemical reaction networks, in Algorithmic Bioprocesses, eds. Condon, Harel, Kok, Salomaa and Winfree, Springer, Berlin, 2009, pp. 543–584. [Sol14] David Soloveichik, The computational power of chemical reaction networks, lecture at Programming with Chemical Reaction Networks: Mathematical Foundations, Banff, June 2014. [SCWB08] David Soloveichik, Matt Cook, Erik Winfree and Jehoshua Bruck, Computation with finite stochastic chemical reaction networks, Natural Computing 7 (2008), 615–633. [ZC08] Gianluigi Zavattaro and Luca Cardelli, Termination problems in chemical kinetics, in CONCUR 2008–Concurrency Theory, eds. Zavattaro and Cardelli, Lecture Notes in Computer Science 5201, Springer, Berlin, 2008, pp. 477–491. The input to such a computer will always be a vector of natural numbers, encoded as the numbers of items of various species. There are some choices as to how the computation proceeds: 255 25 25.3 COMPUTATION AND PETRI NETS Computation with Petri nets • decision problems vs more general computational problems: in a decision problem the output is either ‘yes’, indicated by the presence of a species we’ll call Y , or ‘no’, indicated by the presence of a species called N . In a more general computational problem the answer could be a natural number, or a vector of natural numbers. • uniform vs non-uniform: in the uniform case a single stochastic Petri net is supposed to handle all inputs. In the non-uniform case we can add extra transitions for larger inputs. • deterministic vs probabilistic: in the deterministic case we are guaranteed to get the correct output. In the probabilistic case, it is merely likely. • halting vs stabilizing: does the Petri net ‘tell us’ when it has finished, or not? In the halting case it irreversibly produces some items of a given species that signal that the computation is done. In the stabilizing case it eventually stabilizes to the right answer, but we may not know how long to wait. These choices dramatically affect the computational power of our computer. For example, consider decision problems and uniform computation. Then there are four cases: • deterministic and halting: this has finite computational power. • deterministic and stabilizing: this can decide any ‘semilinear predicate’, as explained in the next section. • probabilistic and halting: this can decide any ‘recursive predicate’. In other words, this class of reactions can solve any decision problem that any Turing machine can solve. • probabilistic and stabilizing: this can decide so-called ‘∆02 predicates’, which are more general than recursive ones. This may sound odd, but if we use Turing machines but don’t require them to signal when they’ve halted, the resulting infinitely long computations can decide predicates that are not recursive. Instead of discussing all four cases in detail, let’s look at one: deterministic stabilizing computations. These give a nice variant on the earlier theme of ‘reachability’. In the reachability problem we asked if starting from some initial complex we can reach some final complex. In deterministic stabilizing computations we ask if starting from some initial complex we must eventually reach a final complex in some set, and then stay in that set. This set encodes a YES answer to our decision problem. In deterministic stabilizing computations, the rate constants of the transitions don’t matter. The reason is that we’re demanding that an outcome occur with certainty, and we don’t care how long it takes. So, even though we’ve been talking about stochastic Petri nets, the concept of deterministic stabilizing computation makes sense for mere Petri nets. 256 25 25.3 COMPUTATION AND PETRI NETS Computation with Petri nets Deterministic stabilizing computations Let us be a bit more precise. Suppose we have a subset X ⊆ Nd , and we want to answer this question: is the vector κ ∈ Nd in the set X? This is a decision problem. What does it mean to solve this problem with a Petri net using a deterministic stabilizing computation? To do this, we represent our vector κ as a bunch of molecules: κ1 of the first kind, κ2 of the second kind, and so on. We may also include a fixed collection of additional molecules to help the reactions run. Mathematically, κ is a complex called the input, and the extra molecules are some fixed complex λ ∈ Nd . Then we choose a Petri net and let it run, starting with the complex κ + λ. The answer to our question will be encoded in some molecules called Y and N . If κ is in X, we want our Petri net to eventually produce at least one Y molecule. If it’s not, we want it to produce at least one N . To make this more precise, we need to define what counts as an output. If we’ve got a complex that contains Y but not N , then the output is YES. If it contains N but not Y , then the output is NO. If it contains both, then the output is undefined. Output-stable states are complexes with YES or NO output such that all complexes reachable from them via transitions in our Petri net give the same output. We say an output-stable-state is correct if this output is the correct answer to the question: is κ in X? Our Petri net gives a deterministic stabilizing computation if for any input, and choosing any state reachable from that input, we can do further chemical reactions to reach a correct output-stable state. In other words: starting from our input, and using any transitions we want, we will eventually reach an output-stable state that gives the right answer to the question “is κ in X?” This sounds a bit complicated, but it’s really not. Let’s look at two examples. Checking an inequality Suppose you want to check two numbers and see if one is greater than or equal to another. Here X = {(κ1 , κ2 ) : κ1 ≤ κ2 } How can you decide if a pair of numbers (κ1 , κ2 ) is in this set? You start with κ1 molecules of type A, κ2 molecules of type B, and one molecule of type Y . Then you use a Petri net with these transitions: A+N →Y B+Y →N If you let these run, the Y switches to a N each time the transition destroys an A. But the N switches back to a Y each time the transition destroys a B. When no more transitions are possible, you are left with either one Y or one N , which is the correct answer to your question! 257 25 25.3 COMPUTATION AND PETRI NETS Computation with Petri nets Checking equality Suppose you want to check two numbers and see if they are equal. Here S = {(κ1 , κ2 ) : κ1 = κ2 } How can you decide if a pair of numbers (κ1 , κ2 ) is in here? This is a bit harder! As before, you start with κ1 molecules of type A, κ2 molecules of type B, and one molecule of type Y. Then you use a stochastic Petri net with these reactions: A+B →Y Y +N →Y A+Y →A+N B+Y →B+N The first transition lets an A and a B cancel out, producing a Y . If you only run this transition, you’ll eventually have some As left and no Bs, or have some Bs left and no As, or you’ll have nothing left but Y s. If you have nothing but Y s, no more transitions are possible and your numbers were equal. The other reactions deal with the cases where you have some As or Bs left over. You can check that no matter what order we run the reactions, we’ll eventually get the right answer! In the end, you’ll have either Y s or N s, not both, and this will provide the correct yes-or-no answer to the question of whether κ1 = κ2 . The power of deterministic stabilizing computations Now you’ve seen some examples of deterministic stabilizing computations. The big question is: what kind of questions can they answer? More precisely, for what subsets X ⊆ Nd can we build a deterministic stabilizing computation that ends with output YES if the input κ lies in X and with output NO otherwise? The answer is: the ‘semilinear’ subsets! [AAE06] Dana Angluin, James Aspnes and David Eistenstat, Stably computable predicates are semilinear, in Twenty-Fifth ACM Symposium on Principles of Distributed Computing, July 2006, ACM Press, New York, 2006, pp. 292–299. A subset S ⊆ Nd is linear if it’s of the form {u0 + n1 u1 + · · · + np up : ni ∈ N} for some fixed vectors ui ∈ Nd . A subset S ⊆ Nd semilinear if it’s a finite union of linear sets. How did Angluin, Aspnes and Eisenstat prove that deterministic stabilizing computations can decide membership in precisely the semilinear subsets? The easy part is showing that membership in any semilinear subset can be decided by some Petri net. The hard part is the converse. This uses a nice fact which is worth knowing for its own sake: 258 25 25.4 COMPUTATION AND PETRI NETS Software Theorem 69 (Dickson’s Lemma). Any subset of Nd has a finite set of minimal elements, where we define x ≤ y if xi ≤ yi for all i. For example, the region above and to the right of the hyperbola here has five minimal elements: It’s very neat to see how Dickson’s lemma helps! You can see how it works in Angluin, Aspnes and Eistenstat’s paper—but they call it ‘Higman’s lemma’. 25.4 Software Finally, no overview of the connection between Petri nets and computation would be complete without a word about free software for working with Petri nets, chemical reaction networks and the like. Jim Stuttard found quite a lot of such software for us, but we will only mention three tools. We shall not explain all the terminology in what follows: unfortunately, there is a lot about Petri nets that we have not discussed! Cain Cain performs stochastic and deterministic simulations of chemical reactions. It can spawn multiple simulation processes to utilize multi-core computers. It stores models, simulation parameters, and simulation results in XML, and it lets you import and export models in Systems Biology Markup Language. It comes equipped with examples from this nice book: [Wil06] Darren James Wilkinson, Stochastic Modelling for Systems Biology, Taylor and Francis, New York, 2006. CPN Tools CPN Tools is probably the most popular tool for editing, simulating, and analyzing colored Petri nets—a generalization of Petri nets where the tokens can have different ‘colors’. Colored Petri nets are explained here: [AS11] Wil van der Aalst and Christian Stahl, Modeling Business Processes: A Petri Net-Oriented Approach, MIT Press, Cambridge, 2011. 259 25 25.5 COMPUTATION AND PETRI NETS Answers [JK09] Kurt Jensen and Lars M. Kristensen, Coloured Petri Nets: Modelling and Validation of Concurrent Systems, Springer, Berlin, 2009. Ordinary Petri nets are a special case. CPN Tools has been developed over the past 25 years. It features incremental syntax checking and code generation, which take place while a net is being constructed. A fast simulator efficiently handles both untimed and timed nets. Full and partial state spaces can be generated and analyzed, and a standard state space report contains information, such as boundedness properties and liveness properties. GreatSPN GreatSPN stands for ‘GRaphical Editor and Analyzer for Timed and Stochastic Petri Nets’. It is a software package for the modeling, validation, and performance evaluation of distributed systems using generalized stochastic Petri nets. The tool provides a framework to experiment with timed Petri net based modeling techniques. It is composed of many separate programs that cooperate in the construction and analysis of Petri net models by sharing files. Using network file system capabilities, different analysis modules can be run on different machines in a distributed computing environment. The creators of this software have also written a book explaining generalized stochastic Petri nets: [MBCDF95] M. Ajmone Marsan, G. Balbo, G. Conte, S. Donatelli and G. Franceschinis, Modelling with Generalized Stochastic Petri Nets, Wiley, New York, 1995. If you enter some personal data, you can download a gzipped tar file containing an electronic version of this book. 25.5 Answers Problem 51. Suppose we have a symmetric monoidal category freely generated by objects A, B, C, D, E, F, G, H and morphisms f: g: h: A⊗B C ⊗D E⊗F → C → E → C ⊗G⊗H Is there a morphism from A ⊗ B ⊗ D ⊗ F to C ⊗ G ⊗ H? Answer. Yes. If you draw the Petri net corresponding to this symmetric monoidal category, you will see it looks exactly like the one we saw in Section 25.1: 260 25 25.5 COMPUTATION AND PETRI NETS Answers after we rename the objects: A = O2 , B = C, C = CO2 , D = NaOH, E = NaHCO2 , F = HCl, G = H2 O, H = NaCl After this renaming the reachability problem here is revealed to be the one we already solved n Section 25.1, with the answer being ‘yes’. What seemed like category theory is secretly chemistry—or perhaps the other way around! 261 26 26 SUMMARY TABLE Summary table state quantum mechanics stochastic mechanics vector ψ ∈ Rn with X |ψi |2 = 1 vector ψ ∈ Cn with X ψi = 1 i i and ψi ≥ 0 observable vector O ∈ Rn n × n matrix O with O† = O where (O† )ij := Oji expected value hψ, Oψi := X hOψi := ψ i Oij ψj i,j symmetry (linear map sending states to states) X Oi ψi i unitary n × n matrix: U U † = U †U = 1 stochastic n × n matrix: X Uij = 1, Uij ≥ 0 i symmetry generator self-adjoint n × n matrix: H=H † infinitesimal stochastic n × n matrix: X Hij = 0, i 6= j ⇒ Hij ≥ 0 i symmetries from symmetry generators equation of motion U (t) = exp(−itH) d ψ(t) = Hψ(t) dt with solution i ψ(t) = exp(−itH)ψ(0) 262 U (t) = exp(tH) d ψ(t) = Hψ(t) dt with solution ψ(t) = exp(tH)ψ(0) REFERENCES REFERENCES In what follows, the pages that refer to a given citation are indicated by numbers in blue. Free online references are in red. These colored items are clickable links. References [AS11] Wil van der Aalst and Christian Stahl, Modeling Business Processes: A Petri Net-Oriented Approach, MIT Press, Cambridge, 2011. (259) [AAE06] Dana Angluin, James Aspnes and David Eistenstat, Stably computable predicates are semilinear, in Twenty-Fifth ACM Symposium on Principles of Distributed Computing, July 2006, ACM Press, New York, 2006, pp. 292–299. Also available at http://www.cs.yale.edu/homes/aspnes/papers/podc2006proceedings.pdf. (258) [ACK] David F. Anderson, George Craciun, and Thomas G. Kurtz, Product-form stationary distributions for deficiency zero chemical reaction networks, Bull. Math. Bio. (2010), 1947–1970. Also available as arXiv:0803.3042. (68, 82) [And] David F. Anderson, A proof of the Global Attractor Conjecture in the single linkage class case. Available as arXiv:1101.0761. (186) [BC07] M. A. Buice and J. D. Cowan, Field-theoretic approach to fluctuation effects in neural networks, Phys. Rev. E 75 (2007), 051919. (58) [BC09] M. A. Buice and J. D. Cowan, Statistical mechanics of the neocortex, Prog. Biophysics and Mol. Bio. 99 (2009), 53–86. (58) [BF] John C. Baez and Brendan Fong, A Noether theorem for Markov processes, J. Math. Phys. 54 (2013), 013301. Also available as arXiv:1203.2035. (94, 239) [BFB66] A. T. Balaban, D. Fǎrcaşiu, and R. Bǎnicǎ, Graphs of multiple 1,2shifts in carbonium ions and related systems, Rev. Roum. Chim. (1966), 11. (120) [BR91] Danail Bonchev and D. H. Rouvray, editors. Chemical Graph Theory: Introduction and Fundamentals, Taylor and Francis, 1991. (141) [Bou98] Zakaria Bouziane, A primitive recursive algorithm for the general Petri net reachability problem, in 39th Annual Symposium on Foundations of Computer Science, IEEE, 1998, pp. 130–136. Also available at http://hal.inria.fr/inria-00073286/PDF/RR-3404.pdf. (253) [BMMS01] Roberto Bruni, José Meseguer, Ugo Montanari and Vladimiro Sassone, Functorial models for Petri nets, Information and Computation 170 (2001), 207–236. Also available at http://eprints.soton.ac.uk/264742/1/prenetsIandCOff.pdf. (254) 263 REFERENCES REFERENCES [Car] J. Cardy, Renormalisation group approach to reaction-diffusion problems. Available as arXiv:cond-mat/9607163. (58) [CDS14] Ho-Lin Chen, David Doty and David Soloveichik, Rate-independent computation in continuous chemical reaction networks, in ITCS 2014: Proceedings of the 5th Innovations in Theoretical Computer Science Conference, ACM, New York, pp. 313–326. Also available at http://www.dna.caltech.edu/∼ddoty/papers/riccrn.pdf. (255) [CSWB09] Matt Cook, David Soloveichik, Erik Winfree and Jehoshua Bruck, Programmability of chemical reaction networks, in Algorithmic Bioprocesses, eds. Condon, Harel, Kok, Salomaa and Winfree, Springer, Berlin, 2009, pp. 543–584. Also available at http://www.dna.caltech.edu/Papers/programmability of CRNs preprint2008.pdf. (255) [Cra15] Gheorghe Craciun, Toric differential inclusions and a proof of the Global Attractor Conjecture. Available as arXiv:1501.02860. (186) [CDSS] Gheorghe Craciun, Alicia Dickenstein, Anne Shiu, and Bernd Sturmfels, Toric dynamical systems. Available as arXiv:0708.3431. (186) [CTF06] Gheorghe Craciun, Yangzhong Tang, and Martin Feinberg, Understanding bistability in complex enzyme-driven reaction networks, Proc. Nat. Acad. Sci. USA 103 (2006), 8697–8702. Also available at http://mbi.osu.edu/files/3813/3528/0136/techreport 50 1.pdf. (177) [DF09] Peter J. Dodd and Neil M. Ferguson, A many-body field theory approach to stochastic models in population biology, PLoS ONE (2009), 4. Available at http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0006855. (58) [Doi76a] M. Doi, Second-quantization representation for classical manyparticle systems, J. Phys. A 9 (1976), 1465–1477. (58) [Doi76b] M. Doi, Stochastic theory of diffusion-controlled reactions, J. Phys. A 9 (1976), 1479–1495. (58) [DSle] P. G. Doyle and J. L. Snell. Random Walks and Electrical Circuits, Mathematical Association of America, Washington DC, 1984. Also available at http://www.math.dartmouth.edu/∼doyle. (152) [EN94] Javier Esparza and Mogens Nielsen, Decidability issues for Petri nets—a survey, Bulletin of the European Association for Theoretical Computer Science 52 (1994), 245–262. Also available at http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.2.3965. (252) 264 REFERENCES REFERENCES [Fei79] Martin Feinberg, Lectures On Reaction Networks, 1979. Available at http://www.che.eng.ohiostate.edu/∼FEINBERG/LecturesOnReactionNetworks/. (177, 220) [Fei87] Martin Feinberg, Chemical reaction network structure and the stability of complex isothermal reactors: I. The deficiency zero and deficiency one theorems, Chemical Engineering Science 42 (1987), 2229–2268. (177, 220) [FR74] M. J. Fischer and Michael O. Rabin, Super-exponential complexity of Presburger arithmetic, Proceedings of the SIAM-AMS Symposium in Applied Mathematics 7 (1974), 27–41. (251) [For11] Daniel B. Forger, Signal processing in cellular clocks, Proc. Nat. Acad. Sci. USA 108 (2011), 4281–4285. Available at http://www.pnas.org/content/108/11/4281.full.pdf. (177) [Fro12] Georg Frobenius, Über Matrizen aus nicht negativen Elemente, S.-B. Preuss Acad. Wiss. Berlin (1912), 456–477. (162) [Fuk80] M. Fukushima, Dirichlet Forms and Markov Processes, NorthHolland, Amsterdam, 1980. (152) [GP98] Peter J. E. Goss and Jean Peccoud, Quantitative modeling of stochastic systems in molecular biology by using stochastic Petri nets, Proc. Natl. Acad. Sci. USA 95 (1998), 6750–6755. (16) [Gub03] Jonathan M. Guberman, Mass Action Reaction Networks and the Deficiency Zero Theorem, B.A. Thesis, Department of Mathematics, Harvard University, 2003. Available at http://www4.utsouthwestern.edu/altschulerwulab/theses/ GubermanJ Thesis.pdf. (219) [Gun03] Jeremy Gunawardena, Chemical reaction network theory for in-silico biologists, 2003. Available at http://vcp.med.harvard.edu/papers/crnt.pdf. (177, 219) [Haa02] Peter J. Haas, Stochastic Petri Nets: Modelling, Stability, Simulation, Springer, Berlin, 2002. (16) [Hor74] Fritz Horn, The dynamics of open reaction systems, in Mathematical Aspects of Chemical and Biochemical Problems and Quantum Chemistry, ed. Donald S. Cohen, SIAM–AMS Proceedings 8, American Mathematical Society, Providence, R.I., 1974, pp. 125–137. (186) [HJ72] F. Horn and Roy Jackson, General mass action kinetics, Archive for Rational Mechanics and Analysis 47 (1972), 81–116. (177, 185, 220) 265 REFERENCES REFERENCES [Jan08] Petr Jančar, Bouziane’s transformation of the Petri net reachability problem and incorrectness of the related algorithm, Information and Computation, 206 (2008), 1259–1263. (253) [JK09] Kurt Jensen and Lars M. Kristensen, Coloured Petri Nets: Modelling and Validation of Concurrent Systems, Springer, Berlin, 2009. (260) [KM27] W. O. Kermack and A. G. McKendrick, A contribution to the mathematical theory of epidemics, Proc. Roy. Soc. Lond. A 772 (1927), 700–721. Available at http://rspa.royalsocietypublishing.org/content/115/772/700.full.pdf. (25) [Kin93] R. Bruce King, Applications of Graph Theory and Topology in Inorganic Cluster Coordination Chemistry, CRC Press, 1993. (141) [Koc10] Ina Koch, Petri nets—a mathematical formalism to analyze chemical reaction networks, Molecular Informatics 29 (2010), 838–843. (16) [Ler08] Jérôme Leroux, The general vector addition system reachability problem by Presburger inductive separators, 2008. Available at http://hal.inria.fr/docs/00/31/93/93/PDF/main.pdf. (253) [Lip76] Roger J. Lipton, The reachability problem requires exponential space, Technical Report 62, Yale University, 1976. Available at http://www.cs.yale.edu/publications/techreports/tr63.pdf. (250) [Log96] S. R. Logan, Chemical Reaction Kinetics, Longman, Essex, 1996. (22) [LR68] Paul C. Lauterbur and Fausto Ramirez, Pseudorotation in trigonalbipyramidal molecules, J. Am. Chem. Soc. 90 (1968), 6722–6726. (124) [Man06] Marc Mangel, The Theoretical Biologist’s Toolbox: Quantitative Methods for Ecology and Evolutionary Biology, Cambridge U. Press, Cambridge, 2006. (27) [MG98] D. C. Mattis and M. L. Glasser, The uses of quantum field theory in diffusion-limited reactions, Rev. Mod. Phys. 70 (1998), 979–1001. (58) [MR92] Zhi-Ming Ma and Michael Röckner, Introduction to the Theory of (Non-Symmetric) Dirichlet Forms, Springer, Berlin, 1992. (112) [MBCDF95] M. Ajmone Marsan, G. Balbo, G. Conte, S. Donatelli and G. Franceschinis, Modelling with Generalized Stochastic Petri Nets, Wiley, New York, 1995. Also available at http://www.di.unito.it/∼greatspn/bookdownloadform.html. (260) 266 REFERENCES REFERENCES [May81] Ernst Mayr, Persistence of vector replacement systems is decidable, Acta Informatica 15 (1981), 309–318. (252) [MM90] José Meseguer and Ugo Montanari, Petri nets are monoids, Information and Computation 88 (1990), 105–155. Available at http://www.sciencedirect.com/science/article/pii/0890540190900138. (254) [New] Michael William Newman, The Laplacian Spectrum of Graphs, Masters Thesis, Department of Mathematics, University of Manitoba, 2000. Available at http://www.seas.upenn.edu/∼jadbabai/ESE680/Laplacian Thesis.pdf. (140) [Opp78] Derek C. Oppen, A 22 upper bound on the complexity of Presburger arithmetic, J. Comput. Syst. Sci. 16 (1978), 323–332. (251) [Pel85] L. Peliti, Path integral approach to birth-death processes on a lattice, J. Phys. France 46 (1985), 1469–1483. (58) [Per07] Oskar Perron, Zur Theorie der Matrizen, Math. Ann. 64 (1907), 248–263. (161) [Pet81] James L. Peterson, Petri Net Theory and the Modeling of Systems, Prentice–Hall, New Jersey, 1981. (251) [Ran97] Milan Randic, Symmetry properties of graphs of interest in chemistry: II: Desargues–Levi graph, Int. Jour. Quantum Chem. 15 (1997), 663–682. (121) [RRTK] Katarzyna Pichòr Ryszard Rudnicki and Marta Tyran-Kamı̀nska, Markov semigroups and their applications, 2002. Available at http://www.impan.pl/∼rams/r48-ladek.pdf. (112) [ST77] S. Sacerdote and R. Tenney, The decidability of the reachability problem for vector addition systems, Conference Record of the Ninth Annual ACM Symposium on Theory of Computing, 2-4 May 1977, Boulder, Colorado, USA, ACM, 1977, pp. 61–76. (250) [Sol14] David Soloveichik, The computational power of chemical reaction networks, lecture at Programming with Chemical Reaction Networks: Mathematical Foundations, Banff, June 2014. Available at http://crn.thachuk.com/images/d/d5/SoloveichikBanff Tutorial.pdf. (5, 240, 255) 2pn [SCWB08] David Soloveichik, Matt Cook, Erik Winfree and Jehoshua Bruck, Computation with finite stochastic chemical reaction networks, Natural Computing 7 (2008), 615–633. Also available at http://dna.caltech.edu/Papers/sCRN computation TR2007.pdf. (255) 267 REFERENCES REFERENCES [THVL78] Uwe Täuber, Martin Howard, and Benjamin P. Vollmayr-Lee, Applications of field-theoretic renormalization group methods to reactiondiffusion problems, J. Phys. A 38 (2005), R79. Also available as arXiv:cond-mat/0501678. (58) [Tri92] Nenad Trinajstic, Chemical Graph Theory, CRC Press, Boca Raton, Florida, 1992. (141) [Wil06] Darren James Wilkinson, Stochastic Modelling for Systems Biology, Taylor and Francis, New York, 2006. (16, 259) [Van07] Nico Van Kampen, Stochastic Processes in Physics and Chemistry, North–Holland, New York, 2007. (58) [ZC08] Gianluigi Zavattaro and Luca Cardelli, Termination problems in chemical kinetics, in CONCUR 2008–Concurrency Theory, eds. Zavattaro and Cardelli, Lecture Notes in Computer Science 5201, Springer, Berlin, 2008, pp. 477–491. Also available at http://lucacardelli.name/Papers/Termination Problems in Chemical Kinetics.pdf (255) 268 Index L1 , 42, 43, 138 L2 , 42, 43 nonclassical ion, 121 solid, 122 water, 12 Ackermann numbers, 252 classical field theory, 13 Anderson–Craciun–Kurtz theorem, 68, classical random walk, see random walk 82–87 coherent state, 69 example of, 89–92, 187–199 colored Petri net, 259 statement of, 86 compartmental model, 27 annihilation operator, 49, 76 complex, 83, 172, 201, 221, 234, 249 Azimuth complex balanced equilibrium, 83, 90– blog, 241 91, 236 Azimuth Project, 5 computer science, 248–259 blog, 4, 86, 131 completeness, 251 network theory series, 4 decidability, 248–252 decision problem, 256–259 balanced graph, 245, 246 deterministic stabilizing computaBergholm, Ville, 240, 241 tion, 257–259 bistability, 177 primitive recursive function, 252– boundary operator, 229 253 Brownian motion, 135, 191 reachability problem, 248–254 bubonic plague, 26 conductance, see electrical circuits connected component, 174, 226–228 category theory, see also diagrammatic and conserved quantities, 239 language, 253–254 and probability distributions, 227 adjoint functors, 219 constant, 239 category as directed multigraph, deficiency, 174 210 definition of, 183 opposite category, 23 equilibrium state, 142 symmetric monoidal category, 23 functions on, 233 tensor product, 254 single, 227 Cauchy sequence, 111 conserved quantity, see Noether’s thechemical computer, 248 orem, 181 chemical reaction network, see reaction contraction semigroup, 111 network Corfield, David, 59, 71 chemistry creation operator, 40, 49, 77 bromine, 178 Desargues graph, 118 decidability, 248–252 diatomic molecules, 178–189 decision problem, 256–259 ethane, 118 deficiency, 184, 205 ethyl cation, 118, 121–134 Deficiency Zero Theorem, 183–186, 200– hydrogen chloride, 122 239 iodine, 178 baby version, 202, 212 ligand, 122 child’s version, 236 269 INDEX INDEX example of, 178 history, 68–69 introduction to, 170–177, 186 proof of, 231–239 statement of, 176, 185 degree, 139 Desargues configuration, 128 Desargues graph, 120–134 Laplacian of, 141 deterministic stabilizing computation, 257–259 definition of, 257 diagrammatic language, 10 Dickson’s lemma, 259 diffusion, 58, 191 directed multigraph, 172, 226 Dirichlet operator, 145–153, 159, 168 and Noether’s theorem, 244–247 definition of, 146 electrical circuits, 145 irreducible, 162, 168 doubly stochastic operator, 245, 246 doubly stochastic semigroup, 246 edge, 137, 139, 172, 210 Egan, Greg, 133, 141, 197 Ehrenfest theorem, 246–247 electrical circuits and Dirichlet operators, 145–152 power, 148 resistor networks, 152 equilibrium, 236 complex balanced, 83 complex balanced, 83, 90–91, 236 expected value, 12, 105 and Noether’s theorem, 97–101 definition of, quantum, 105 definition of, stochastic, 95 Markov chains, 98 of Hamiltonian, quantum, 148 positive, 159 quantum vs stochastic, 158 standard deviation, 12 time derivative of, 60 time independent, 95 falling power, 33 Ferraris, Federica, 5 Feynman diagrams, 73–81 and logistic growth, 66–68 and probabilities, 35 and stochastic Petri nets, 34 with animals, 13 Fong, Brendan, 3, 4, 82, 86 formal variable, 39, 65 generalized stochastic Petri net, 260 generating function, see power series Global Attractor Conjecture, 185–186 graph strongly connected, 168 graph Laplacian, see Laplacian graph theory balanced graph, 245, 246 connected component, 174, 226– 228 Desargues graph, 120–134 random walk on, 137 Desargues’ theorem, 127 directed graph, 210 directed multigraph, 172, 210, 211 category as, 210 edge, 137, 139, 172, 210 Levi graph, 128 quiver, 210 simple graph, 139, 147 strongly connected component, 226– 227 terminal, 228 strongly connected graph, 160 vertex, 137, 139, 172, 210 degree of, 139 indegree and outdegree of, 245 weakly reversible graph, 225 weighted graph, 223, 245 group theory unitary 1-parameter, 107, 108, 111 word problem, 254 Hamiltonian, 76 270 INDEX INDEX of weighted graph, 224, 229, 245, 246 heat equation, 135, 139, 141 Hilbert space, see also vector space complex, 36 finite-dimensional, 108 Fock space, 36 from vector space, 107 one-parameter unitary groups on, 108 Hille–Yosida theorem, 110–112 linkage class, see connected component Lotka–Volterra equations, 18 model, 11, 16 labelling, 30 Laplacian graph, 135–144 definition of, 139 Desargues graph, 141 Levi graph, 128 linear operator, see operator linear subset, 258 Noether’s theorem, 92–102 for Dirichlet operators, 244–247 for Markov chains, 99 quantum version, 115, 116, 242 quantum vs stochastic, 115–117, 242–243 stochastic version, 97, 243 nonlinearity Markov chain, 98–100 Markov process, see also random walk, stochastic mechanics, master equation, 95 equilibrium theorem, 227 Markov semigroup, 95, 109–111 mass-action kinetics, see chemical reindegree, 245 action network index free vector notation, 27 master equation, 32, 33, 44 infinitesimal doubly stochastic operaand Markov processes, 94 tor, 246 equilibrium, 86, 89 infinitesimal stochastic of stochastic Petri net, 73 Laplacian, 135, 140, 145 vs Hamiltonian, 51 sign convention, 150 matrix infinitesimal stochastic operator, 96– exponentiation, 4, 212, 213, 234 100, 109–116, 146, 156, 166, irreducible, 161, 165, 167, 168 168, 224 isometry, 157 definition of, 46 magic of matrix multiplication, 95, input, 73, 248 96 input complex, 83 matrix algebra, 224 integration, 42 nonnegative, 160–162 and measure space, 104 off-diagonal entries, 38, 46, 47 finite sums, 104 permutation, 113, 166, 167 partial fractions, 64 positive, 159 vs sums, 42 positive definite, 159 intensity matrix, see infinitesimal stochasstochastic, 167 tic operator, 46 unitary, 167 irreducible Dirichlet operator, 168 measure space, 42, 59, 96, 104, 109, irreducible matrix, 165, 167, 168 111, 113, 115 isometry, 43, 106 finite, 106 probability distributions on, 138 Jain, Arjun, 166, 197 multigraph, see directed multigraph 271 INDEX INDEX equation, 213 matrix exponentiation, 4, 212–219, 234 quantum probabilities, 2 rate equation, 4, 200, 232 relation to linear operator, 237 number operator, 52, 192 labelling, 30 master equation, 11 rate constant, 10, 73 reachability, 248–254 species, 10, 73 stochastic, 11 definition of, 73 generalized, 260 observable token, 30, 259 quantum, 105 transition, 10 stochastic, 105 vector addition system, 252 one-parameter unitary group, 44, 108, Poisson distribution, 41 135, 141, 145 and coherent states, 69 continuous, 107 and the deficiency zero theorem, open quantum system, see quantum me68 chanics independent pair, 91 operator standard deviation, 70, 71 Dirichlet, 168 polarization identity, 117 doubly stochastic, 245 power series, 188 infinitesimal doubly stochastic, 246 and the Hamiltonian, 65 infinitesimal stochastic, 96–100, 109– basis of, 57 116, 146, 156, 166, 168, 224 convergence of, 49, 79 definition of, 46 derivative of, 52 isometry, 106 formal, 49 self-adjoint, 105, 157 master equation, 84 stochastic, 43, 45, 106, 157 probability distribution, 39, 76 definition of, 43 properties of, 59–61 unitary, 106 quantum harmonic oscillator, 69 opposite category, 23 Presburger arithmetic, 253 outdegree, 245 primitive recursive function, 252–253 output, 73, 248 projective geometry, 127 output complex, 83 output-stable state, 257 quantum field theory, see also Feynman diagram partial fractions, 64 annihilation operator, 49, 76 permutation matrix, 113, 166, 167 coherent state, 69 Perron’s theorem, 159–160 creation operator, 40, 49, 77 statement of, 160 Fock space, 36 Perron–Frobenius theorem number operator, 52, 192 and quantum states, 153 quantum mechanics and the deficiency zero theorem, decoherence, 121 200, 227–232 Ehrenfest theorem, 246 statement of, 161 Hamiltonian, 108 Petri net, 10–11 harmonic oscillator, 69 colored, 259 isometry, 44, 106, 107, 112, 157 definition of, 11, 73, 248 definition of, 157 272 INDEX INDEX Schrödinger’s equation, 36, 108, 116, renormalization group, 58 141, 156, 212 resistance, see electrical circuits definition of, 36 reversible reaction network, 183 self-adjoint operator, 105 Schrödinger, Erwin, 69 unitary operator, 106 self-adjoint operator, 105, 157 quantum walk, see random walk definition of, 146 quiver, 210 semigroup random walk 1-parameter, 107, 109 and electrical networks, 152 contraction, 111 as Markov processes, 135, 140, 191, doubly stochastic, 246 222 Markov, 95, 109–111 classical particle, 145 semilinear subset, 258 definition of, 212 SI model, 24 Desargues graph, 121, 135–141 simple graph, 139, 147 heat equation, 141 SIR model, 25–26 on line segment, 191 SIRS model, 26–27 on reaction network, 215 Soloveichik, David, 5, 255 quantum walk, 135, 145 source, 172, 201, 223, 234 rank-nullity theorem, 206 species, 10, 73, 170, 201, 221, 234, 248 rate constant, 73, 223, 234 standard deviation rate equation and variance, 71 definition of, 12 Poisson distribution, 70, 71 stochastic Petri net, 13 state, 223 equilibrium, 89 stochastic, 76 in terms of matrix exponentiation, stochastic matrix, 167 217 stochastic mechanics reachability, 248–254 master equation, 156 definition of, 249 stochastic operator, 43, 45, 106, 157 problem, 248–254 definition of, 43 definition of, 250 doubly stochastic, 245, 246 reaction network, 170–177, 201 stoichiometric chemical species, 89 compatibility class, 175, 177, 184 deficiency of, 174, 184 positivity, 184 definition of, 172 subspace, 184 random walk on complexes, 215 Stone’s theorem reversible, 183 baby version, 108 species, 170 for unitary groups, 108 stochastic generalization to contraction semidefinition of, 172 groups, 111 stoichiometric compatibility class, strongly connected component, 226–227 184 terminal, 228 stoichiometric subspace of, 184 strongly connected graph, 160, 168 weakly reversible, 176 and irreducibility, 161, 162 reaction network deficiency Stuttard, Jim, 5, 240, 259 definition of, 174 symmetric monoidal category, 23 273 INDEX INDEX target, 172, 201, 223, 234 tensor product, 254 token, 30, 259 transition, 10, 73, 201, 221, 223, 234, 248 transition rate matrix, see infinitesimal stochastic operator unbounded operator, 108 unitary matrix, 167 unitary operator, 106 vector, see also vector space nonnegative, 160 positive definition of, 160 span of, 207 vector addition system, see also Petri net, 252 vector space, 107, 174 basis, 155 of probability distributions, 138 rank-nullity theorem, 206 spanned by, 207 vertex, 137, 139, 172, 210 degree of, 139 weakly reversible graph, 225 reaction network, 176 weighted graph, 223, 245 274