Riemannian game dynamics are dynamics for one-population games induced by the payoff vector field. Under some circumstances, such dynamics display Hamiltonian properties.
#hamiltonian-game-dynamics #evolutionary-game-theory #replicator-dynamics
Evolutionary game dynamics is a well-established interdisciplinary subject with far-reaching applications to genetics, ecology, chemistry, physics of complex systems, sociobiology and economics12.
Broadly speaking it models the evolution of a population of individuals divided into sub-populations mutually interacting with each other, the interaction being effectively captured by so-called fitness functions.
The fraction of the individuals belonging to a subpopulation $i$ is given by a number $x_i$ between $0$ and $1$, so that the state $x$ of the entire population at any given time is described by a point on the standard simplex $\Delta$. The fitness of a subpopulation when the population is at some state is then a real number $f_i(x)$ depending on the subpopulation and the population state, and the fraction of individuals belonging to a given subpopulation expands or shrink over time according to a dynamical system on the simplex
\[\dot{x} = X(x) \tag{1}\]the vector field $X$ being generated in some way by the fitness functions $f$.
Mertikopoulos and Sandholm3 showed that one way of generating the dynamical system $(1)$ in the interior of the simplex from the fitness functions $f$ is by means of a Riemannian metric $g$, demanding the motion to minimize the difference between a cost function and a gain function: \(X(x) = \text{argmin}_{v \in T_x{\Delta}} \left( \frac{1}{2} ||v||_x^2 - v \cdot f(x) \right) \tag{RGD-1}\) This is equivalent to orthogonally projecting (with respect to $g$) the vector field $f^{\sharp}$ on the simplex, where $\sharp$ is the sharp isomorphism associated with the metric $g$ and $f$ is seen as a $1$-form. The vector field $(\text{RGD-1})$ is then equivalent to
\[X = f^{\sharp} - \frac{f^{\sharp} \cdot \mathbb{1}}{n \cdot \mathbb{1}}n \tag{RGD-2}\]where $\mathbb{1} = (1, \dots, 1)$, and $n = \mathbb{1}^{\sharp}$ is a vector field normal to the simplex.
If $g$ is the Shahshahani metric45 \(g(x) = \sum_{i,j} \frac{\delta_{ij}}{x_i} dx^i \otimes dx^j\) then $(\text{RGD-2})$ becomes the replicator field6 \(X^i(x) = x^i \left( f_i(x) - \sum_j x^j f_j(x) \right)\) which says that the growth rate of a sub-population is equal to the difference between its fitness and the average fitness of the population.
The explicit expression of the fitness depends on the system under consideration. In this work we investigate the case in which the fitness depends linearly on the population state \(f(x) = Ax\) and the game is zero-sum, namely the fitness matrix $A$ is anti-symmetric.
A breakthrough paper by Akin and Losert[^Akin84] showed that anti-symmetric zero-sum replicator systems yield a symplectic foliation on the standard simplex and the following dichotomy holds: Either
Focusing on the first case, Alishah and Duarte8 showed that the symplectic foliation of the simplex arises from a stratified Poisson structure9 on the simplex, in which the replicator system is Hamiltonian. This result was recently generalised in the context of Dirac / Big-isotropic structures10.
In my master thesis (May 2020) I investigated the Hamiltonian nature of zero-sum replicator systems in presence of interior fixed points with respect to the cubic Poisson structure on the standard simplex described by Alishah and Duarte.
To this end some familiarity with Poisson manifolds is required, with particular focus on the methods of singular Poisson reduction, used to derive the stratified Poisson structure for the standard simplex. After a brief detour into classical concepts of Evolutionary Games and Population Dynamics, I proved the main result, namely the Hamiltonian character of the replicator vector field with respect to the derived Poisson structure, and discussed some simple examples.
In collaboration with Prof. Gabriele Benedetti I am working along the following directions:
Upon some homogeneity assumptions, the replicator field in the interior of the simplex $\mathring{\Delta}$ can be obtained from the field $x_i f_i(x)$ in $\mathbb{R}^n_{>0}$ via the Poisson quotient map $x \mapsto x / \left(\sum_i x_i \right)$ onto $\mathring{\Delta}$. We study the Hamiltonian structure of this system in $\mathbb{R}^n_{>0}$ with respect to a suitable Poisson structure, and motivate the appearance of the Kullback-Leibler divergence as Hamiltonian function from the fact that it is a Casimir function of such Poisson structure. Starting from an idea by Alishah and Duarte8 we furthermore developed an algorithm to generate a Hamiltonian zero-sum replicator system with any chosen fixed point.
Next, starting from an idea by Mertikopoulos and Sandholm3 on Bregman divergences, we are investigating whether such Hamiltonian description generalizes beyond the replicator case to arbitrary conservative Riemannian game dynamics, the answer appearing to be affirmative.
Finally, we want to find periodic orbits of zero-sum replicator systems with interior fixed points by variational methods based on Clarke’s duality, exploiting the convexity of the Hamiltonian function11. In low dimension one can exploit the convexity further to show the existence of a global surface of section for the flow using the work of Hofer, Wysocki and Zehnder12. In particular we want to
Here is an extremely non-exhaustive list of relevant works beyond the ones explicitly cited so far.
This research was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany ́s Excellence Strategy EXC 2181/1 - 390900948 (the Heidelberg STRUCTURES Excellence Cluster).
J. Hofbauer and K. Sigmund, Evolutionary games and population dynamics, Cambridge University Press, Cambridge, 1998. ↩
W. H. Sandholm. Population games and evolutionary dynamics. MIT press, 2010. ↩
P. Mertikopoulos and W. H. Sandholm. Riemannian game dynamics. Journal of Economic Theory, 177:315–364, 2018 ↩ ↩2
S. Shahshahani. A new mathematical framework for the study of linkage and selection. American Math- ematical Soc., 1979 ↩
E. Akin, The differential geometry of population genetics and evolutionary games. In Mathematical and Statistical Developments of Evolutionary Theory, pp. 1-93. Springer, Dordrecht, 1990. ↩
P. Schuster and K. Sigmund. Replicator dynamics. Journal of theoretical biology, 100(3):533–538, 1983 ↩
John C Baez and Blake S Pollard. “Relative entropy in biological systems”. In: Entropy 18.2 (2016), p. 46 ↩
H. N. Alishah and P. Duarte, Hamiltonian evolutionary games, J. Dyn. Games 2 (2015), no. 1, 33-49 ↩ ↩2
Juan-Pablo Ortega, Tudor S. Ratiu, and Rui Loja Fernandes. “The momentum map in Poisson geometry”;. In: American journal of mathematics 131.5 (2009), pp. 1261-1310 ↩
H. N. Alishah. Hamiltonian formalism of the inverse problem using dirac geometry and its application on linear systems, 2020 ↩
Ekeland, I. (2012). Convexity methods in Hamiltonian mechanics (Vol. 19). Springer Science & Business Media. ↩
Hofer, H., Wysocki, K., & Zehnder, E. (1998). The Dynamics on Three-Dimensional Strictly Convex Energy Surfaces. Annals of Mathematics, 148(1), second series, 197-289. doi:10.2307/120994 ↩
F. Savarino and C. Schnörr, Continuous-Domain Assignment Flows, preprint available at arXiv:1910.07287, 2019. ↩
Oechssler, J. and Riedel, F. (2001). Evolutionary dynamics on infinite strategy spaces. Economic theory, 17(1):141–162 ↩