## Archive for March, 2014

### Lyapunov exponents

March 25, 2014

In the literature on chaos and strange attractors one concept which plays a prominent role is that of Lyapunov exponents. I came across this repeatedly but never understood the definition. I think I have now understood the reason that I had problems. The fact that I have spent so much time with differential geometry in the past years sometimes makes me see the mathematical world through ‘differential geometric spectacles’. I felt that dynamical systems were objects which naturally live on smooth manifolds and that the definition of important concepts should not be dependent on coordinates or the presence of a preferred metric. The usual definitions of Lyapunov exponents appeared to me strongly tied to Euclidean space although I had seen a couple of comments in the literature (without further justification) that the definition was coordinate independent. In what follows I will give a manifestly coordinate independent definition. Incidentally, the definition of Lyapunov exponents is not invariant under changes of the time coordinate and at one time this led to confusion among people studying chaos in cosmological models. The source of the problem was the lack of a preferred time coordinate in general relativity.

Consider a dynamical system defined by a smooth vector field on a manifold $M$. Let $A$ be a compact subset of the manifold which is invariant under the flow $\phi$ generated by the vector field. The aim here is to define the maximum Lyapunov exponent of a point $x_0\in A$. The derivative of the flow, $J_t=D_x\phi (t,x_0)$ is a linear mapping from $T_{x_0}M$ to $T_{\phi (t,x_0)}M$. In the Euclidean space picture $J_t$ is treated as a matrix and this matrix is multiplied by its transpose. What is this transpose in an invariant setting? It could be taken to be the mapping from $T^*_{\phi (t,x_0)}M$ to $T^*_{x_0}M$ naturally associated to $J_t$ by duality. The product of the matrices could be associated with the composition of the linear mappings but unfortunately the domains and ranges do not match. To overcome this I introduce a Riemannian metric $g$ on a neighbourhood of $A$. It is then necessary to show at the end of the day that the result does not depend on the metric. The key input for this is that since $A$ is compact the restrictions of any two metrics $g_1$ and $g_2$ to $A$ are uniformly equivalent. In other words, there exists a positive constant $C$ such that $C^{-1}g_1(v,v)\le g_2(v,v)\le Cg_1(v,v)$ for all tangent vectors $v$ at points of $A$. Once the metric $g$ has been chosen it can be used to identify the tangent and cotangent spaces with each other at the points $x_0$ and $\phi (t,x_0)$ and thus to compose $J_t$ and its ‘transpose’ to get a linear mapping $B(t)$ on the vector space $T_{x_0}M$. This vector space does not depend on $t$. The eigenvalues $\lambda_i (t)$ of the mapping $B(t)$ are easily shown to be positive. The maximum Lyapunov exponent is the maximum over $i$ of the limes superior for $t\to\infty$ of $\frac{1}{t}$ times the logarithm of $\sqrt{\lambda_i (t)}$. Note that the ambiguity of a multiplicative constant in the definition of $B(t)$ becomes an ambiguity of an additive constant in the definition of the logarithms and because of the factor $\frac{1}{t}$ this has no effect on the end result.

In general if the maximum Lyapunov exponent at a point $x_0$ is positive this is regarded as a sign of instability of the solution starting at that point (sensitive dependence on initial conditions) and if the exponent is negative this is regarded as a sign of stability. Unfortunately in general these criteria are not reliable, a fact which is known as the Perron effect. This is connected with the question of reducing the study of the asymptotic behaviour of a non-autonomous linear system of ODE to that of the autonomous systems obtained by freezing the coefficients at fixed times.