Convergence of a particular fixed point iteration scheme

Setup

I have the following non-linear system of equations: $$\mathbf{x} P(\mathbf{x}) = 0$$ where

• $\mathbf{x} \in \mathbb{R}_{>0}^n$ is a probability distribution, i.e., $\sum_i x_i = 1$, and
• $P(\mathbf{x})$ is an $n \times n$ matrix that depends on $\mathbf{x}$ (hence the non-linearity.) $P$ is the transition rate (infinitesimal generator) matrix of a Markov chain, but it is probably not important.
• Not sure if important: $P_{ij}, i \ne j$ is a strictly convex function of $\mathbf{x}$, and $P_{ii} = - \sum_j P_{ij}$ is concave.

Furthermore I know the following things.

• This system has a unique fixed / stationary point, and
• The dynamical system $\mathbf{\dot{x}} = \mathbf{x} P(\mathbf{x})$ converges to this fixed point from any starting point.

Problem

I want to show that the following iterative scheme converges to the fixed point.

1. pick any initial $\mathbf{x}_1$, e.g., $\mathbf{x}_1 = [1/n, \ldots, 1/n]$
2. given $\mathbf{x}_k$, find $\mathbf{x}_{(k+1)}$ by setting $P_k = P(\mathbf{x_k})$ solving the linear system $\mathbf{x} P_k = 0$

I have done a lot of numerical simulations to convince myself that this iterative scheme converges, but I would like to prove it formally.

Any help is appreciated!

• Have you tried looking into uniformization? I think I remember that there are results which show that the uniformized Markov chain (your linear system) converges to the same stationary distribution as the continuous-time Markov chain given the uniformization constant is sufficiently large. – Bernhard Jun 23 '15 at 11:48
• @BernhardGeiger I tried going through a discrete-time Markov chain formulation using uniformization, but it didn't help. Basically, letting $\tilde{P}$ be the matrix of transition probabilities obtained with uniformization, solving the linear system reduces to a matrix multiplication $\mathbf{x}Q$, where $Q = \lim_{m \to \infty} \tilde{P}^m$. But I couldn't manage to analyze $Q$... – lum Jun 23 '15 at 12:00
• Hmm, I see. But you want to show that your iterative scheme converges, right? I think the proof for convergence is possible via uniformization and related results. You would still need to let the iterative scheme run to get the invariant distribution... Plus, if you know that there is a unique fixed point, then the matrix $Q$ should have all its rows being equal to this fixed point $\mathbf{x}$, shouldn't it? – Bernhard Jun 23 '15 at 12:28
• But $Q$ is a function of $\mathbf{x}$. The key part is to show that the sequence of linearizations leads to the desired fixed point (uniformization or not.) But maybe I'm not getting your point? If you want to expand on it in an answer (even incomplete), I'd be happy to have a deeper look! – lum Jun 23 '15 at 13:08

I have to admit that I don't have the answer, it is just an educated guess: I just recommended to look into the proofs for uniformization, because there you also solve for the fixed-point of a dynamical system (CTMC) via linearization (DTMC). I think it might be possible to use uniformization also for non-homogeneous Markov chains, as in your case, but there you should look into the proofs.

With non-homogeneity I mean what you call non-linearity: Let $\tilde{P}$ be a function of $\mathbf{x}_{k}$, i.e., include the non-linearity in uniformization. You might need to choose the uniformization constant properly, i.e., according to the extreme cases for $\mathbf{x}_{k}$. If you can then show that none of your changes (w.r.t. standard uniformization) violates conditions made in the proofs, then you know that your iterative scheme converges to the fixed point you are interested in.

On top of that, it might be worth looking into discretization of dynamical systems...

• Great, thanks. I'll have a deeper look over the next days and report back ;-) – lum Jun 24 '15 at 19:55
• Any progress? Just curious... – Bernhard Jul 1 '15 at 7:12
• I'll write my progress so far as an answer. – lum Jul 7 '15 at 8:49

I tried to follow @Bernhard's recommendations, and this is the progress I made so far. Uniformization in itself didn't help me much. The CTMC converges, so clearly the corresponding uniformized (discrete) chain converges as well.

Consider the uniformized (linearized) chain. Instead of making a single "small" change at each discrete step (direct consequence of the small uniformization constant), I'm letting this linearized chain run until convergence. Then, uniformize again, and repeat the process.

However, @Bernhard's answer prompted me to look into results on nonhomogeneous / nonlinear Markov chains.

1. Nonhomogeneous Markov chains. There's a flurry of results, but unfortunately none of them seems to apply to my case. In particular, most results require asymptotical homogeneity of the transition matrix, so it doesn't advance me much (it is clear that asymptotical homogeneity is equivalent to convergence my method!) See e.g. Seneta's book, Non-negative Matrices and Markov Chains

2. Believe it or not, but there is actually some recent theory on nonlinear Markov chains. In a few words, it's a stochastic "way of thinking" of any mapping in the probability simplex. Unfortunately, it is difficult to check the conditions for convergence, see On Ergodic Properties of Nonlinear Markov Chains

In summary: nothing conclusive so far!

• Thanks for the clarification...Sorry that I could not help! – Bernhard Jul 7 '15 at 11:03
• I'm very thankful still - it pointed me in an interesting direction. I haven't abandoned all hope yet ;-) – lum Jul 7 '15 at 11:44