Finite Calls to Stability | 31 Oct. 2019 |

Consider an unknown dynamical system x_{k+1}=Ax_k parametrized by Ain mathsf{Sym}(n,mathbf{R}). In this short note we look at a relatively simple, yet neat method to assess stability of A via finite calls to a particular oracle. Here, we will follow the notes by Roman Vershynin (RV2011).

Assume we have no knowledge of A, but we are able to obtain langle Ay, y rangle for any desired yin mathbf{R}^n. It turns out that we can use this oracle to bound rho(A) from above.

First, recall that since A is symmetric, |A|_2=sigma_{mathrm{max}}(A)=|lambda_{mathrm{max}}(A)|=rho(A). Now, the common definition of |A|_2 is sup_{xin mathbf{R}^nsetminus{0}}|Ax|_2/|x|_2. The constraint set mathbf{R}^nsetminus{0} is not per se convenient from a computational point of view. Instead, we can look at a compact constraint set: |A|_2=sup_{xin mathbf{S}^{n-1}} |Ax|_2, where mathbf{S}^{n-1} is the sphere in mathbf{R}^n. The latter formulation has an advantage, we can discretize these compact constraints sets and tame functions over the individual chunks.

To do this formally, we can use varepsilon-nets, which are convenient in quantitative analysis on compact metric spaces (X,d). Let mathcal{N}_{varepsilon}subseteq X be a varepsilon-net for X if forall xin X;exists yin mathcal{N}_{varepsilon}:d(x,y)leq varepsilon. Hence, we measure the amount of balls to cover X.

For example, take (mathbf{S}^{n-1},ell^n_2) as our metric space where we want to bound |mathcal{N}_{varepsilon}| (the cardinality, i.e., the amount of balls). Then, to construct a minimal set of balls, consider an varepsilon-net of balls B_{varepsilon/2}(y), with d(y_1,y_2)geq varepsilon;forall y_1,y_2in mathcal{N}_{varepsilon}. Now the insight is simply that mathrm{vol}(B_{varepsilon/2})|mathcal{N}_{varepsilon}|leq mathrm{vol}(B_{1+varepsilon/2})-mathrm{vol}(B_{1-varepsilon/2}). Luckily, we understand the volume of n-dimensional Euclidean balls well such that we can establish

     |mathcal{N}_{varepsilon}| leq frac{mathrm{vol}(B_{1+varepsilon/2})-mathrm{vol}(B_{1-varepsilon/2})}{mathrm{vol}(B_{varepsilon/2})}= (1+2/varepsilon)^n-(1-2/varepsilon)^n.

Which is indeed sharper than the bound from Lemma 5.2 in (RV2011).

Next, we recite Lemma 5.4 from (RV2011). Given an varepsilon-net for mathbf{S}^{n-1} with varepsilonin[0,1) we have

     |A|_2 = sup_{xin mathbf{S}^{n-1}}|langle Ax,xrangle | leq frac{1}{(1-2varepsilon)}sup_{xin mathcal{N}_{varepsilon}}|langle Ax,x rangle |

This can be shown using the Cauchy-Schwartz inequality. Now, it would be interesting to see how sharp this bound is. To that end, we do a 2-dimensional example (x=(x_1,x_2)). Note, here we can easily find an optimal grid using polar coordinates. Let the unknown system be parametrized by

   A = left[   begin{array}{ll}   0.1 & -0.75    -0.75 & 0.1    end{array}right]

such that rho(A)=0.85. In the figures below we compute the upper-bound to |A|_2 for several nets mathcal{N}_{varepsilon}. Indeed, for finer nets, we see the upper-bound approaching 0.85. However, we see that the convergence is slow while being inversely related to the exponential growth in |mathcal{N}_{varepsilon}|.


Overall, we see that after about a 100 calls to our oracle we can be sure about A being exponentially stable. Here we just highlighted an example from (RV2011), at a later moment we will discuss its use in probability theory.