Sharp Norm Approximation of Covering Markov Operators

Abstract

We study the operator-norm approximation of Markov operators associated with finite graph coverings. Let $q:Y\to X$ be a connected covering of degree $d\geq 2$ between finite connected graphs, and let $A_q:C(Y)\to C(X)$ be its fiber-averaging operator. We show that $A_q$ is not a finite convex combination of extreme Markov operators, although such combinations approximate $A_q$ arbitrarily well in operator norm. We determine the exact distances from $A_q$ to convex combinations of at most $m$ extreme Markov operators, with arbitrary or equal convex weights. If $m=kd+r$, where $k\geq 0$ and $1\leq r\leq d$, these distances are $2(d-r+1)/[d(k(d-r+1)+1)]$ and $2r(d-r+1)/(dm)$, respectively. Both distances are attained and depend only on the covering degree $d$ and the number of terms $m$. We also determine when equal weights are optimal and quantify the additional approximation error when they are not. Our proofs use total-variation estimates and continuous rearrangements of labelled weighted atoms. In particular, the unrestricted errors have a common scaled limit, while the equal-weight errors exhibit residue-dependent asymptotic behavior.