\documentclass[10pt]{amsart} \usepackage[margin=27mm]{geometry} \usepackage[T1]{fontenc} \usepackage{lmodern,microtype,amsmath,amssymb} \usepackage[colorlinks=true,allcolors=blue]{hyperref} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \hypersetup{ pdftitle={Observation cost of nonequilibrium in nonnormal networks}, pdfauthor={Samuil Petkov} } \begin{document} \title[Observation cost of nonequilibrium in nonnormal networks]{Observation cost of nonequilibrium in nonnormal networks:\\ a matched exponent for an explicit four-state generator} \author{Samuil Petkov} \address{Department of Physics, \'Ecole normale sup\'erieure, Universit\'e PSL, Paris, France} \email{samuil.petkov@phys.ens.psl.eu} \date{1 November 2026} \begin{abstract} We consider detection of nonequilibrium from passive, causally scheduled snapshots of a stationary continuous-time Markov chain, with gaps between snapshots of at least $\tau$. False alarms must be controlled uniformly over every finite stationary detailed-balance chain observed through a deterministic labelling, with no restriction on hidden dimension, rates, stationary label law or mixing time. Aznagulov showed that for normal generators the optimal snapshot cost grows like $e^{2\gamma_{\rm irr}\tau}$, and that for nonnormal generators its exponent lies in a spectral window. For an explicit nonnormal four-state generator $Q_4$, with a real slow mode and residual damping $\beta$, we prove that $\tau^{-1}\log N^*\to2\gamma$, where $\gamma=(701-\sqrt{158761})/126\in(2,\beta)$ is a relaxation rate of an explicit reversible birth--death comparator; for $Q_4$ this exponent lies strictly inside the window $[2,2\beta]$. The comparator comes from an ordered-eigenfunction completion lemma: among $\pi$-reversible generators with a prescribed strictly monotone eigenfunction and eigenvalue, a birth--death generator maximizes the entire Dirichlet form. The upper bound uses symmetric two-lag scores whose stationary mean is nonpositive under every hidden reversible model, together with a positive-operator tilt lemma that gives the moment-generating bound $e^{32n\theta^2}$ for a deterministic number $n$ of dependent transitions, uniformly over the null class. We also show that the two-lag pair-law approximation exponent of $Q_4$ by hidden reversible models is $\gamma$, that the cost exponent $2\gamma$ holds uniformly over a four-parameter box of nonnormal deformations, and that it extends to interacting path families on $\{1,2,3,4\}^d$ built from tensor lifts. More generally, if $u$ has distinct entries and is an eigenvector of $Q$ and of its $L^2(\pi)$-adjoint with eigenvalue $-\lambda$, the exponent is $2\gamma^*$ whenever $\gamma^*<\beta$, where $\gamma^*$ is the restricted gap of the completion. \medskip \noindent\textbf{Mathematics Subject Classification (2020).} Primary 62M02; Secondary 60J27, 82C05. \noindent\textbf{Keywords.} nonequilibrium detection; detailed balance; nonnormal generators; continuous-time Markov chains; hypothesis testing; birth--death processes. \end{abstract} \maketitle \section{Introduction} What controls the optimal cost of detecting nonequilibrium when relaxation and circulation do not commute? We consider a finite irreducible row generator $Q$ started in its stationary law $\pi$, with all states even under time reversal. An observer records passive snapshots of the state at causally chosen times whose gaps are at least $\tau$, and must decide whether the record could have come from an equilibrium system. The null hypothesis contains all finite stationary detailed-balance chains with deterministic labels, without bounds on dimension, rates or mixing time. The question is the exponential rate in $\tau$ of the minimal expected number of snapshots, and the target is a matching lower and upper bound. Aznagulov \cite{Aznagulov2026} introduced this observation contract and null class. For an irreversible normal generator, the minimal expected record length is of order $e^{2\gamma_{\rm irr}\tau}\log(1/\alpha)$ \cite[Eq.~(3)]{Aznagulov2026}, where $\gamma_{\rm irr}$ is the smallest damping among non-real eigenvalues. For general $Q$, the exponent is bracketed between twice the spectral gap of the additive reversible part and twice the slowest damping among irreversibly visible eigenvalues, the upper bound being stated when the slowest visible level is a single real eigenvalue or a single complex pair \cite[Eq.~(12)]{Aznagulov2026}. The passive causal observation contract, the unrestricted stationary hidden-equilibrium null, the causal row-relative-entropy lower-bound method and the concentration of antisymmetric Poissonized currents \cite[Eq.~(8)]{Aznagulov2026} are taken from that work. This note determines the exponent for explicit nonnormal targets in which that window does not close. The main results are the following. \begin{enumerate} \item An ordered-eigenfunction completion lemma (Lemma~\ref{lem:completion}): given a stationary law and a strictly increasing mean-zero function $u$, an explicit birth--death generator $B$ with $Bu=-\lambda u$ maximizes the Dirichlet form among all reversible generators with the same eigenpair. \item For an explicit nonnormal generator $Q_4$ on four states, the two-lag pair-law approximation exponent by hidden reversible models equals $\gamma$, a relaxation rate of the completion $B$ (Theorem~\ref{thm:pair}). \item A positive-operator tilt lemma (Lemma~\ref{lem:posop}): if two bounded symmetric two-lag scores have nonpositive stationary mean under \emph{every} hidden reversible model, then their sum along one stationary dependent trajectory with a deterministic number of transitions has moment-generating function at most $e^{32n\theta^2}$, uniformly over the null class. \item The matched cost exponent $2\gamma$ for $Q_4$ (Theorem~\ref{thm:main}). For $Q_4$ the window of \cite[Eq.~(12)]{Aznagulov2026} is $[2,2\beta]$, and $2<2\gamma<2\beta$. \item The same exponent uniformly over a four-parameter box $|\theta|\le10^{-3}$ of nonnormal deformations of $Q_4$ (Theorem~\ref{thm:box}), and for interacting path families on $\{1,2,3,4\}^d$ obtained by tensor lifts (Theorem~\ref{thm:path}). \item For a class of generators with a two-sided slow eigenvector $Qu=Q^\sharp u=-\lambda u$, the exponent is $2\gamma^*$ whenever $\gamma^*<\beta$, where $\gamma^*$ is the restricted gap of the completion; the lower bound $2\min\{\beta,\gamma^*\}$ holds throughout the class, and this expression is not the exponent in general when $\gamma^*\ge\beta$ (Theorem~\ref{thm:class}). \end{enumerate} The scores in (3) are symmetric under endpoint exchange and generally nonzero on diagonal pairs, so they are outside the antisymmetric-score hypothesis of \cite[Eq.~(8)]{Aznagulov2026}; the lemma uses the quantifier over all hidden stationary laws, which permits reversible changes of stationary law in its proof. \medskip \noindent\textbf{Scope.} The exponents are proved for the displayed targets only: the four-state generator $Q_4$, the box around it, the path families built from it, and the class of Theorem~\ref{thm:class} in the regime $\gamma^*<\beta$. No general nonnormal spectral formula, sharp prefactors, or optimal dependence on $\alpha$ is asserted. For the path families, no simultaneous limit in $d$ and $\tau$ is claimed, the history-comparison constant depends on $d$, and general noncommuting interactions and unbounded degree are not covered. Our comparison with earlier work is limited to the works cited in Section~\ref{sec:prior}. \section{Setting} \label{sec:setting} All rates use one fixed physical time unit, and logarithms are natural. For a probability vector $\pi$ with positive entries, $\langle f,g\rangle_\pi=\sum_i\pi_if_ig_i$ and $\|f\|_\pi^2=\langle f,f\rangle_\pi$; $D_\pi=\operatorname{diag}(\pi)$, and $\mathbf 1$ is the constant vector. For a matrix, $\|\cdot\|_{\max}$ is the largest absolute entry. \begin{definition}[Observation contract and null class] \label{def:contract} Let $Q$ be a finite irreducible reciprocal row generator with stationary law $\pi$, started from $\pi$. A \emph{passive causal procedure with minimum gap $\tau$} records the state at an initial time and at subsequent times whose gaps are at least $\tau$; each gap, the decision to stop, and the final decision may depend on the past record and on private randomness. Let $N\in\{0,1,2,\ldots\}\cup\{\infty\}$ be the number of sampled transitions before stopping, with $N=\infty$ if the procedure never stops. The \emph{null class} $\mathcal H$ consists of all observations $\ell(Z_t)$ of finite stationary continuous-time Markov chains $Z$ satisfying detailed balance, through deterministic label maps $\ell$ into the observed state space; the hidden dimension, rates, stationary label marginal and mixing time are unrestricted, and reducible stationary models are permitted. For $0<\alpha<1/2$, $N^*(Q,\tau,\alpha)$ is the infimum of $\mathbb E_QN$ over passive causal procedures with minimum gap $\tau$ whose probability of stopping and rejecting equilibrium is at most $\alpha$ under every member of $\mathcal H$ and at least $1-\alpha$ under $Q$, with $\inf\emptyset=+\infty$. \end{definition} Counting the initial snapshot as well gives $N+1$ observations; this additive change does not affect any logarithmic exponent below. \begin{remark}[Comparison with the contract of \cite{Aznagulov2026}] \label{rem:contract} In \cite{Aznagulov2026}, $N^*(Q,\tau_0,\alpha)$ is the least expected number $\mathbb E_QN$ of observed transitions over tests with false-positive probability at most $\alpha$ for every hidden equilibrium model and false-negative probability at most $\alpha$ under $Q$; gaps are at least $\tau_0$, are chosen from the past record and private randomness, and the process starts in stationarity. Definition~\ref{def:contract} counts the same quantity under the same two error constraints. The requirement ``rejects with probability at least $1-\alpha$ under $Q$'' coincides with ``false-negative probability at most $\alpha$'' for every procedure with $\mathbb E_QN<\infty$, which are the only procedures that affect the infimum when it is finite. Allowing reducible stationary hidden models can only enlarge $\mathcal H$, hence only increase $N^*$. Every lower bound below uses an irreducible hidden model, and every upper bound controls the enlarged class, so all exponents below hold for either reading of the null class. Our results are stated for $0<\alpha<1/2$; \cite[Eq.~(3)]{Aznagulov2026} is stated for $0<\alpha\le1/4$. \end{remark} We use the following elementary bound for relative entropy: for probability vectors $p,r$ with $r>0$, \begin{equation} D_{\rm KL}(p\Vert r)\le\sum_j\frac{(p_j-r_j)^2}{r_j}. \label{eq:chisq} \end{equation} \section{The ordered-eigenfunction completion lemma} \begin{lemma}[Ordered completion] \label{lem:completion} Let $n\ge3$, $\pi_i>0$, $\sum_i\pi_i=1$, and $u_1<\cdots0$. Define a reversible birth--death generator $B$ by its conductances \[ c_k=\pi_k B_{k,k+1}=\pi_{k+1}B_{k+1,k} =-\lambda\frac{\sum_{i\le k}\pi_i u_i}{u_{k+1}-u_k}, \] with no other off-diagonal edges and rows summing to zero. Then $c_k>0$, $Bu=-\lambda u$, and every $\pi$-reversible generator $R$ satisfying $Ru=-\lambda u$ obeys \[ \langle v,-Rv\rangle_\pi\le\langle v,-Bv\rangle_\pi \quad\hbox{for every real }v. \] Thus $B$ maximizes the entire Dirichlet form within this constrained class, and in particular maximizes the gap restricted to $\{\mathbf1,u\}^{\perp_\pi}$. \end{lemma} \begin{proof} Every proper prefix $S_k=\sum_{i\le k}\pi_i u_i$ is negative, since the weighted mean on a strict lower segment is below the full mean zero. Hence $c_k>0$. Taking $S_0=S_n=0$, direct subtraction gives \[ (-Bu)_i=\frac{-\lambda S_{i-1}+\lambda S_i}{\pi_i}=\lambda u_i. \] For any $v$, choose $z$ up to a constant by \[ z_{k+1}-z_k=\frac{(v_{k+1}-v_k)^2}{u_{k+1}-u_k}. \] Weighted Cauchy--Schwarz on consecutive increments yields \[ (v_j-v_i)^2\le(u_j-u_i)(z_j-z_i),\qquad i0. \] Consequently every same-state generator $R$ reversible with respect to the same uniform $\pi$ and retaining $Ru=-u$ has restricted gap at most $42611/17745$, below the residual damping of $Q_4$. \end{proposition} The proposition concerns generators on the same four states. The hidden models in the remaining sections are unrestricted, and the proposition is not used in the statistical results. \section{Two-lag approximation by unrestricted hidden equilibrium models} \label{sec:pair} For any finite stationary detailed-balance hidden CTMC $\mathcal R$ and deterministic label map into $\{1,2,3,4\}$, let \[ J_{\mathcal R}(t)_{ij} =\Pr\{\ell(Z_0)=i,\ell(Z_t)=j\}. \] The stationary label marginal need not equal $\pi$. Define \[ d_2(\tau)=\inf_{\mathcal R}\max_{k=1,2} \left\|J_{\mathcal R}(k\tau)-D_\pi e^{k\tau Q_4}\right\|_{\max}. \] The same model supplies both lags, but its dimension, rates, stationary law and labeling may change with $\tau$. Reducible stationary models are permitted. \begin{theorem}[Pair-approximation exponent] \label{thm:pair} For this fixed target and null class, \[ \lim_{\tau\to\infty}-\frac1\tau\log d_2(\tau)=\gamma. \] \end{theorem} This statement concerns joint pair expectations, not statistical detection on a dependent trajectory; the statistical results are in Sections~\ref{sec:posop}--\ref{sec:main}. \begin{proof} \emph{Upper approximation bound.} The identity-labeled generator $B$ is admissible. It has the same projectors $\Pi,U$ as $Q_4$ and residual decay at rates $\gamma,\Gamma$. Since $\beta>\gamma$, their finite spectral decompositions give \begin{equation} \|e^{tQ_4}-e^{tB}\|_{\max}\le C_0e^{-\gamma t}, \qquad d_2(\tau)\le C_0'e^{-\gamma\tau}. \label{eq:upperapprox} \end{equation} \emph{A fixed pointwise certificate.} Choose a real vector $w$ with $Bw=-\gamma w$ and $\|w\|_\pi=1$. The distinct eigenvalues and symmetry of $B$ imply $w\perp_\pi\mathbf1,u$ and $Fw=w$. Define $\psi_1=0$ and \[ \psi_{k+1}-\psi_k =\frac{(w_{k+1}-w_k)^2}{u_{k+1}-u_k}. \] Weighted Cauchy--Schwarz gives \begin{equation} (w_i-w_j)^2\le(u_i-u_j)(\psi_i-\psi_j) \quad\hbox{for every }i,j. \label{eq:certificate} \end{equation} Using $-Bu=u$, consecutive-edge equality and $-Bw=\gamma w$ gives \[ \langle\psi,u\rangle_\pi =\langle\psi,-Bu\rangle_\pi =\sum_{k=1}^3 c_k(\psi_{k+1}-\psi_k)(u_{k+1}-u_k) =\langle w,-Bw\rangle_\pi=\gamma. \] \emph{Uniform consequences of a putative faster approximation.} Fix $\gamma