Low-rank approximation of non-Hermitian operators from spectral entangling strength

A dimension-independent operator-norm approximation bound for every alpha below one third, with an optimal rank exponent and no Hermiticity or locality assumption.

Yuxuan Zhang · Version 1.0, 18 November 2026

Manuscript (PDF) · Source and supplementary checks (ZIP) · Release metadata

If an operator produces a rapidly decaying Schmidt spectrum from every product input, including inputs with local ancillas, must the operator itself be approximable by a short sum of product operators?

This manuscript gives a dimension-independent answer for arbitrary complex operators at sufficiently low spectral order. For every \(0<\alpha<1/3\), every operator \(X\) on finite-dimensional \(A\otimes B\), and every integer \(D\geq0\), it proves

\[\min_{\operatorname{OSR}(Y)\leq D}\|X-Y\|_{\mathrm{op}} \leq J_\alpha(X)(D+1)^{-\left(1/(2\alpha)-3/2\right)}.\]

Here \(\operatorname{OSR}\) is the number of product operators needed to write \(Y\). The strength \(J_\alpha(X)\) is the largest \(\ell_{2\alpha}\) norm of the unnormalized output Schmidt coefficients over all normalized product inputs with arbitrary local ancillas. Neither the constant nor the rank exponent depends on the local dimensions.

What the theorem adds

Kim et al., Spectral Small-Incremental Entangling, PRX 16, 041035 (2026) introduce this strength and discuss its relation to operator approximation. Their published Note added reports the Hermitian positive result and a non-Hermitian obstruction at \(\alpha\geq1/3\). The present proof supplies the positive guarantee for general non-Hermitian operators throughout \(0<\alpha<1/3\).

The rank-one obstruction already reported in that source also shows that the power \(1/(2\alpha)-3/2\) cannot be increased uniformly. The theorem’s upper constant one is not claimed to be optimal.

Priority remains unconfirmed. The source cites an unpublished follow-up by Kim, Kusunoki and Kuwahara as reference 93. Its complete manuscript was not found in the public sources checked on 18 November 2026, including the team’s publication and preprint lists. Possible overlap cannot be excluded. This release makes no claim of being the first solution.

Why the dimension factor disappears

Truncating the usual operator Schmidt decomposition controls a Hilbert–Schmidt error. That need not give the desired dimension-independent operator-norm bound. Instead, the proof chooses local input densities that maximize the output spectral strength, with a small positive floor on their eigenvalues. Undoing those input filters gives a tensor decomposition

\[X=\sum_s\lambda_s A_s\otimes B_s.\]

First variations of the two input densities bound both weighted column square sums, \(\sum_s\lambda_s^{2\alpha}A_s^\dagger A_s\) and \(\sum_s\lambda_s^{2\alpha}B_s^\dagger B_s\). A Cauchy–Schwarz estimate then controls the full operator norm of the discarded terms. The tail decays precisely when \(\alpha<1/3\). Finally, a compactness argument removes the positive floor while preserving the rank bound.

The estimates apply to arbitrary input vectors in the operator norm, including entangled ones. The derivative calculation also treats repeated and zero Schmidt coefficients: positive input filters keep the realignment rank fixed throughout the differentiation.

Accuracy and scope

For fixed \(\alpha\), the sufficient rank scales as

\[D=O_\alpha\!\left((J_\alpha(X)/\epsilon)^{2\alpha/(1-3\alpha)}\right).\]

For example, the accuracy exponent is \(2/3\) at \(\alpha=1/6\), \(2\) at \(\alpha=1/4\), and \(6\) at \(\alpha=3/10\). It diverges as \(\alpha\) approaches \(1/3\).

This is an existence and rank theorem. The proof does not give a polynomial-time method to find the globally maximizing inputs. The approximant need not be positive, Hermitian, unitary, or a physical channel. No locality assumption is imposed. The related bounded-strength evolution result has a different hypothesis and conclusion and is not used in this proof.

Checks and review status

The source package includes the full LaTeX manuscript, a source comparison, fixed-seed diagnostics, and a hash-verifying reproduction runner. Extract it into an empty directory and run python3 reproduce.py with NumPy and SciPy installed; the checks require no network access.

There are 12 complex factorization cases, 24 finite-difference gradient checks, two local stationarity diagnostics, and eight detected invalid substitutions. An exact rational control rejects an incorrectly stronger rank exponent. The local optimizations do not certify global maxima, and these finite checks do not replace the analytic proof.

The manuscript has received sequential internal proof and source review. Independent expert review and formal verification have not been performed. The author directed the research and manuscript preparation. AI assistance contributed to derivation, literature comparison, computation, checking and drafting.