Operator Schmidt rank under bounded spectral entangling strength
A squared-logarithmic rank bound for full quantum dynamics at fixed time and operator-norm accuracy, without spatial locality or bounds on local Hamiltonians.
Yuxuan Zhang · Version 1.0, 17 November 2026
Full manuscript (PDF) · Source and supplementary checks (ZIP) · Release metadata
An interaction may generate little entanglement from every product input, yet its full time-evolution operator can still require a growing number of product terms. How fast can that number grow with dimension?
For an arbitrary finite bipartite Hamiltonian
\[H=H_A\otimes I_B+I_A\otimes H_B+V, \qquad V=V^*,\qquad \mathcal J(V)\le1,\]the manuscript gives an operator-norm approximation satisfying
\[\left\|e^{-itH}-\widetilde U\right\|\le\epsilon, \qquad \operatorname{OSR}(\widetilde U) \le C_{t,\epsilon}\,[\log(2D)]^2, \qquad D=d_A d_B.\]Here operator Schmidt rank is the number of product operators needed to write the approximation. The strength \(\mathcal J\) is the largest sum of output Schmidt coefficients when the interaction acts on a normalized product input, including arbitrary local ancillas. Time and accuracy are fixed independently of dimension. Both local Hamiltonians are arbitrary; their norms need not be bounded across the family.
This is an existence theorem for approximating the entire evolution in operator norm. It does not provide an efficient algorithm for finding the factors, and the approximant need not itself be unitary.
The question it addresses
Kim et al., Spectral Small-Incremental Entangling, published in PRX 16, 041035 (2026), show in Proposition 9 that bounded spectral entangling strength does not guarantee a fixed rank independent of dimension. They then ask whether a fixed-time, fixed-error approximation can require rank growing as a positive power of \(D\).
The bound above answers that question negatively: for every fixed \(c>0\), \((\log D)^2=o(D^c)\). It is compatible with their example, which rules out constant rank and admits a logarithmic-dimensional approximation.
The step that controls dimension
The proof first establishes a uniform approximation theorem for Hermitian operators. If \(\mathcal J(V)\le1\), then for every \(\delta>0\) there is a nearby operator
\[W=\sum_s A_s\otimes B_s, \qquad \|V-W\|\le\delta,\]whose four local square sums \(\sum_s A_s^*A_s\), \(\sum_s A_sA_s^*\), \(\sum_s B_s^*B_s\), and \(\sum_s B_sB_s^*\) are all bounded by the same dimension-independent number. One explicit bound is \(O(\delta^{-4}\log^5(2/\delta))\).
A theorem of Arhancet and Kriegler supplies a completely positive dilation controlled in the Hilbert–Schmidt norm. Positive weights built from its diagonal maps define local dyadic spectral bands. The proof bounds each pair of bands in two ways: one estimate decays with the sum of their indices, the other with their difference. Together they control every direction in the index plane. The tail is summable without a bound on the weights’ condition numbers, which is where dimension disappears from the approximation error.
The evolution-strength inequality \(\mathcal J(e^{-itH})\le e^{|t|}\) is already Corollary 6 of Kim et al. Applying the new approximation theorem to the two Hermitian parts of the full unitary gives a balanced tensor decomposition. Correlated Gaussian local factors then produce unbiased rank-one product samples. Their two matrix second moments are bounded independently of dimension; clipping and matrix Bernstein concentration give the squared logarithm.
Parameters and relation to other results
With \(T=|t|\), \(q=1+T+\log(1/\epsilon)\), and \(0<\epsilon\le1\), the proof supplies the explicit, nonoptimal form
\[\operatorname{OSR}(\widetilde U) \le C e^{10T}\epsilon^{-10}q^{10} \log(2D)\,[\log(2D)+q].\]The universal constant is large. Improving the error dependence and deciding whether one logarithm always suffices remain quantitative questions.
Mackay et al., Optimal low-rank compression of quantum dynamics prove stronger compression results under spatial locality and bounded local-strength assumptions. The present result instead assumes only the bipartite strength bound. The \(\alpha<1/2\) result announced in the Note added of Kim et al. also has a different hypothesis from the endpoint treated here.
Supplementary checks and review status
The package includes the complete LaTeX source and two standalone checkers. Run python3 reproduce.py after extracting it into an empty directory. Python, NumPy, SciPy and SymPy are required; the calculations need no network connection. The runner verifies file hashes before executing the checks.
The local-band checker runs 374 finite assertions, using floating-point matrix controls and a rational large-star reduction. The direct-transfer checker runs 71 exact assertions, including finite quadrature for the Gaussian moment identities and checks of the error constants. The two checkers detect three and four classes of invalid substitution, respectively. These checks supplement the analytic proof; they do not certify an all-dimensional theorem by enumeration.
The manuscript has received sequential internal proof and source review. Independent expert review and formal verification have not been performed. The full unpublished work listed as reference 93 of Kim et al. was unavailable for comparison, and worldwide first-publication priority is not established. This is a research preprint, with no claim of journal or catalog acceptance.
The author directed the research and manuscript preparation. AI assistance contributed to derivation, literature comparison, implementation, checking and drafting.