In chaotic quantum dots, quantum wave effects such as weak localization, conductance fluctuations, and time-delay statistics are computed with a semiclassical diagrammatic expansion. This method writes each measurable quantity as a sum over pairs of classical trajectories. It organizes the correlated pairs into topological diagrams whose contributions shrink as powers of $1/N$, where $N$ is the number of open scattering channels. The expansion is exact in principle, but the number of diagrams grows factorially with order. In practice, people stop at some order and hope the remainder is small. Nobody tracks how much error that stopping choice injects when the series is used repeatedly, for example inside an optimizer that fits cavity parameters to measured transport statistics.
Optimization theory has a mature answer to exactly this situation. The field of inexact subproblem solving studies what happens when an inner computation (a proximal step, a lower-level solve) is only approximate. It shows that an outer algorithm keeps its convergence rate if the inner errors shrink on a prescribed schedule and can be certified by a computable quantity such as a duality gap. This note proposes that the truncated semiclassical series be treated as such an inexact inner oracle. It also proposes that its missing "duality gap" be supplied by physics. A truncated series stops respecting unitarity, meaning conservation of probability for the scattering matrix. The size of that violation can be computed from the diagrams already evaluated, and we propose to use it as the certificate that decides when to add the next order.
Core Claim
Take a truncated semiclassical expansion at order $K$. Evaluate two things independently from diagrams: a transmission quantity and its reflection or sum-rule partner. The mismatch between them is a free, computable residual. We claim this residual tracks the true truncation error well enough to drive a provably convergent tolerance schedule in an outer inverse-fitting loop. We further claim it does so better than the standard rule for truncating asymptotic series, which is to watch the size of the last term.
Concretely, lead $1$ has $N_1$ channels and the transmission block is $T$. The truncated series gives $\widehat{\mathrm{Tr}\,T^n}^{(K)}$ and, from separate reflection diagrams (including coherent backscattering), $\widehat{\mathrm{Tr}\,R^n}^{(K)}$. Unitarity forces $R = 1 - T$ on that block, so we define
$$ r_K(\theta) = \Big|\, \widehat{\mathrm{Tr}\,R^n}^{(K)} - \sum_{j=0}^{n}\binom{n}{j}(-1)^j\, \widehat{\mathrm{Tr}\,T^j}^{(K)} \Big| , $$
where $\theta$ collects the physical parameters being fitted, for example tunnel-barrier reflectivities or the ratio $\tau_E/\tau_D$ of Ehrenfest time to dwell time. The analogous residual for time delays uses the exactly known mean Wigner time $\tau_D$. The residual is zero for the full series, and it is computable at every order.
The outer loop increases $K_k$ at iteration $k$ only when $r_{K_k}$ exceeds a tolerance $\epsilon_k$. That tolerance follows the decay the inexact-optimization literature requires for rate preservation, for instance
$$ \sum_k \epsilon_k < \infty \quad\text{(plain)}, \qquad \epsilon_k = O(k^{-2-\delta}) \quad\text{(accelerated)} . $$
The first condition comes from the summable-error criteria of New inertial proximal gradient methods for unconstrained convex optimization problems. The second comes from the rate-preservation results of Inexact first-order primal-dual algorithms.
Why This Intersection, and What Is New
The file's search found no prior work joining the two fields. Two adjacent papers set the bar.
A frozen Gaussian approximation-based multi-level particle swarm optimization for seismic inversion is the closest precedent. It uses a semiclassical wave solver at two fixed fidelities, 300 versus 3,000 Gaussian wave packets, inside a population search, and switches between them on a heuristic percentage schedule. It motivates our proposal, because variable fidelity saved most of the compute. It does not supply our method: it has no certificate of forward-model error, no convergence guarantee, and no diagrammatic series.
Improved sampling and validation of frozen Gaussian approximation with surface hopping algorithm for nonadiabatic dynamics controls error in a semiclassical path sum by Monte Carlo pruning and branching. It stays within $O(M^{-1/2})$ sampling error in the number of sampled paths $M$, and it has no outer optimizer.
Our proposal differs from both in three ways. The fidelity knob is diagram order rather than sample count or packet count. The switching rule is a physics-derived certificate rather than a fixed schedule. The outer method is a first-order algorithm with inexact-oracle guarantees.
On the physics side, several papers supply the method. Universality in chaotic quantum transport: The concordance between random-matrix and semiclassical theories shows that untying encounters corresponds to multiplying permutations by transpositions. This gives an enumerable recursion for generating diagrams order by order, which we reuse. Transport moments beyond the leading order supplies the ribbon-graph bookkeeping, in which trajectories are edges and the regions where they run close together ("encounters") are vertices.
Other papers motivate the proposal. Efficient semiclassical approach for time delays states that time-delay diagrams were enumerated only to fifth order in $1/N$ because of combinatorial explosion. Density of states of chaotic Andreev billiards reports the same blockage. Together they show that order is a costly resource worth rationing.
The Berkolaiko–Kuipers result is also the main threat. Exact agreement with random matrix theory relies on massive cancellations between diagrams, so partial sums can misbehave. That same fact is our argument against the last-term rule: a term can be small because of cancellation, not because the remainder is small. Unitarity residuals test exactly the conservation law that those cancellations enforce, which is the untested link the file flags as Gap 4.
Research Program
We will build a diagram generator from the permutation–transposition recursion. It will cover the unitary and orthogonal symmetry classes, transmission moments $\mathrm{Tr}\,T^n$ for $n \le 4$, and Wigner time-delay moments. We will extend it to leads with tunnel barriers, where terms are expanded jointly in $1/N$ and barrier reflectivity following Conductance fluctuations in chaotic systems with tunnel barriers and Effect of a tunnel barrier on time delay statistics. We will also extend it to finite Ehrenfest time, where diagrams acquire $\exp(-\tau_E/\tau_D)$-type factors following Ehrenfest-time dependence of quantum transport corrections and spectral statistics.
Each truncated observable is a closed-form function of $\theta$. Its gradient with respect to $\theta$ is therefore exact at fixed $K$, and the only inexactness is truncation.
The outer problem is a regularized least-squares fit of $\theta$ to a target set of moments, with box constraints handled by a proximal step. It is nonconvex, so we use the adaptive-momentum proximal method APGnc+ from Convergence Analysis of Proximal Gradient with Momentum for Nonconvex Optimization. That method tolerates inexact gradients whose errors diminish at a controlled rate.
The analytic contribution is a lemma that bounds truncation error in the objective gradient by a constant times $r_K$ on a stated parameter region. We expect it to hold for $N$ large relative to $n$, where the $1/N$ series for fixed-$N$ rational moments converges. Where the lemma cannot be proved, the constant is estimated empirically and reported as such.
Baselines
The naive reference fixes the lowest order that includes quantum interference, meaning the diagonal term plus weak localization, for every outer iteration. It is the cheapest option and shows how much bias a common physics default leaves in fitted parameters.
The "exact" reference fixes the highest order our generator reaches within a set wall-clock budget for all iterations. It plays the role of solving the inner problem to machine precision, and it defines the accuracy target.
The heuristic multi-fidelity baseline switches from low to high order at a fixed fraction of the iteration budget, as in A frozen Gaussian approximation-based multi-level particle swarm optimization for seismic inversion. That paper reports 56 hours for the always-high-fidelity run versus 6 hours for the multi-level schedule on the smoothed Marmousi model, at comparable best objective value. These figures are for seismic inversion, not quantum transport. For truncated semiclassical series in a fitting loop, no reported result was found on Lacuna.
The strongest competitor on our specific question is the classical asymptotic-series rule: add orders until the magnitude of the last computed term falls below $\epsilon_k$. It uses the same schedule, the same outer optimizer, and costs nothing extra, so it isolates whether the unitarity residual is the ingredient that matters.
A stochastic baseline replaces the series with Monte Carlo sampling of random scattering matrices from the matching ensemble, including barrier-dressed variants. It uses stochastic gradients at matched wall-clock.
Finally, an oracle ablation drives the schedule with the true truncation error, which is known wherever closed-form random-matrix results exist. This gives an upper bound on what any certificate could achieve.
All methods share the same targets, parameter initializations (20 seeds), outer optimizer, step-size tuning grid, and total budget measured in diagram evaluations and wall-clock. Beating the strongest competitor means reaching the exact reference's parameter error with at least 2× fewer diagram evaluations than the last-term rule, with no seed diverging. A secondary requirement is that the residual-driven schedule's cost be within 1.5× of the oracle schedule.
Why not just…?
Why not just use random matrix theory directly? In the ergodic, ideal-lead regime, closed-form moments exist and would beat any series. We use that regime only as a testbed where the truth is known. The proposal matters where random matrix theory does not apply: finite Ehrenfest time, where results depend on system-specific trajectory structure (Ehrenfest-time dependence of counting statistics for chaotic ballistic systems), and secondary gaps in Andreev billiards. For tunnel barriers, ensemble sampling does apply and is a real threat at small $N$. Our stochastic baseline measures the crossover in $N$ where its sampling noise and matrix-diagonalization cost exceed series cost. If no such crossover appears, we will restrict our claims to the Ehrenfest regime.
Why not just truncate when the last term is small? This rule is standard, and it is our primary ablation. Cancellations between diagrams can make individual orders small or even zero at some orders for certain symmetry classes while the remainder stays large. The rule would then stop too early. If the experiments show the last-term rule matches the unitarity residual in cost and reliability, the certificate adds nothing, and we will report that as the finding.
Why not just always use the highest affordable order? Diagram count grows factorially (Efficient semiclassical approach for time delays), so early iterations far from the solution waste most of their cost. The inexact-optimization results say those iterations tolerate larger error without losing the outer convergence rate. The fixed-high-order baseline quantifies this waste directly.
Decisive Checks
The first check comes before any optimization. Across a grid of $N$, $n$, symmetry class, barrier reflectivity, and $\tau_E/\tau_D$, we compute $r_K$ and the true truncation error, measured against closed forms or against numerical simulation of open quantum maps with matched parameters. We then report their rank correlation and the fraction of cases in which $c\,r_K$ upper-bounds the true error for a single fitted constant $c$. If that fraction is below 90%, or if the rank correlation is no better than that of the last-term magnitude, the core claim fails.
The second check runs the full fitting comparison on two inverse problems. The first recovers per-lead barrier reflectivities from time-delay moments. The second recovers $\tau_E/\tau_D$ from transmission moments. The metrics are cost to reach the target parameter error and final bias.
The third check is a stress test near small $N$, where the $1/N$ series may diverge. There we expect the residual to grow rather than shrink. Showing that it signals divergence, and so warns the optimizer instead of silently corrupting it, is part of the claim.
References
- New inertial proximal gradient methods for unconstrained convex optimization problems
- Inexact first-order primal-dual algorithms
- A frozen Gaussian approximation-based multi-level particle swarm optimization for seismic inversion
- Improved sampling and validation of frozen Gaussian approximation with surface hopping algorithm for nonadiabatic dynamics
- Universality in chaotic quantum transport: The concordance between random-matrix and semiclassical theories
- Transport moments beyond the leading order
- Efficient semiclassical approach for time delays
- Density of states of chaotic Andreev billiards
- Conductance fluctuations in chaotic systems with tunnel barriers
- Effect of a tunnel barrier on time delay statistics
- Ehrenfest-time dependence of quantum transport corrections and spectral statistics
- Convergence Analysis of Proximal Gradient with Momentum for Nonconvex Optimization
- Ehrenfest-time dependence of counting statistics for chaotic ballistic systems