arxiv: 2604.21210 · v1 · submitted 2026-04-23 · 🪐 quant-ph · cs.LG

Recognition: unknown

The Feedback Hamiltonian is the Score Function: A Diffusion-Model Framework for Quantum Trajectory Reversal

Sagar Dubey , Alan John

Authors on Pith no claims yet

Pith reviewed 2026-05-09 22:47 UTC · model grok-4.3

classification 🪐 quant-ph cs.LG

keywords quantum trajectoriesfeedback controlscore functiondiffusion modelstime reversalGirsanov theoremcontinuous monitoring

0 comments

The pith

The feedback Hamiltonian equals the score function of the quantum trajectory distribution.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proves that the García-Pintos feedback Hamiltonian H_meas = r A / τ is identical to the functional derivative of the log path probability with respect to the density matrix. This equivalence shows why that Hamiltonian tilts the distribution of continuously monitored quantum trajectories and statistically reverses their order in time. The result also imports score-based machine learning methods to estimate the same object when real experiments deviate from perfect efficiency, zero delay, or Gaussian noise.

Core claim

We prove that δ log P_F / δ ρ = r A / τ = H_meas by applying Girsanov's theorem to the measurement record, Fréchet differentiation in the space of trace-class operators, and Kähler geometry on the pure-state manifold. The García-Pintos feedback Hamiltonian is thereby the score function of the quantum trajectory distribution, exactly the quantity Anderson's reverse-time diffusion theorem requires for trajectory reversal. The equality extends to multi-qubit systems with independent channels, where the score decomposes as a sum of local operators.

What carries the argument

The functional derivative δ log P_F / δ ρ of the forward path probability, proven equal to the feedback Hamiltonian r A / τ

If this is right

Varying the feedback gain X produces a continuous one-parameter family of path measures whenever [H, A] is nonzero.
X = -2 recovers the backward process in leading-order linearization around the forward dynamics.
Machine-learning score estimators can replace the analytic formula when unit efficiency, zero delay, or Gaussian noise fail.
In multi-qubit systems the score remains a sum of local operators acting on independent measurement channels.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Score-matching algorithms from diffusion models could be trained directly on quantum trajectory data to learn reversal Hamiltonians without explicit analytic expressions.
The continuous family of path measures suggests a tunable interpolation between forward and reversed dynamics that might be tested in quantum thermodynamics or information erasure protocols.
If learned scores remain effective under time delay or lossy detection, the framework could extend reversal techniques to a broader class of open quantum systems.

Load-bearing premise

The derivation assumes unit-efficiency measurements, zero delay between record and feedback, and Gaussian noise.

What would settle it

Numerical or experimental sampling of trajectory ensembles to compute the empirical gradient of log probability and direct comparison against r A / τ under the stated ideal conditions would confirm or refute the equality.

read the original abstract

In continuously monitored quantum systems, the feedback protocol of Garc\'ia-Pintos, Liu, and Gorshkov reshapes the arrow of time: a Hamiltonian $H_{\mathrm{meas}} = r A / \tau$ applied with gain $X$ tilts the distribution of measurement trajectories, with $X < -2$ producing statistically time-reversed outcomes. Why this specific Hamiltonian achieves reversal, and how the mechanism relates to score-based diffusion models in machine learning, has remained unexplained. We compute the functional derivative of the log path probability of the quantum trajectory distribution directly in density-matrix space. Combining Girsanov's theorem applied to the measurement record, Fr\'echet differentiation on the Banach space of trace-class operators, and K\"ahler geometry on the pure-state projective manifold, we prove that $\delta \log P_F / \delta \rho = r A / \tau = H_{\mathrm{meas}}$. The Garc\'ia-Pintos feedback Hamiltonian is the score function of the quantum trajectory distribution -- exactly the object Anderson's reverse-time diffusion theorem requires for trajectory reversal. The identification extends to multi-qubit systems with independent measurement channels, where the score is a sum of local operators. Two consequences follow. First, the feedback gain $X$ generates a continuous one-parameter family of path measures (for feedback-active Hamiltonians with $[H, A] \neq 0$), with $X = -2$ recovering the backward process in leading-order linearization -- a structure absent from classical diffusion, where reversal is binary. Second, the score identification enables machine learning (ML) score estimation methods -- denoising score matching, sliced score matching -- to replace the analytic formula when its idealizations (unit efficiency, zero delay, Gaussian noise) fail in real experiments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows the García-Pintos feedback Hamiltonian equals the score function of the quantum trajectory distribution, which lets them import diffusion-model tools, but the result stays tied to unit-efficiency and zero-delay assumptions.

read the letter

The main point is that the feedback Hamiltonian H_meas = r A / τ is exactly the functional derivative of the log path probability, δ log P_F / δρ. This makes it the score function that Anderson's reverse-time diffusion theorem needs for trajectory reversal. The derivation uses Girsanov on the measurement record, Fréchet differentiation on trace-class operators, and Kähler geometry on the state manifold, and it extends straightforwardly to independent multi-qubit channels as a sum of local terms. They also note that the gain X parametrizes a continuous family of path measures, with X = -2 recovering the backward process to leading order, which has no direct classical analog. That identification is new relative to the cited literature and gives a clean reason why this particular Hamiltonian works for reversal. The paper states its idealizations plainly and flags that real experiments will need ML score estimators once efficiency drops or delay appears. The math is presented without circularity or hidden fitting, and the multi-qubit extension is a direct consequence rather than an add-on. The soft spot is that the practical payoff still depends on relaxing those assumptions; without that step the ML replacement remains a stated possibility rather than a demonstrated one. No numerical validation or explicit examples appear in the abstract, so robustness under perturbation is not yet shown. This is for people working on continuous quantum feedback or on importing generative-model methods into open-system control. A reader who already knows García-Pintos et al. and wants the score-function link will find the central step useful. I would send it to peer review. The formal steps are grounded enough and the connection is specific enough that referees can check the derivation and ask for the next practical layer.

Referee Report

2 major / 3 minor

Summary. The manuscript claims to prove that the García-Pintos feedback Hamiltonian H_meas = r A / τ equals the score function δ log P_F / δ ρ of the quantum trajectory distribution. The derivation combines Girsanov's theorem applied to the measurement record, Fréchet differentiation on the Banach space of trace-class operators, and Kähler geometry on the pure-state manifold. This identification shows that the feedback protocol implements the reverse-time diffusion required by Anderson's theorem, extends to multi-qubit systems as a sum of local operators, and generates a continuous one-parameter family of path measures parameterized by feedback gain X, with X = -2 recovering the backward process to leading order. The result is presented under idealizations of unit-efficiency measurements, zero delay, and Gaussian noise, with the suggestion that ML score-matching methods can replace the analytic formula when these fail.

Significance. If the central identification holds, the work establishes a precise bridge between quantum feedback control and score-based diffusion models, providing a first-principles explanation for why a specific Hamiltonian achieves statistical time reversal. The derivation is parameter-free and directly invokes standard theorems (Girsanov, Fréchet, Kähler structure), which is a strength. The continuous family of path measures for X is a structural feature absent from classical diffusion and may have broader implications for quantum trajectory engineering. The explicit link to ML estimators offers a practical route to non-ideal experiments.

major comments (2)

[Main derivation (theorem on score identification)] The step equating the Fréchet derivative of log P_F with respect to ρ to the operator r A / τ (presumably in the section deriving the main result) is load-bearing; the manuscript should explicitly verify that the functional derivative lands in the space of bounded operators and commutes appropriately with the measurement channel under the stated Gaussian-noise assumption.
[Multi-qubit extension paragraph] The extension to multi-qubit systems claims the score is a sum of local operators for independent channels; this needs an explicit check that cross terms vanish when [H, A] ≠ 0, as this underpins the locality statement and the applicability of local ML estimators.

minor comments (3)

[Abstract and §1] The abstract and introduction should recall Anderson's reverse-time diffusion theorem in one sentence so that readers outside the diffusion-model community can follow the central claim without external lookup.
[Preliminaries] Notation for the path probability P_F and the functional derivative δ/δρ should be introduced with a brief reminder of the underlying measure space before the main theorem.
[Outlook section] The discussion of ML score estimators (denoising score matching, sliced score matching) would benefit from one concrete reference to a quantum-compatible implementation or a short remark on how the quantum state space is discretized for training.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their positive evaluation and the precise technical comments, which have helped us strengthen the rigor of the derivation and the multi-qubit extension. We address both points below and have revised the manuscript accordingly.

read point-by-point responses

Referee: [Main derivation (theorem on score identification)] The step equating the Fréchet derivative of log P_F with respect to ρ to the operator r A / τ (presumably in the section deriving the main result) is load-bearing; the manuscript should explicitly verify that the functional derivative lands in the space of bounded operators and commutes appropriately with the measurement channel under the stated Gaussian-noise assumption.

Authors: We agree that an explicit verification strengthens the load-bearing step. In the revised manuscript we have inserted a new paragraph immediately following the application of Girsanov’s theorem that computes the Fréchet derivative of log P_F on the Banach space of trace-class operators. Under the Gaussian-noise assumption the resulting operator is bounded (being a scalar multiple of the bounded observable A) and commutes with the measurement channel in the precise sense required by the change-of-measure formula. The added calculation uses only the stated idealizations and does not alter the theorem statement. revision: yes
Referee: [Multi-qubit extension paragraph] The extension to multi-qubit systems claims the score is a sum of local operators for independent channels; this needs an explicit check that cross terms vanish when [H, A] ≠ 0, as this underpins the locality statement and the applicability of local ML estimators.

Authors: We concur that an explicit demonstration is warranted. The revised Section IV now contains a direct expansion of the joint path probability for independent channels, showing that all cross terms in the Fréchet derivative vanish identically because the measurement records factorize. This cancellation is independent of whether [H, A] commutes and follows solely from the tensor-product structure of the multi-qubit state and the locality of each channel. The updated text therefore confirms both the sum-of-local-operators form and the validity of local score-matching estimators. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation uses external theorems to equate score to Hamiltonian

full rationale

The central claim computes δ log P_F / δ ρ directly from the quantum trajectory path measure via Girsanov's theorem, Fréchet differentiation on trace-class operators, and Kähler geometry on the projective manifold, then shows equality to the known García-Pintos feedback Hamiltonian r A / τ. This is a one-directional proof from the stochastic path probability to the operator identification, with no redefinition of the score in terms of H_meas, no fitted parameters renamed as predictions, and no load-bearing self-citations. The cited external results (Girsanov, Fréchet, Kähler) are standard mathematical tools independent of the target result. The paper explicitly flags its idealizations (unit efficiency, zero delay, Gaussian noise) as limitations to be relaxed, confirming the derivation is not self-referential.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The central claim rests on standard results from stochastic calculus and functional analysis rather than new postulates or fitted quantities.

axioms (3)

standard math Girsanov's theorem applies to the measurement record
Invoked to change measure between forward and feedback-modified processes
standard math Fréchet differentiation is valid on the Banach space of trace-class operators
Used to compute the functional derivative of log path probability
domain assumption Kähler geometry on the pure-state projective manifold
Supports the geometric structure of the density-matrix manifold

pith-pipeline@v0.9.0 · 5633 in / 1419 out tokens · 191493 ms · 2026-05-09T22:47:24.002887+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages · 2 internal anchors

[1]

L. P. García-Pintos, Y.-K. Liu, and A. V. Gorshkov, Reshaping the quantum arrow of time, Phys. Rev. X16, 011028 (2026). doi:10.1103/l18s-9vmh

work page doi:10.1103/l18s-9vmh 2026
[2]

B. D. O. Anderson, Reverse-time diffusion equation models,Stoch. Process. Appl.12, 313 (1982). doi:10.1016/0304-4149(82)90051-5

work page doi:10.1016/0304-4149(82)90051-5 1982
[3]

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, Score-based generative modeling through stochastic differential equations, inInternational Conference on Learning Representations(ICLR, 2021); arXiv:2011.13456

work page internal anchor Pith review Pith/arXiv arXiv 2021
[4]

X. Liu, J. Zhuang, W. Hou, and Y.-Z. You, Measurement-based quantum diffusion models, arXiv:2508.08799 [quant-ph] (2025)

work page internal anchor Pith review arXiv 2025
[5]

R. Nasu, G. Tanaka, and A. Tsuchiya, Quantum reversibility meets classical reverse diffusion, arXiv:2510.18512 [quant-ph] (2025)

work page arXiv 2025
[6]

Zhang, P

B. Zhang, P. Xu, X. Chen, and Q. Zhuang, Quantum diffusion models: Score reversal is not free in Gaussian dynamics, arXiv:2603.06488 [quant-ph] (2026)

work page arXiv 2026
[7]

H. M. Wiseman and G. J. Milburn,Quantum Measurement and Control(Cambridge University Press, 2009). doi:10.1017/CBO9780511813948

work page doi:10.1017/cbo9780511813948 2009
[8]

A connection between score matching and denoising autoencoders.Neural computation, 23(7):1661–1674, 2011

P. Vincent, A connection between score matching and denoising autoencoders,Neural Comput. 23, 1661 (2011). doi:10.1162/NECO_a_00142

work page doi:10.1162/neco_a_00142 2011
[9]

Cartan,Differential Calculus(Houghton Mifflin, 1971)

H. Cartan,Differential Calculus(Houghton Mifflin, 1971)

1971
[10]

Amari and H

S. Amari and H. Nagaoka,Methods of Information Geometry(AMS/Oxford University Press, 2000). doi:10.1090/mmono/191

work page doi:10.1090/mmono/191 2000
[11]

V. P. Belavkin, Quantum stochastic calculus and quantum nonlinear filtering,J. Multivar. Anal.42, 171 (1992). doi:10.1016/0047-259X(92)90042-E

work page doi:10.1016/0047-259x(92)90042-e 1992
[12]

Fagnola and V

F. Fagnola and V. Umanità, Generators of detailed balance quantum Markov semigroups,Infin. Dimens. Anal. Quantum Probab. Relat. Top.10, 335 (2010). doi:10.1142/S0219025710004152

work page doi:10.1142/s0219025710004152 2010
[13]

Y. Song, A. Garg, J. Shi, and S. Ermon, Sliced score matching: A scalable approach to density and score estimation, inUncertainty in Artificial Intelligence(UAI, 2020); arXiv:1905.07088. 14

work page arXiv 2020