REVIEW
Learning Local Fermionic Lindbladians under Parity Superselection
Quantum Physics · Submitted 30 Sep 2026 · 81 references
Pith's verdict: Genuinely new protocol for learning interacting fermionic Lindbladians under parity superselection; the core algebra checks out, but marker packing gates the headline no-external-modes claim and the finite-lattice regime boundary is never mapped.
T0 · AI review
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
Calibrating three internal marker Majoranas lets local parity-covariant fermionic open-system generators be learned from even measurements alone, with no external modes, at sample counts matching the short-time lower bound.
The reading
What carries the argument
The load-bearing objects are the parity-dressing map $J_\eta(\Gamma_A)=i^{\pi(A)}\Gamma_A\eta^{\pi(A)}$, which turns an odd Majorana monomial into an even Hermitian unitary by adjoining one marker Majorana $\eta$; the internal-marker calibration triplet, three single-mode marker cores pairwise separated by more than the interaction range $r$, whose pair derivatives $s_{ij}=b_i+b_j$ solve for the marker self-terms and make the correction identity of Theorem 5.9 exact; and signed Fierz inversion, $\chi_{A,B}=d^{-3}\sum_{C,D}(-1)^{\pi(D)\pi(B)}\Lambda_{C,D}\,\mathrm{tr}[\Gamma_C\Gamma_B\Gamma_D\Gamma_A]$, combined with the restricted-data diagonal-extension identity that expresses a restricted dressed array as the global coefficient plus extensions of strictly larger support, giving a descending recursion from large to small supports. Around these sit pair-block process shadows, which extract even transfer curves from common data with bounded second moments, the parallel even-block marginalization lemma that lets one full-system sample serve many disjoint regions, and the semidefinite fit that promotes noisy coefficients to a valid GKSL generator.
Extended reading notes
Core claim
Time-independent, parity-covariant, $k$-mode-local Lindbladians on $m$ fermionic modes can be reconstructed from even preparations and measurements with uninterrupted short-time evolution, without external modes when the mode graph is known and satisfies the marker-packing assumption. Three pairwise-separated single-mode marker cores provide the calibration: the pair derivatives satisfy $s_{ij}=b_i+b_j$ (Proposition 5.8), fixing the marker self-terms $b_i$, and the internal-marker correction $\Lambda^{R|i}_{C,D}=\Lambda_{C,D}+\mathbf{1}\{\pi(C)=\pi(D)=1\}\delta_{C,D}b_i$ (Theorem 5.9) turns measured derivatives into the ideal external-marker data that the restricted signed Fierz inversion of Theorem 4.5 consumes. The descending recursion over support sizes then estimates every induced Hamiltonian coefficient $h_A$ and Kossakowski entry $G_{A,B}$ to entrywise error $\varepsilon$ with $\widetilde{\mathcal{O}}_{k,r,\Delta}(\bar\alpha^2\varepsilon^{-2}\log(m/\delta))$ samples (Theorem 6.5), while a semidefinite fit returns a valid generator with diamond-norm error $\varepsilon$ at an extra factor $m^2$ that Corollary 6.7's lower bound matches; without a supplied graph, one idle ancillary mode per system mode plus pair operations recovers all-to-all structure under an approximate coefficient-degree bound $d$ with local error $\varepsilon$ using $\widetilde{\mathcal{O}}_k(\bar\alpha^2 d^2 m^{\lfloor k/2\rfloor}\varepsilon^{-2}\log(m/\delta))$ samples (Theorem 7.2).
Load-bearing premise
The geometric no-external-modes result (Theorem 6.5) rests on Assumption 6.1: every $k$-local region of diameter at most $r$ must have three single-mode marker cores inside a ball of radius $4r+3$, pairwise separated by more than the interaction range, because only under that termwise separation do the calibration identities of Proposition 5.8 and Theorem 5.9 hold exactly, and on small, dense, or boundary-heavy graphs this packing fails, leaving the ancillary-mode protocol of Section 7 as the fallback.
Editorial extensions
If this is right
- Dissipative Fermi–Hubbard models with hopping, interaction, loss, injection, dephasing, pair loss, and incoherent hopping fall under Theorem 6.5, so their canonical coefficients are learnable from even data at logarithmic sample counts (Example 6.6).
- Fixed-support even observables — occupations, double occupations, currents, spin observables, and even pairing correlations — can be predicted simultaneously with logarithmic system-size sample complexity at fixed time and accuracy (Corollary 8.2).
- For a prescribed region, only a surrounding finite volume needs reconstruction: the sample count, local readouts, and classical inference cost become independent of the total system size, although every sample still evolves the full system (Corollary 8.3).
- Exponentially decaying interactions on periodic lattices add only polylogarithmic overhead to entrywise and diamond-norm learning (Corollary 6.9).
- Without a known graph, the approximate coefficient-degree condition restores structure learning with one idle ancilla per mode and an extra $m^{\lfloor k/2\rfloor}$ factor, and thresholding recovers the exact coefficient support when a magnitude gap separates zero from nonzero entries (Theorem 7.2).
Reading between the lines
- Where the marker separation is only approximate, one would expect residual cross-terms of order $\bar\alpha$ to appear in Theorem 5.9's identity and the entrywise guarantee of Theorem 6.5 to degrade continuously toward the ancillary-mode protocol; this transition could be probed numerically on small lattices that violate Assumption 6.1.
- Because the calibration removes the markers' own contributions exactly, the same triplet scheme may allow fermionic Hamiltonian- and Gaussian-learning protocols to drop external reference modes on geometric graphs, potentially extending Heisenberg-limited fermionic learning to leaner experimental access.
- The even-curve framework and robust derivative recovery of Proposition F.3 are time-local in structure, so an extension to time-dependent parity-covariant fermionic noise appears plausible; the authors do not claim such an extension.
- The $m^{\lfloor k/2\rfloor}$ pairing-coverage factor in Theorem 7.2 is a cost of the present ancillary-pairing architecture rather than a lower bound for collective even measurements, so collective access could plausibly push the graph-free sample scaling toward the geometric bound.
Original abstract
Parity superselection forbids direct measurement of odd Majorana observables. We learn time-independent, parity-covariant, $k$-mode-local Lindbladians on $m$ modes using even preparations and measurements with uninterrupted short-time evolution. Internal markers turn odd probes into even observables, and pair measurements among three separated markers calibrate their dynamical contributions. Signed Fierz inversion recovers canonical coefficients, while semidefinite fitting yields a valid generator. For finite-range models on known bounded-degree graphs with suitable marker access and a supplied weighted-strength bound $\bar\alpha$, entrywise error $\varepsilon$ is achieved using $\widetilde{\mathcal{O}}(\bar\alpha^2\varepsilon^{-2}\log(m/\delta))$ samples, without external modes or known nonzero coefficient locations. Recovery to diamond-norm error $\varepsilon$ costs an additional factor $m^2$, matching lower bounds for short-time experiments on fresh systems up to logarithmic and fixed geometric factors. Without a supplied graph, one idle ancillary mode per system mode and potentially nonlocal pair operations give total absolute coefficient error at most $\varepsilon$ per mode using $\widetilde{\mathcal{O}}_k(\bar\alpha^2\mathsf d^2m^{\lfloor k/2\rfloor}\varepsilon^{-2}\log(m/\delta))$ samples, under a supplied approximate coefficient-degree bound $\mathsf d$ and controlled weak-coefficient tails. All guarantees hold with probability at least $1-\delta$. For geometric models, fixed-support even observables can be predicted with logarithmic system-size sample complexity at fixed time and accuracy. We also give finite-volume and exponential-tail extensions.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
Parity superselection forbids direct measurement of odd Majorana observables. We learn time-independent, parity-covariant, $k$-mode-local Lindbladians on $m$ modes using even preparations and measurements with uninterrupted short-time evolution. Internal markers turn odd probes into even observables, and pair measurements among three separated markers calibrate their dynamical contributions. Signed Fierz inversion recovers canonical coefficients, while semidefinite fitting yields a valid generator. For finite-range models on known bounded-degree graphs with suitable marker access and a supplied weighted-strength bound $\bar\alpha$, entrywise error $\varepsilon$ is achieved using $\widetilde{\mathcal{O}}(\bar\alpha^2\varepsilon^{-2}\log(m/\delta))$ samples, without external modes or known nonzero coefficient locations. Recovery to diamond-norm error $\varepsilon$ costs an additional factor $m^2$, matching lower bounds for short-time experiments on fresh systems up to logarithmic and fixed geometric factors. Without a supplied graph, one idle ancillary mode per system mode and potentially nonlocal pair operations give total absolute coefficient error at most $\varepsilon$ per mode using $\widetilde{\mathcal{O}}_k(\bar\alpha^2\mathsf d^2m^{\lfloor k/2\rfloor}\varepsilon^{-2}\log(m/\delta))$ samples, under a supplied approximate coefficient-degree bound $\mathsf d$ and controlled weak-coefficient tails. All guarantees hold with probability at least $1-\delta$. For geometric models, fixed-support even observables can be predicted with logarithmic system-size sample complexity at fixed time and accuracy. We also give finite-volume and exponential-tail extensions.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | Parity superselection forbids direct measurement of odd Majorana observables. | background 1.00 | 0.05 | 0.03 | asserted 1.00 |
| s1 | We learn time-independent, parity-covariant, $k$-mode-local Lindbladians on $m$ modes using even preparations and measurements with uninterrupted short-time evolution. | contribution 0.94 | 0.95 | 0.34 | asserted 1.00 |
| s2 | Internal markers turn odd probes into even observables, and pair measurements among three separated markers calibrate their dynamical contributions. | method 1.00 | 0.86 | 0.07 | asserted 1.00 |
| s3 | Signed Fierz inversion recovers canonical coefficients, while semidefinite fitting yields a valid generator. | method 0.99 | 0.86 | 0.08 | asserted 1.00 |
| s4 | For finite-range models on known bounded-degree graphs with suitable marker access and a supplied weighted-strength bound $\bar\alpha$, entrywise error $\varepsilon$ is achieved using $\widetilde{\mathcal{O}}(\bar\alpha^2\varepsilon^{-2}\log(m/\delta))$ samples, without external modes or known nonzero coefficient locations. | result 0.96 | 0.91 | 0.32 | asserted 1.00 |
| s5 | Recovery to diamond-norm error $\varepsilon$ costs an additional factor $m^2$, matching lower bounds for short-time experiments on fresh systems up to logarithmic and fixed geometric factors. | result 0.96 | 0.84 | 0.15 | asserted 1.00 |
| s6 | Without a supplied graph, one idle ancillary mode per system mode and potentially nonlocal pair operations give total absolute coefficient error at most $\varepsilon$ per mode using $\widetilde{\mathcal{O}}_k(\bar\alpha^2\mathsf d^2m^{\lfloor k/2\rfloor}\varepsilon^{-2}\log(m/\delta))$ samples, under a supplied approximate coefficient-degree bound $\mathsf d$ and controlled weak-coefficient tails. | result 0.92 | 0.91 | 0.12 | asserted 0.98 |
| s7 | All guarantees hold with probability at least $1-\delta$. | result 0.91 | 0.79 | 0.05 | asserted 1.00 |
| s8 | For geometric models, fixed-support even observables can be predicted with logarithmic system-size sample complexity at fixed time and accuracy. | result 0.96 | 0.86 | 0.12 | asserted 0.98 |
| s9 | We also give finite-volume and exponential-tail extensions. | contribution 0.85 | 0.89 | 0.04 | asserted 0.97 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Figures reproduced from arXiv:2610.00448 by the authors.
References
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.