{"id":"2ea546b8-a35f-42d9-b507-ae4b198f67a2","arxiv_id":"2607.27534","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"t-doped Gaussian and Gaussian-entanglable unitaries are learnable in poly(m) forward-only queries using coherent probes and heterodyne detection, while general m-mode unitaries require Ω(E^{2m}) queries.","lead":"This theory paper proves that two structured families of non-Gaussian quantum light devices can be reverse-engineered with only polynomially many experiments, while a fully general device still requires exponentially many. These are the first efficient forward-only learning protocols for multimode unitaries whose non-Gaussian parts and entanglement can both be extensive.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main poly(m) bound rests on an unproved perturbative error-propagation assumption: Step errors are analyzed assuming earlier steps exact, and the triangle-inequality composition in Theorem S27 is asserted rather than derived. This is a proof gap, not a contradiction.","rationale":"The reader's weakest_assumption precisely identifies the same load-bearing issue: the poly(m) query bound is conditional on an explicit perturbative error-propagation approximation and on energy-finitude conditions that are assumed rather than derived. My reading of Supplemental Note II A and Theorem S27 confirms that the step-error bounds are asserted under the assumption that earlier steps are exact, while the actual protocol uses approximate counter-rotations and approximate passive frames. The triangle-inequality composition in (S459)–(S469) is exact as a norm inequality, but the individual terms are not proven for the actual estimators. This is the most load-bearing concern because the main positive results, Theorems 1 and 2, inherit their polynomial scaling directly from this composition. The concern is addressable: a non-perturbative error analysis, or a numerical check on a small explicit instance, would settle whether error amplification is real. The structural theorems (Darmois–Skitovich, activation, uncalibrated-probe learning) and the general lower bound appear to rest on more self-contained arguments and are not where I would place the risk. The reader's CONDITIONAL verdict is therefore appropriate; I do not see a reason to move to ACCEPT or REJECT on the current manuscript.","tokens_in":69612,"tokens_out":7834,"duration_ms":94393,"concrete_test":"Perform a two-stage error-propagation test on a fixed 2-mode Gaussian-entanglable unitary U_ge = U_S (W_1 ⊗ I) U_O, with U_S a known two-mode squeezer, U_O a 50:50 beamsplitter, and W_1 a Kerr or cubic-phase gate. Inject a controlled error δS into the Step-1 symplectic estimate, then run Steps 2 and 3 on the resulting data in a finite-cutoff simulation, and compute the actual ∥U_ge − U~_ge∥. Vary δS at fixed target ε. If the achieved error is not bounded by the predicted sum ε_gd + ε_ps + ε_ng + ε_G and does not decrease at the predicted rate, then the perturbative composition in Theorem S27 is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Supplemental Note II A states an explicit 'perturbative error-propagation assumption': when analyzing reconstruction error at a given step, all preceding steps are assumed exact, and total error is taken to be the sum of independent step errors. Theorem S27 then proves the poly(m) query bound by the telescoping triangle inequality (S459–S469). The triangle inequality itself is exact, but the four individual bounds are not established for the actual estimators. For example, the third term (S466–S467) bounds ∥W'_j − W~'_j∥ in the energy-constrained diamond norm using the uniform intermediate bound E** defined from the exact U_final in (S452). The local estimate W~'_j, however, is produced from data that in the actual protocol have passed through approximate counter-rotations U~_S and U~_O2; no argument shows that the single-mode reconstruction theorem S23 / Proposition 10 applies to those data with the same E**. If Step-1 or Step-2 estimation errors change the effective input energy to the local channels, or if local reconstruction error depends on them, the sum in (S469) need not close. Thus the central claim M = poly(m) is conditional on an unproved approximation, in addition to the assumed energy finiteness N_dyn, E_II, E**. This is a genuine gap between the stated theorem and the supplied proof, not evidence the result is false.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims a tractability frontier for bosonic unitary learning. It proves a lower bound Ω(E^{2m}) for general m-mode unitaries under an energy constraint, and provides two positive results: t-doped Gaussian unitaries and Gaussian-entanglable unitaries can be learned with poly(m) queries using forward-only coherent-state probes, heterodyne detection, Gaussian operations, and classical post-processing. The proofs rely on a multimode quantum Darmois–Skitovich theorem, an almost-sure activation theorem for non-Gaussian unitaries, and a method for learning from uncalibrated coherent inputs. The main theorems are supported by extensive supplemental derivations and explicit sample-complexity expressions, but they are conditional on several additional energy-finiteness conditions and on an explicitly stated perturbative error-propagation assumption.","tokens_in":69981,"tokens_out":6381,"duration_ms":74421,"significance":"If fully established, the results are significant: they extend efficient bosonic process learning beyond the Gaussian class and identify the irreducible mixing of non-Gaussianity with multimode entanglement as the key complexity obstruction. The structural theorems—particularly the multimode Darmois–Skitovich theorem and the almost-sure activation theorem—are likely to be of independent interest beyond the learning application. The paper is also commendable for providing explicit protocols, explicit query-complexity bounds, and a detailed supplement. However, the central polynomial-query claims are not yet fully proven because the error-composition argument rests on an unproved perturbative approximation, and several energy conditions are assumed rather than derived from the stated physical constraints.","major_comments":[{"comment":"The proof of the polynomial query bound uses an explicit 'perturbative error-propagation assumption' (Supplemental Note II A), under which each step's error is computed assuming previous steps are exact and total error is the sum. In Theorem S27, Eq. (S459) applies the triangle inequality, but the individual bounds (S460)–(S468) are only derived for exact intermediate channels. In particular, the third term (S466)–(S467) uses the uniform intermediate energy bound E** defined from the exact U_final in Eq. (S452), whereas the actual data for the local reconstruction W~'_j are obtained after approximate counter-rotations U~_S and U~_O2. No argument shows that Theorem S23 applies to those data with the same E**, nor that the local reconstruction error is Lipschitz in the preceding errors. Thus the composition (S469) is asserted, not derived. This is load-bearing for Theorems 1 and 2.","section":"Supplemental Note II A; Theorem S27, Eqs. (S459)–(S469)"},{"comment":"The polynomial bounds are conditional on energy-finiteness conditions that are not implied by Eq. (1): N_dyn, E_II, and E**. In Theorem 2, E** is defined via Eq. (S452) from the exact U_final, which is unknown during the protocol; no verifiable condition is given. Theorem 1 similarly requires Proposition 10's N_dyn < ∞, a supremum over an unbounded set of states that is not guaranteed by bounded probe/output energies. The class of unitaries for which the protocols run in poly(m) is therefore not characterized by the stated physical assumptions, and no check is offered to certify these conditions.","section":"Theorems 1, 2 and Eq. (1)"},{"comment":"The benchmark is presented as a tractability frontier, but the upper bound (24) relies on the unproved assumption 'Assume that the cutoff-L coherent-probe reconstruction is stable,' and the effective cutoff dimension d_L with L = max{K, 256 E_U(K)/ε²} can be much larger than the d_E used in the lower bound. Hence the upper bound is not a matching converse, and the frontier between the general lower bound and the two polynomial families is not quantitatively closed. The lower bound itself appears sound, but the 'frontier' claim should be calibrated.","section":"Theorem 11, Eqs. (23)–(24)"},{"comment":"The single-mode uncalibrated learning theorem assumes N_dyn < ∞ for W_gf and N_out < ∞. These are used to set the cutoff R and to control the polar completion (Lemma S22). For many non-Gaussian unitaries the dynamical bound is finite, but for unitaries with unbounded energy growth N_dyn can be infinite. The theorems restrict to the finite case, which is an extra structural promise not visible in the main-text statement 'an arbitrary single-mode unitary.' The abstract and Section II overstate the reach of Theorem 9 if the finite-energy condition is not highlighted there.","section":"Theorem S23 / Proposition 10"}],"minor_comments":[{"comment":"Typographical errors: 'Gaussian untiaries' (Section I.B), 'constract' (Section I), 'Gaussinity' (Supplemental Remark S16), 'exppoly' (Fig. 1 caption).","section":"Throughout"},{"comment":"The parameter C in Eq. (24) is not specified; an order bound with an unknown constant C≥5 makes the comparison with the lower bound loose. Please state the best available explicit value or explain why it is protocol-dependent.","section":"Theorem 11"},{"comment":"The quantity E** is defined in the supplement after Eq. (S452) but is cited in the main-text Theorem 2. Please define it in the main text or give a precise reference to the supplemental equation.","section":"Theorem S27 / Eq. (S452)"}],"recommendation":"major_revision","confidential_remarks":"This is a technically rich manuscript with several intriguing results and a large amount of useful supplementary material. The main concern is that the central polynomial-query bound is not fully proven: the perturbative error-propagation assumption is explicitly stated but not justified, and the energy conditions E** and N_dyn are assumed rather than derived. These issues are likely fixable within the manuscript's scope—for example, by proving Lipschitz/robustness properties of the reconstruction steps or by explicitly adding the perturbative assumption and the energy conditions to the theorem statements—but as it stands the main theorems are conditional on unproved approximations. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe headline is that this is a serious, ambitious paper that will move the CV learning literature, but the main polynomial bounds are conditional on a stated first-order error-propagation assumption that is not proved. The reader's conditional verdict is on the mark, and the stress-test about Theorem S27 is the soft spot to worry about.\n\nWhat's genuinely new: the two learnable families (t-doped Gaussian and Gaussian-entanglable unitaries), the multimode quantum Darmois–Skitovich theorem, the almost-sure activation theorem, and the uncalibrated-coherent-probe learner. These are real theorems with real proofs in the supplement, and the paper is transparent about the assumptions it needs. The Ω(E^{2m}) lower bound for general unitaries is a clean embedding argument and looks right. The authors build on their own state-learning results—that's not circularity, and they cite the external machinery properly.\n\nWhere it gets soft: the query-complexity theorems (Thm 1 and 2) are assembled by summing per-step errors under an explicit perturbative approximation. That's fine as a heuristic, but as a proof it doesn't close. The triangle inequality step is exact; each of the four terms, especially the local non-Gaussian reconstruction term, is not actually established for the real data, which have passed through approximate counter-rotations. The bound E** is defined from the exact U_final, but the estimators see a slightly different channel. If step errors change the effective input energy to the local channels, the sum need not close. The paper says \"we adopt a perturbative error-propagation assumption\"—good honesty, but then Theorem S27 as stated goes beyond what the proof supports. This is a proof gap, not a contradiction, and I'd be surprised if it can't be filled with a more careful recursive analysis, but it is the thing a referee must press on.\n\nAlso, the energy finiteness conditions (N_dyn, E_II, E**) are assumed rather than derived from Eq. (1). They are stated transparently, so I don't hold that against the authors, but it means the theorems are conditional, not unconditional.\n\nBottom line: this paper deserves a serious referee. The structural results alone are worth publishing, and the learning results are likely right in spirit. I'd send it to peer review with a request to focus on the error-composition step and to either prove a recursive bound or state the theorems with the first-order assumption made explicit as a hypothesis. For my own reading, I'd bring it to the group and cite it, with a caveat on the main theorem.","headline":"A serious, novel paper whose main polynomial learning claims are conditional on a stated but unproved perturbative error-propagation assumption; deserves peer review, with the error-budget step as the focus.","tokens_in":70458,"tokens_out":2990,"would_cite":true,"duration_ms":30693,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P45","81V80","62E10"],"pacs":["03.67.-a","42.50.Ar"],"model":"deepseek-v4-flash","headline":"The paper tries to establish that the real obstacle to learning multimode bosonic unitaries is not non-Gaussianity by itself, but the irreducible mixing of non-Gaussian features with multimode entanglement. It claims that two broad non-Gaus","keywords":["bosonic unitary learning","continuous-variable quantum information","non-Gaussianity","Gaussian-entanglable unitaries","t-doped Gaussian unitaries","Darmois–Skitovich theorem","coherent-state process tomography","query complexity"],"falsifier":"Compute, for a concrete non-Gaussian single-mode gate W embedded in a Gaussian-entanglable unitary with growing m, the exact output energy after the passive input layer under a fixed per-mode input energy; if the finite uniform intermediate evaluation-energy bound E** diverges with m, the polynomial query bound of Theorem 2 does not apply. Alternatively, construct a two-step instance where the perturbative error-propagation assumption underestimates the accumulated error by more than a constant factor—for example by numerically simulating the protocol with the estimated Gaussian counter-rotati","tokens_in":69426,"feed_emoji":"⚛️","tokens_out":4404,"duration_ms":48525,"temperature":0.7,"pith_summary":"This paper establishes a tractability frontier for learning unknown multimode bosonic unitaries from input–output experiments, using only coherent-state probes, local heterodyne detection, Gaussian unitaries, and classical post-processing. Its central claim is that non-Gaussianity per se is not the obstruction to efficient learning; the obstruction is the irreducible mixing of non-Gaussianity with multimode entanglement. The authors prove that two families—t-doped Gaussian unitaries (Gaussian layers interleaved with a constant number of local non-Gaussian gates) and Gaussian-entanglable unitaries (arbitrary local single-mode unitaries sandwiched between passive and general Gaussian layers)—can be learned with poly(m) queries, even though the second family can have extensive non-Gaussianity and strong entanglement. In contrast, they prove a general m-mode unitary with per-mode input energy E requires at least Ω(E^{2m}) channel uses. If correct, efficient characterization and verification of structured photonic processors with local nonlinear operations becomes feasible.","feed_headline":"Non-Gaussianity alone doesn't block learning bosonic circuits","feed_subtitle":"t-doped and Gaussian-entanglable unitaries are learnable with poly(m) queries; general ones need E^{2m}.","key_machinery":"The protocol rests on three structural tools. First, a multimode quantum Darmois–Skitovich theorem: under nontrivial passive (linear-optical) mixing, a product input can remain product only if all participating input states are Gaussian; this lets the learner identify the passive layers up to a limited gauge (permutations, local phase rotations, and mixing within covariance-degenerate Gaussian blocks). Second, an almost-sure activation theorem: for any non-Gaussian unitary, a random coherent probe produces a non-Gaussian output with probability one, so the exceptional coherent inputs that hide non-Gaussianity form a measure-zero set. Third, a self-calibrating single-mode unitary learner that","core_discovery":"The paper's main results are two theorems. Theorem 1 shows that a t-doped Gaussian unitary U_doped = U_S (U_X ⊗ D) U_O, where the non-Gaussian block U_X acts on at most κt = O(1) modes, can be learned to error ε in energy-constrained diamond norm using M = poly(m) forward queries with coherent states and local heterodyne detection. Theorem 2 shows the same for Gaussian-entanglable unitaries U_ge = U_S (⊗_j W_j) U_O, where each W_j is an arbitrary single-mode unitary and the non-Gaussian layer can act on all m modes; the query complexity is poly(m, E*, 1/ε, log(1/δ)). Both protocols are forward-only, meaning they never require access to U† or U^T. The lower bound of Ω(E^{2m}) for general unit","pith_inferences":["Editorial inference: If the perturbative error-propagation assumption—that each step's error is computed assuming earlier steps are exact and errors add linearly—can be replaced by rigorous cross-step error bounds, the polynomial scaling would hold under weaker hypotheses; currently the triangle-inequality composition of errors in the proof relies on this assumption.","Editorial inference: The Darmois–Skitovich theorem is exact; a finite-error version of it, as the paper itself notes, would extend the learning protocol to lossy or noisy settings where coherent outputs are only approximately product.","Editorial inference: The lower bound Ω(E^{2m}) suggests that any efficient learner for a wider class of unitaries must exploit a promise that limits the number of modes on which non-Gaussian gates act jointly; families with blockwise non-Gaussian couplings of growing width are natural candidates for further study.","Editorial inference: The uncalibrated-probe learner implies that calibration and process tomography can be merged; in practice this could reduce alignment requirements in photonic experiments, but the current protocol's complexity involves high powers of 1/ε, so numerical optimization of the constants would be a practical next step."],"forward_implications":["Structured photonic processors that interleave Gaussian operations with local nonlinear gates can be characterized with polynomially many forward queries, even when the number of modes is large.","The exponential lower bound for general unitaries means the learnability boundary is set by the mixing of non-Gaussianity and entanglement, not by either resource alone.","The quantum Darmois–Skitovich theorem gives a mode-resolved test for whether a passive linear-optical network genuinely mixes modes, which could be used as a diagnostic for hidden mixing.","The activation theorem resolves probe selection: a random coherent probe is sufficient to reveal non-Gaussian dynamics, so adversarial hiding inputs are negligible.","Learning from uncalibrated coherent probes removes the need for separate source calibration in coherent-state process tomography, up to an unavoidable phase-rotation gauge."],"fun_headline_variants":["Polynomial learning for two non-Gaussian unitary classes","Bosonic learning: mixing non-Gaussianity and entanglement is the true cost","General bosonic unitaries need exponential queries; two classes don't","Non-Gaussianity alone isn't the obstacle; mixing with entanglement is"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The polynomial query bounds assume that several energy quantities remain finite—the dynamical energy bound N_dyn, the output second-moment bound E_II, the per-probe output bound N_out, and the paper-specific finite uniform intermediate evaluation-energy bound E**—and that estimation errors from different steps add without cross-step amplification; if an intermediate energy blows up or errors amplify, the triangle-inequality composition no longer closes and the poly(m) claims","fun_headline_variants_meta":{"raw":{"variants":["Polynomial learning for two non-Gaussian unitary classes","Bosonic learning: mixing non-Gaussianity and entanglement is the true cost","General bosonic unitaries need exponential queries; two classes don't","Non-Gaussianity alone isn't the obstacle; mixing with entanglement is"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001173,"raw_usage":{"total_tokens":4735,"prompt_tokens":838,"completion_tokens":3897,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":3822}},"tokens_in":582,"tokens_out":3897,"duration_ms":26111,"temperature":1.0,"reasoning_tokens":3822,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:10:11.540713+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, for a concrete non-Gaussian single-mode gate W embedded in a Gaussian-entanglable unitary with growing m, the exact output energy after the passive input layer under a fixed per-mode input energy; if the finite uniform intermediate evaluation-energy bound E** diverges with m, the polynomial query bound of Theorem 2 does not apply. Alternatively, construct a two-step instance where the perturbative error-propagation assumption underestimates the accumulated error by more than a constant factor—for example by numerically simulating the protocol with the estimated Gaussian counter-rotati","supporting_citations":[],"review_version":1}