{"id":"ec4b6edd-9590-4916-bc5e-c03f50542e82","arxiv_id":"2505.00102","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A coherent averaging protocol over randomly imperfect interferometers is proven to converge to the ideal boson sampling distribution, supported by a new trace distance bound for vacuum-heralded linear networks.","lead":"Researchers show that averaging multiple noisy copies of a boson sampler can recover the ideal sampling distribution when the fabrication errors are random and unbiased. The result is a concrete error-mitigation strategy for sampling experiments, but it pays for this with an exponentially small success probability.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence to the ideal boson sampling distribution depends on zero-mean, uncorrelated noise; a systematic fabrication bias would make the averaged unitary converge to the wrong matrix, so Eq. (4.12) would fail.","rationale":"I read the paper as claiming: under stochastic unitary fabrication errors, the unitary-averaging protocol produces a boson-sampling distribution converging to the ideal one as N→∞, with Theorem 1 giving a quantitative TVD bound. The proof of Theorem 1 is credible: the homogeneity ϕ(cM)=c^nϕ(M), the tensor-power Lipschitz bound (5.10), and the TVD≤state-vector-distance step (5.15) are all sound. The actual bottleneck is not the representation theory but Eq. (3.9). The central limit makes U_avg converge to E(U_i), and the protocol can only correct the effect of a scalar prefactor. If E(U_i) is not a scalar multiple of U, the limit is a different boson-sampling distribution. Real fabrication errors can have a systematic component, as the paper's own benchmarking discussion (Sec. VII.C) acknowledges. Thus the convergence claim is conditional on zero-mean i.i.d. noise and equal path lengths; without an explicit caveat the abstract overstates the scope. I also noted that Eq. (4.11) uses p_uni in the denominator but p_post in the definition of k; this is a minor inconsistency that does not affect the N→∞ limit but should be corrected. Because the mathematical derivation is internally consistent under the stated model, and the scope limitation is already reflected in a CONDITIONAL verdict, no verdict change is needed. The proposed biased-noise simulation would settle whether the concern actually lands.","tokens_in":16424,"tokens_out":23139,"duration_ms":243915,"concrete_test":"Run a Monte Carlo simulation of the UA protocol for a 4-mode, n=2 boson sampler, as in Fig. 4, but with the noise in Eq. (3.2) modified to δθ_{i,j} ~ N(μ,ν) and δφ_{i,j} ~ N(η,ν) for μ=η=0.02, ν=0.01, keeping all other Sec. III assumptions. Compute the average TVD between D_U and the heralded UA distribution for N=10, 30, 100, 300, 1000. If the TVD saturates at the positive value ‖D_U - D_Ũ‖ instead of decaying to zero, the zero-mean assumption behind Eq. (3.9) is load-bearing and the abstract's unqualified convergence claim must be restricted to unbiased stochastic errors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central convergence claim (Eq. 4.12) depends entirely on Eq. (3.9), E(U_i) = (1-ν/2)^d U, which is derived in Sec. III from the zero-mean, independent Gaussian noise model (3.4) plus the uniform-depth equal-path-length assumption. If fabrication noise has any systematic component, e.g., E(δθ_{i,j}) = μ ≠ 0 or E(δφ_{i,j}) = η ≠ 0, then E(U_i) = Ũ is not proportional to U. By the law of large numbers U_avg → Ũ, and the normalization in Eq. (4.11) cancels only scalar factors; it cannot map Ũ back to U. Consequently the TVD in Eq. (4.12) tends to ‖D_U - D_Ũ‖ > 0, not 0. The paper provides no calibration or bias-estimation step, and Sec. VII.C itself lists 'systematic deviations, process drift' among fabrication variability sources, so the zero-mean assumption is not a harmless technicality. This is the load-bearing link between the physical noise model and the claimed convergence; no amount of averaging removes a bias.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unitary averaging protocol for mitigating stochastic unitary errors in boson sampling. The main theoretical result is Theorem 1, an upper bound on the total variation distance between the output distributions of two invertible vacuum-heralded linear optical transformations, derived using Schur-Weyl duality. The authors apply this bound to the averaged unitary U_avg = (1/N)Σ U_i and claim that in the N→∞ limit the output distribution converges to the ideal boson sampling distribution, with a success probability lower-bounded by an expression exponential in the photon number and interferometer depth. The paper also discusses applications to linear combinations of unitaries and benchmarking fabrication repeatability.","tokens_in":16622,"tokens_out":19136,"duration_ms":195071,"significance":"If the convergence claim can be made rigorous, the work is significant: it extends quantum error mitigation from expectation-value estimation to sampling tasks, provides a quantitative distance bound for vacuum-heralded linear optical networks, and offers connections to LCU and photonic quantum computing. The proof of Theorem 1 appears largely sound, and the numerical demonstrations support the qualitative behavior. However, the central convergence statement currently relies on treating a second-order Taylor approximation as an exact proportionality, and there are inconsistencies in the application of Theorem 1. These issues are fixable, and the paper has the potential to be a useful contribution after revision.","major_comments":[{"comment":"The convergence claim in Eq. (4.12) is not justified as written. Equation (3.7) is a second-order Taylor approximation yielding E(y) ≈ (1−ν/2)y0, and Eq. (3.9) inherits this approximation. In Section IV.B this approximate proportionality is treated as exact: Eq. (4.8) uses (1−ν/2)^{dn} and p_uni = (1−ν/2)^{2dn}, and Eq. (4.12) concludes the TVD is ≤ 0. For Gaussian noise the exact expectation is E[cos(θ+δ)] = cosθ e^{−ν/2}, with analogous expressions for sin and e^{iφ}, so the exact proportionality is E(U_i) = e^{−νd/2} U and the associated success probability is e^{−νdn}. With the stated p_uni, the normalized averaged unitary converges to (e^{−νd/2}/(1−ν/2)^d) U, leaving an O(ν²) residual, so the limit in Eq. (4.12) is not zero. Please either use exact Gaussian moments throughout (with p_uni = e^{−νdn}) or explicitly state that the convergence is up to O(ν²) and adjust Eq. (4.12) accordingly.","section":"Section III, Eqs. (3.7)-(3.9); Section IV.B, Eq. (4.12)"},{"comment":"The constant k in Eq. (4.11) is defined as max(∥U∥, ∥U_avg p_post^{−1/2n}∥), but Theorem 1 requires k to be the maximum of the norms of the two normalized matrices appearing in the difference, namely max(∥U∥, ∥U_avg p_uni^{−1/2n}∥) when B is normalized by p_uni. Since p_post ≥ p_uni, we have p_post^{−1/2n} ≤ p_uni^{−1/2n}, so the defined k can be smaller than the required constant and the stated bound may fail. The normalizations p_uni and p_post should be used consistently throughout the application of the theorem.","section":"Section IV.B, Eq. (4.11)"},{"comment":"The mitigation result depends critically on the zero-mean assumption E(δθ)=E(δφ)=0 in Eq. (3.4). If fabrication noise contains a systematic component, Eq. (3.9) fails and the averaged unitary converges to a biased matrix that the scalar normalization p_uni cannot correct; the TVD in Eq. (4.12) would then tend to a positive value, not zero. Section VII.C itself lists systematic deviations and process drift as sources of fabrication variability, so the paper should explicitly state this limitation and discuss whether a calibration or bias-estimation step is needed before the protocol can be applied to real devices.","section":"Section III, Eq. (3.4); Section VII.C"},{"comment":"Theorem 1 requires A,B ∈ GL(m,C), and the paper assumes U_avg is invertible for finite N without proof. This is not automatic for an average of unitaries. If U_avg is singular, the theorem does not apply and the convergence conclusion is not established. The authors should either condition the statement on invertibility of U_avg or argue that the singular set has measure zero under the continuous Gaussian noise model, so that the assumption holds with probability one.","section":"Section IV.B and Theorem 1"}],"minor_comments":[{"comment":"Since a norm is nonnegative, the inequality ∥D_U − D_Uavg∥ ≤ 0 is equivalent to equality; the text should say '= 0' rather than '≤ 0'.","section":"Section IV.B, Eq. (4.12)"},{"comment":"The complexity expression 'O(log(n)n32n)' appears to be a typesetting error; it should presumably be O(log(n) n^3 2^n) or similar. Please check the formula.","section":"Section VII.A"},{"comment":"The norm signs in Eq. (5.12) appear garbled in the manuscript; please ensure the displayed equation is typeset correctly.","section":"Section V.B, Eq. (5.12)"},{"comment":"The phrase 'Without loss of generality' before specifying Gaussian noise is misleading, because the exact form of Eq. (3.9) depends on the Gaussian moment structure; higher moments of a general zero-mean distribution would enter at O(ν²). This is related to the first major comment.","section":"Section III, Eq. (3.4)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a worthwhile theoretical contribution, but the central convergence claim needs to be tightened: the noise-model proportionality and the normalization constants must be handled consistently. The issues are substantial enough to require another round of revision, but they do not appear to invalidate the overall approach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nWhat you should know: this is a real result, not packaging. The new mathematical content is Theorem 1, which extends Arkhipov's TVD bound from unitary interferometers to invertible vacuum-heralded linear transformations. The proof via Schur–Weyl duality and the tensor-power Lipschitz bound is sound as far as it goes. The application of unitary averaging to boson sampling is a genuine new application, and the paper is honest about the exponential success-probability cost.\n\nWhere it does well: the bound in Eq. (4.10) is a genuine generalization, the homomorphism convergence proof is clean, and the invertibility requirement is stated explicitly. The LCU connection and the benchmarking application are useful extras. The numerics are minimal (2-mode, two photons, no error bars), but this is not primarily a numerical paper.\n\nSoft spots: the convergence claim in Eq. (4.12) is only as strong as the noise model. Equation (3.9), E(U_i) = (1 − ν/2)^d U, requires zero-mean, uncorrelated Gaussian noise. If fabrication has a systematic bias—and the paper itself lists systematic deviations and process drift as variability sources in Sec. VII.C—then U_avg converges to a biased matrix and the normalization cannot restore the target. No calibration or bias-estimation step is provided. That does not invalidate the theorem, but it does mean the 'mitigation' claim is conditional on a modeling premise, and the paper could do more to flag that. Also, the proof assumes U_avg is invertible; for finite N this is an event they do not condition on, though in the infinite limit the scalar multiple of U is invertible. Minor: encoding/decoding noise suppression is deferred to the self-cited Ref. [17]; a few lines of justification would help.\n\nOverall, the reader's conditional verdict is right. The core derivation holds under the stated model. The paper deserves refereeing: the theorem is new, the proof is accessible, and the main limitations are addressable in revision. I would send it to peer review, asking for (i) an explicit statement that zero-mean noise is load-bearing, (ii) a discussion of what happens under biased noise and whether a rescaling or calibration step can help, and (iii) either more numerics or a comment on whether the 2-mode, 2-photon example is representative.\n\nI would cite the generalized bound in my own work. Worth bringing to the reading group.","headline":"A real generalization of Arkhipov's robustness bound to vacuum-heralded networks, with an honest application to boson sampling; the convergence claim is conditional on zero-mean noise, which the paper should flag more clearly.","tokens_in":17218,"tokens_out":1961,"would_cite":true,"duration_ms":21913,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Stochastic fabrication errors in boson samplers can be averaged away: as the number of coherently averaged noisy unitaries grows, the output distribution converges to the ideal one.","keywords":["boson sampling","quantum error mitigation","unitary averaging","stochastic unitary errors","vacuum heralding","Schur-Weyl duality","linear optical quantum computing","trace distance bound"],"falsifier":"Fabricate many copies of the same interferometer with controlled per-beam-splitter parameter noise, characterize each noisy unitary by full process tomography, and form the empirical mean $(1/N)\\sum_j U_j$; if the renormalized mean is not proportional to the target unitary as $N$ grows, or if the total-variation distance between the averaged and ideal distributions does not decrease with increasing $N$ at fixed noise variance, the central convergence claim fails.","tokens_in":16162,"feed_emoji":"⚛️","tokens_out":9017,"duration_ms":90990,"temperature":0.7,"pith_summary":"Boson samplers are vulnerable to random fabrication errors that make each implemented interferometer deviate from the intended unitary. This paper claims those stochastic deviations can be mitigated by running $N$ noisy copies of the device inside a passive linear-optical network that coherently averages their unitaries and heralds on vacuum auxiliary modes. It proves that the total-variation distance between the averaged sampler's output distribution and the ideal boson sampling distribution is bounded by a quantity that vanishes as $N\\to\\infty$, provided the per-beam-splitter noise is unbiased and independent. If correct, this gives a concrete error-mitigation strategy for sampling experiments, producing an error-suppressed quantum state rather than a classical estimate, at the price of an exponentially small success probability.","feed_headline":"Average many noisy interferometers, recover ideal boson sampling","feed_subtitle":"Coherently averaging noisy interferometers converges sampling to the ideal distribution, with a proven bound.","key_machinery":"The engine is the unitary averaging framework: $N$ copies of the noisy $m$-mode interferometer $U_j$ are interleaved with $N$-dimensional discrete Fourier transforms, encoding each input mode into $N$ modes, and after propagation the $N-1$ auxiliary modes per input are heralded onto vacuum. This passively implements the generally non-unitary averaged operator $U_{\\mathrm{avg}}=(1/N)\\sum_j U_j$ on the original modes with probability $p_{\\mathrm{post}}$. The proof ingredient is a Schur-Weyl duality bound: the $n$-boson homomorphism $\\phi$ satisfies $\\|\\phi(A)-\\phi(B)\\| \\le n k^{n-1}\\|A-B\\|$ because the symmetric-subspace norm is controlled by the full tensor-power norm, and the total-variation distance between distributions is bounded by the norm distance between normalized output states. Theorem 1 then converts this into the trace-distance bound involving $A p_A^{-1/2n}$ and $B p_B^{-1/2n}$. The noise model enters only through $\\mathbb{E}(U_i)=(1-\\nu/2)^d U$, which is what makes the infinite-$N$ average coincide with the target after renormalization.","core_discovery":"On the paper's own terms, the central result is Theorem 1: for any two invertible vacuum-heralded linear transformations $A,B\\in\\mathrm{GL}(m,\\mathbb{C})$ with success probabilities $p_A,p_B>0$, the total-variation distance between the corresponding $n$-photon output distributions obeys $\\|D_A-D_B\\| \\le n k^{n-1}\\|A p_A^{-1/2n} - B p_B^{-1/2n}\\|_{\\mathrm{op}}$, where $k$ bounds the normalized operator norms. Applying this to $A=U$ and $B=U_{\\mathrm{avg}}=(1/N)\\sum_j U_j$, the unitary-averaging network's effective transformation, and using the noise model $\\mathbb{E}(U_i)=(1-\\nu/2)^d U$, the paper concludes that in the limit $N\\to\\infty$ the bound vanishes: the mitigated distribution converges to the ideal boson sampling distribution. For finite $N$ the theorem supplies a quantitative rate in terms of the distance between the target and the averaged unitary, scaled by the photon number and the normalization factor $k^{n-1}$. The cost is that this is a probabilistic protocol: under a uniform-depth implementation the success probability is lower-bounded by $(1-\\nu/2)^{2dn}$, decaying exponentially in photon number and interferometer depth.","pith_inferences":["If fabrication drift is systematic rather than zero-mean, a modified protocol that first estimates the biased mean unitary, or that calibrates the noise before averaging, would be needed; the unbiased-noise premise is what makes $U_{\\mathrm{avg}}$ point at $U$.","The $\\mathrm{GL}(m,\\mathbb{C})$ invertibility assumption, needed for the Schur-Weyl decomposition, is likely a technical rather than physical restriction: small generic perturbations of an invertible target remain invertible, so the bound should extend by continuity to near-non-invertible cases.","A finite-$N$ version of the bound could serve as a manufacturing acceptance test: compare the operator norm of $U_{\\mathrm{avg}}/p_{\\mathrm{post}}^{1/2n}$ to $U$ to certify chips before running a full boson sampling experiment.","Where particle loss rather than parameter noise dominates, combining unitary averaging with loss-tolerant encodings could yield a combined mitigation bound; the paper notes loss is easier to detect but does not prove such a combined result."],"forward_implications":["The protocol turns stochastic unitary errors into heralded loss while outputting a genuine error-suppressed quantum state, so the improved state can be reused in further computation rather than only yielding expectation values.","The success probability is at least $(1-\\nu/2)^{2dn}$ in the uniform-depth implementation; for small per-element noise this leaves an intermediate-scale regime where the method is practical despite the exponential cost.","Theorem 1 extends the known robustness bound for unitary-only interferometers to vacuum-heralded networks, quantifying how close the output distributions of any two invertible heralded interferometers are.","By choosing the encoding and decoding matrices, the same network realizes linear combinations of unitaries with passive optics and vacuum heralding, which also covers any $m\\times m$ matrix as a combination of at most four unitaries.","With $N$ nominally identical chips, the protocol is a fabrication-repeatability witness: identical chips leave the ancilla modes in vacuum, while any chip-to-chip discrepancy produces nonzero ancilla photon counts."],"supporting_citations":[{"why":"Supplies the prior robustness bound for unitary boson samplers that Theorem 1 generalizes to vacuum-heralded networks.","marker":"[35]"},{"why":"Defines boson sampling probabilities via permanents and the $n$-boson homomorphism used in the proof.","marker":"[23]"},{"why":"Introduces the unitary averaging protocol on which the error-mitigation scheme is built.","marker":"[15]"},{"why":"Earlier demonstration that averaging noisy linear-optical networks suppresses errors, motivating the present analysis.","marker":"[16]"},{"why":"Shows encoding and decoding errors can be suppressed to first order and discusses loss protection via parity encoding, supporting the idealized noise model.","marker":"[17]"},{"why":"Provides the Clements decomposition used to write noisy unitaries as products of beam splitters and to define the uniform depth $d$.","marker":"[37]"},{"why":"Provides the explicit homomorphism from single-boson to $n$-boson evolution used in the Schur-Weyl bound.","marker":"[39]"}],"fun_headline_variants":["Average noisy interferometers to recover ideal boson sampling","Unitary averaging mitigates stochastic errors in boson sampling","Error mitigation for boson sampling via unitary averaging","Averaging noisy networks restores ideal boson sampling","Proven bound: averaging noisy interferometers yields ideal sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire convergence argument rests on the hardware noise being zero-mean, independent, and identically distributed across beam splitter parameters; if fabrication errors carry a systematic bias or correlations, the averaged unitary converges to the wrong interferometer and the mitigated distribution misses the ideal one.","fun_headline_variants_meta":{"raw":{"variants":["Average noisy interferometers to recover ideal boson sampling","Unitary averaging mitigates stochastic errors in boson sampling","Error mitigation for boson sampling via unitary averaging","Averaging noisy networks restores ideal boson sampling","Proven bound: averaging noisy interferometers yields ideal sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000814,"raw_usage":{"total_tokens":3636,"prompt_tokens":1083,"completion_tokens":2553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":2475}},"tokens_in":699,"tokens_out":2553,"duration_ms":18012,"temperature":1.0,"reasoning_tokens":2475,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:52:27.913994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fabricate many copies of the same interferometer with controlled per-beam-splitter parameter noise, characterize each noisy unitary by full process tomography, and form the empirical mean $(1/N)\\sum_j U_j$; if the renormalized mean is not proportional to the target unitary as $N$ grows, or if the total-variation distance between the averaged and ideal distributions does not decrease with increasing $N$ at fixed noise variance, the central convergence claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the prior robustness bound for unitary boson samplers that Theorem 1 generalizes to vacuum-heralded networks."},{"cited_title":"Singh, A","cited_arxiv_id":null,"evidence_quote":"Introduces the unitary averaging protocol on which the error-mitigation scheme is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier demonstration that averaging noisy linear-optical networks suppresses errors, motivating the present analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Clements decomposition used to write noisy unitaries as products of beam splitters and to define the uniform depth $d$."}],"review_version":1}