{"id":"4912c0ac-dfce-4721-8d03-aa0bb5bf8f03","arxiv_id":"2602.24102","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Under combined loss and dephasing, GKP codes beat number-phase codes only when dephasing is below roughly 1/100 of the loss strength; beyond that, number-phase codes win.","lead":"This paper uses an evolutionary search to tune the parameters of two families of bosonic error-correcting codes, then maps which family survives better under combined photon loss and dephasing noise. It finds a crossover where number-phase codes win when dephasing is about one hundred times weaker than loss, offering a practical rule for code choice.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Concern: The reported 'two orders of magnitude' boundary is set by the energy caps (GKP Δ_min=0.18; NP s≤5, n≤4), not demonstrated to be stable as those caps are relaxed.","rationale":"The reader's weakest assumption is exactly this energy-cap dependence, and I agree. The attack is not that the numerics are wrong but that the claimed quantitative boundary is not demonstrated to be independent of the optimization domain. This is a load-bearing weakness of the abstract's headline number, but not of the qualitative comparison—the symmetry-based expectations and the broad shape of the fidelity landscape already support 'GKP for loss-dominated, NP for dephasing-dominated.' A single targeted rerun with relaxed caps would settle the concern. If the rerun shows instability, the verdict should remain CONDITIONAL; if it passes, the result could be upgraded. Thus no change to the reader's verdict is needed.","tokens_in":16374,"tokens_out":6364,"duration_ms":69050,"concrete_test":"Choose a fixed loss slice, e.g., γt=0.1, and relocate the boundary κt* (where ΔF~=0) under progressively relaxed caps: Δ_min=0.18→0.14→0.11→0.09 (mean photon ≈15, ≈25, ≈41, ≈62) and NP s_max=5→7→9 with correspondingly increased n_max and Fock truncation. Re-run the same CMA-ES optimization and compute κt* by bisection or dense sampling. If κt* shifts by more than a factor of 2 between the original caps and the most relaxed caps, or if the sign of ΔF~ at κt≈γt/100 changes, the reported boundary is an artifact of the energy truncation. If κt* is stable and the scaling κt*≈γt/100 persists, the concern is retired. To make the test unambiguous, release the state-preparation and QEC-matrix code used for Figs. 4–5, or a minimal script reproducing one boundary point, so the near-optimal-fidelity evaluations can be independently checked.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Sec. III C and Abstract) is that the GKP/NP performance boundary lies at κt ≈ γt/100. This number is read off a fidelity-difference contour obtained by maximizing within the fixed domains of Eqs. (25)–(26), and both families hit those domain walls in exactly the regime that sets the boundary. Sec. III A states that at κt=0.0001 the GKP optimizer 'saturates the lower bound of Δ', i.e., Δ=0.18 (≈14.9 photons), the highest energy the GPU budget allows. In the same loss-dominated corner, the NP optimizer returns s=5 and n=4.00 (Fig. 3h), the upper edges of its search box, and Sec. III B explicitly says 'if larger values of s were accessible, the optimal solution would revert to the symmetric case f=1/2.' Thus the relative fidelity at the points used to interpolate the boundary is not converged: one or both codes are being compared at truncation-imposed energies, not at energies that are optimal for the noise. Nothing in the paper quantifies how ΔF shifts as Δ_min→0 or s_max→∞. Since the headline number is extracted from this not-converged contour, it cannot be called an intrinsic boundary; it is at best a computational-budget-dependent estimate. The qualitative ordering (GKP wins under loss, NP wins under dephasing) is robust and supported by symmetry arguments and Ref. [29]; the concern is directed at the quantitative 'two orders of magnitude' claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a GPU-accelerated numerical optimization framework, combining the near-optimal fidelity metric of Ref. [32] with CMA-ES, to optimize the parameters of Gottesman-Kitaev-Preskill (GKP) and number-phase (NP) codes under a simultaneous photon-loss and dephasing channel. It reports optimized parameter trends, fidelity maps, and a performance boundary in the (γt, κt) plane, summarized as the claim that the crossover appears when dephasing is approximately two orders of magnitude weaker than photon loss (κt ≈ γt/100). The paper also introduces strict-advantage regions using the two-sided bound of the near-optimal fidelity, and it provides careful truncation and Kraus-count controls in the appendices.","tokens_in":16679,"tokens_out":5275,"duration_ms":56709,"significance":"If the quantitative boundary were established, the paper would provide a useful practical guide for choosing between GKP and NP encodings and would be a valuable methodological contribution: the use of strict fidelity bounds, analytical Kraus truncation estimates, and double-precision GPU implementations are genuinely careful. The qualitative ordering (GKP better under loss, NP better under dephasing) is plausible and consistent with symmetry arguments and with prior numerical studies. However, the headline quantitative claim rests on optima that saturate the computational search bounds, and the paper's own text documents those saturations. Because the central new claim is the specific 'two orders of magnitude' boundary, the current manuscript does not yet support that claim as an intrinsic property of the code families. The methodology is sound, but the interpretation of the numerical results needs revision or substantial additional convergence analysis.","major_comments":[{"comment":"The 'approximately two orders of magnitude' boundary is read from a contour obtained by maximizing within the domains Δ∈[0.18,0.6], s≤5, n≤4. The manuscript itself states that in the loss-dominated regime the GKP optimizer saturates the lower bound Δ=0.18 (Sec. III A), and that the NP optimizer returns s=5, n=4.00 at the upper edges of its domain (Fig. 3(h), Sec. III B), with the caveat that 'if larger values of s were accessible, the optimal solution would revert to the symmetric case f=1/2.' Thus the fidelity difference that defines the boundary is not converged in code energy: one or both codes are being compared at truncation-imposed energies rather than at energies that are optimal for the noise. Equation (23) bounds the fidelity metric for fixed parameters; it does not bound the error induced by the finite optimization domain. Since the central claim is extracted from this non-conv","section":"Sec. III C, Fig. 5, Eqs. (25)-(26)"},{"comment":"The NP codes are optimized only over envelopes θ_n = ⟨n|α,r⟩ derived from displaced-squeezed Gaussian states. Equation (14) explicitly shows that cat and binomial codes correspond to different choices of θ_n, so the comparison is not between GKP and the full NP code family defined in Sec. II A 2, but between GKP and a Gaussian-envelope subclass of NP codes. The abstract and Sec. IV state the conclusion as a property of 'NP codes' generally. A better-performing non-Gaussian envelope could shift the boundary, so the quantitative claim is conditional on this ansatz. The paper should either justify that the Gaussian envelope is sufficient for the noise regimes considered, or soften the conclusions to the optimized subclass.","section":"Sec. II A 2, Eq. (13)"},{"comment":"The black contour ΔF̃_opt = 0 in Fig. 5 is used to identify the boundary, but Eq. (23) only relates F̃_opt to F_opt through a two-sided bound. The sign of F̃_GKP − F̃_NP does not necessarily equal the sign of F_GKP − F_NP when the difference is smaller than the bound width. The strict advantage regions defined by F_lower > F_upper are conservative and well defined, but the 'approximately two orders of magnitude' line is not certified in the same way. Please report the width of the two-sided bound along the black contour and determine whether the ΔF̃=0 contour lies within the undetermined band; if so, the boundary should be presented as an uncertainty band rather than a sharp quantitative statement.","section":"Sec. III C, Eq. (23), Fig. 5"}],"minor_comments":[{"comment":"The phrases 'fundamental advantage' and 'intrinsically outperforms' are too strong given the parameter-domain restrictions documented in Eqs. (25)-(26) and the Gaussian-envelope restriction in Eq. (13). Recommend using 'within the optimized parameter families' or 'for the code families studied.'","section":"Abstract and Sec. IV"},{"comment":"Typo 'Fidliety' should be 'Fidelity'; also 'one to two orders of magnitude speedup' appears twice in consecutive sentences. Minor editorial issues.","section":"Fig. 4 caption and Sec. II C"},{"comment":"The statement 'approximately two orders of magnitude smaller' would benefit from a numerical range (e.g., the fitted κt/γt interval) and an explicit statement of the optimization-domain dependence. As written, the precision implied by 'approximately two orders' exceeds what the reported data establish.","section":"Sec. III C"}],"recommendation":"major_revision","confidential_remarks":"The numerical work is careful and the strict-advantage construction is a good feature. The key problem is that the manuscript's headline quantitative claim is extracted from a contour where both code families hit their search-domain boundaries, and the paper itself contains the evidence for this saturation. This is a load-bearing issue but fixable: additional stability/convergence scans or a more cautious interpretation would resolve it. I see no reason to doubt the qualitative ordering, but the specific 'two orders of magnitude' number should not appear as an established fact until the energy-cap dependence is addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper actually does something useful: it runs a serious parameter optimization over GKP and NP code families under combined loss-dephasing, using a near-optimal fidelity with two-sided bounds and careful Kraus truncation. The qualitative result—GKP wins when loss dominates, NP wins when dephasing dominates—is robust and consistent with earlier symmetry arguments and Ref. [29]. The new quantitative map (Fig. 5) and the idea of a crossover at κt ~ γt/100 are plausible, and the GPU+CMA-ES pipeline is reusable. That is worth something.\n\nThe soft spot is real, and it is the one that matters. The boundary is extracted in a regime where both optimizers are pressing against their search boxes. For GKP, Δ hits the lower bound 0.18 at weak dephasing; for NP, s=5 and n=4 hit the upper edges in the loss-dominated corner, and the paper itself says a larger s would revert f to 1/2. So the relative fidelity at the points that define the boundary is not converged in code energy. The “two orders of magnitude” figure has no uncertainty estimate, and the contour is linearly interpolated on a coarse grid. Without released code or data, the central number should be treated as a computational-budget-dependent estimate, not a property of the code families. The paper's own text is honest about the s issue but does not quantify its effect.\n\nThat said, the flaw is not fatal. The qualitative ordering holds, and the method is documented well enough to reproduce in principle. A serious referee should ask for a sensitivity analysis with higher energy caps (or an extrapolation), plus code/data release, before the quantitative boundary is used as a guide. But the paper deserves the referee.\n\nFor whom? Experimentalists in bosonic QEC will get a useful rule of thumb with a caveat; theorists may care about the optimization methodology. I would not cite the boundary number as a fact, but the qualitative comparison and the pipeline are worth a look.\n\nRecommendation: send to peer review with a request for additional analysis of the domain dependence.","headline":"A careful numerical comparison with a real caveat: the headline two-orders-of-magnitude boundary is set by the energy caps, not by converged code performance.","tokens_in":17315,"tokens_out":1811,"would_cite":false,"duration_ms":18392,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P70","81P45"],"pacs":["03.67.Pp"],"model":"deepseek-v4-flash","headline":"Under simultaneous photon loss and dephasing, optimized GKP codes beat optimized number-phase codes only when dephasing is roughly two orders of magnitude weaker than loss; above that ratio, number-phase codes win.","keywords":["bosonic quantum error correction","GKP code","number-phase code","photon loss","dephasing","near-optimal fidelity","parameter optimization","CMA-ES"],"falsifier":"Repeat the optimization with a lower GKP energy bound (Δ_min = 0.10, ~50 photons) and larger s (s=6,7). If the κt/γt ratio at the crossing changes by more than a factor of ~2, or if the GKP advantage region widens as energy grows, the claimed 'approximately two orders of magnitude' boundary is an artifact of the truncated search domain rather than an intrinsic code-family separation.","tokens_in":16098,"feed_emoji":"⚛️","tokens_out":3700,"duration_ms":32906,"temperature":0.7,"texified_at":"2026-08-05T20:58:46.944888+00:00","pith_summary":"This paper asks which of two families of bosonic quantum error-correcting codes—Gottesman-Kitaev-Preskill (GKP) codes, which protect against displacement errors, and number-phase (NP) codes, which protect against rotation errors—is fundamentally better when a quantum harmonic oscillator suffers both photon loss and dephasing at once. The authors' central claim is that the crossover is sharp: after optimizing each code's free parameters, GKP wins when the dephasing strength is roughly a hundred times weaker than the photon-loss strength, and NP wins when dephasing is stronger. They also find that GKP fidelity degrades much more quickly under dephasing than NP fidelity does under loss. If true, this gives experimentalists a concrete rule of thumb—measure the ratio of dephasing to loss and pick the encoding accordingly—and locates where an altogether different encoding with mixed symmetry might beat both.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":6929,"prompt_tokens":792,"completion_tokens":6137,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":792,"completion_tokens_details":{"reasoning_tokens":5424}},"feed_headline":"GKP codes win only when dephasing drops below ~loss/100","feed_subtitle":"Optimized parameter search maps the crossover between translation-protected and rotation-protected bosonic encodings.","key_machinery":"The load-bearing tool is the near-optimal channel fidelity, defined from the QEC matrix M as $(1/d_L^2) \\| \\mathrm{Tr}_L \\sqrt{M} \\|_F^2$, with the two-sided bound $(1-\\tilde{F}_{\\mathrm{opt}})/2 \\le 1-F_{\\mathrm{opt}} \\le 1-\\tilde{F}_{\\mathrm{opt}}$; it approximates true optimal fidelity well in the regime of interest and is cheap enough (matrix operations, not SDP) to embed in a CMA-ES evolutionary search over code parameters. The optimized parameter sets are {α, β, Δ} for GKP and {f, s, r, n} for NP; the fidelity difference contour and strict-advantage boundaries locate the crossover.","core_discovery":"For a single bosonic mode under combined photon loss (rate γ) and dephasing (rate κ), the paper finds numerically that optimal GKP and optimal NP codes have a well-defined performance boundary: GKP codes outperform NP codes in loss-dominated channels, NP codes outperform in dephasing-dominated channels, and the crossover occurs when κt is approximately two orders of magnitude smaller than γt. The GKP code is substantially more sensitive to dephasing than the NP code is to loss. The result is obtained by maximizing a near-optimal fidelity metric over code parameters—lattice geometry and finite-energy envelope for GKP; lattice spacing, skewness, and Gaussian envelope for NP—and locating where","pith_inferences":["The 'two orders of magnitude' boundary is likely to shift if the energy caps are relaxed: GKP optimization saturates Δ_min = 0.18 (~14.9 photons) in the loss-dominated regime, and NP s ≤ 5 in high-loss/low-dephasing; an extended search could move the crossover and reveal the asymptotic scaling.","Because the metric is near-optimal fidelity rather than threshold or logical error rate per gate, the boundary might not directly translate to fault-tolerance thresholds; a testable extension is to repeat the comparison at fixed logical error rate or with finite squeezing and measurement-level GKP error correction.","The paper's suggestion of interpolating codes on the boundary could be made concrete by constrained optimization over states with partial rotational symmetry, checking whether a hybrid code outperforms both families across a band around κt ≈ γt/100.","A practical hardware prediction: as dephasing is suppressed, the optimal GKP energy (mean photon number) should grow without bound in the pure-loss limit, while the optimal NP energy saturates; this is testable in current circuit-QED platforms."],"forward_implications":["In a laboratory where dephasing is not negligible, the choice between GKP and NP encodings can be made from the ratio κt/γt: pick GKP only when κt is roughly ≤ γt/100, otherwise pick NP.","Near the boundary, no code of either family is optimal, so codes that interpolate between translational and rotational symmetry (non-Gaussian states with mixed symmetry) are predicted to outperform both.","The optimization protocol (near-optimal fidelity + CMA-ES) transfers to other bosonic code families and noise models without re-derivation.","Optimal parameters show interpretable trends: GKP stays close to the hexagonal lattice and only rotates its excitation direction under strong combined noise; NP codes shift lattice skewness (f = 2/5, 3/5) to effectively extend number-direction distance when loss dominates.","The near-optimal fidelity's two-sided bound guarantees that the reported advantage regions are unambiguous (strict lower-upper separation)."],"fun_headline_variants":["GKP beats NP codes when dephasing is loss/100 or lower","Crossover at dephasing = loss/100: GKP for loss, NP for dephasing","Bosonic code boundary: loss favors GKP, dephasing favors NP","Where to use GKP vs NP codes: dephasing must be ~100x weaker than loss"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The boundary location is computed within bounded parameter ranges—GKP energy capped at about 15 photons and NP lattice spacing capped at s=5—and in the loss-dominated regime the GKP optimum sits exactly at the energy cap, so the claim that the crossover sits at κt ≈ γt/100 rests on these search cutoffs, not on demonstrated behavior at higher energies.","fun_headline_variants_meta":{"raw":{"variants":["GKP beats NP codes when dephasing is loss/100 or lower","Crossover at dephasing = loss/100: GKP for loss, NP for dephasing","Bosonic code boundary: loss favors GKP, dephasing favors NP","Where to use GKP vs NP codes: dephasing must be ~100x weaker than loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000419,"raw_usage":{"total_tokens":1977,"prompt_tokens":708,"completion_tokens":1269,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":1175}},"tokens_in":452,"tokens_out":1269,"duration_ms":9695,"temperature":1.0,"reasoning_tokens":1175,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:02:33.014609+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the optimization with a lower GKP energy bound (Δ_min = 0.10, ~50 photons) and larger s (s=6,7). If the κt/γt ratio at the crossing changes by more than a factor of ~2, or if the GKP advantage region widens as energy grows, the claimed 'approximately two orders of magnitude' boundary is an artifact of the truncated search domain rather than an intrinsic code-family separation.","supporting_citations":[],"review_version":1}