{"id":"19ccb6cb-92e6-4977-abf2-f22f6e8f23a2","arxiv_id":"2510.07396","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Haar-random quantum codes lose correctability exactly at the hashing bound, and the spectral band structure predicts a higher detection threshold for postselected error correction.","lead":"This paper studies what happens to the density matrix of a random quantum error-correcting code when each qudit suffers a bit of noise. It finds that recovery fails exactly at the classic hashing bound, explains this through a band structure in the eigenvalue spectrum, and shows that postselecting on low-weight errors can protect information past that threshold.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Postselected-recovery claim rests on an unverified band-alignment assumption; the authors explicitly state they have no analytic ansatz for the postselected coherent information.","rationale":"The reader's verdict is CONDITIONAL, and my stress-test identifies the same load-bearing assumption: the postselected-recovery claim depends on band isolation and eigenbasis alignment between ρ_Q and ρ_RQ past p_c, which the paper supports only heuristically and numerically to N=13. The paper itself explicitly disclaims an analytic treatment of the postselected coherent information. This does not overturn the central hashing-bound result, which has independent Weingarten-calculus support, so the verdict should remain CONDITIONAL. If the proposed exact-diagonalization check at larger N confirms the postselected coherent information remains maximal below the Rényi threshold, the secondary claim would be on much firmer ground; if not, the claim should be downgraded or explicitly reframed as a conjecture.","tokens_in":30946,"tokens_out":4530,"duration_ms":44067,"concrete_test":"Perform exact diagonalization for a single logical qubit encoded into N physical qubits, for N=15,17,19, sampling at least 200 Haar-random codes per N. For α=2, construct M_2 from the eigenbasis of ρ_Q, form σ_{Q,2} and σ_{RQ,2}, and compute the coherent information I_c(R⟩Q) for p in (0.20, 0.35), spanning p_c=0.189 and p_c^{(2)}=(3-√3)/4≈0.317. Check whether I_c stays at 1 within statistical error for p<p_c^{(2)}, and whether the transition width collapses as 1/N. Also compute the overlap between the low-weight band subspace of ρ_RQ and the corresponding eigenbasis of ρ_Q; if this overlap is not close to 1 across the relevant p-range, the heuristic 'eigenvector correspondence' underlying the postselected-recovery claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The primary hashing-bound threshold is independently supported by the Weingarten calculation and by numerics, so I do not object to that part. The load-bearing weak point is the secondary claim that postselected error correction survives past p_c up to p_d^{(α)}. That claim is a statement about the coherent information of the postselected states σ_{Q,α} and σ_{RQ,α}, not merely about the Rényi entropy of ρ_Q. To compute I_c(R⟩Q) for the postselected state one needs both S_vN(σ_{Q,α}) and S_vN(σ_{RQ,α}), where σ_{RQ,α} = (1_R ⊗ M_α) ρ_RQ (1_R ⊗ M_α^†)/Tr[ρ_Q^α]. The authors state in Sec. V of the main text: 'We have not been able to derive an analytic ansatz for the coherent information of the post-selected state beyond this saddle-point level: since the postselection protocol involves reweighting the eigenvectors of ρ_RQ by the eigenvalues of ρ_Q, we do not have an explicit expression for the eigenvalues (and therefore the entropies) of M_αρ_RQ M_α^†.' The Supplement repeats this: 'We do not have an ansatz for this quantity.' Their numerical evidence is only N≤13 (Fig. 4(d)). The heuristic substitute assumes a 'first approximation' in which each eigenvector in band w^* of ρ_RQ corresponds to q^k degenerate eigenvectors in band w^* of ρ_Q. That is precisely the band-isolation/alignment assumption the reader flagged. If the eigenbases of ρ_Q and ρ_RQ are not aligned across bands, or if inter-band reservoir mixing perturbs the low-weight blocks for p>p_c, the postselected coherent information could drop below k before p_c^{(α)}, invalidating the secondary claim even though the hashing-bound threshold itself stands. This is an admitted gap, not an invented one, and it is the least secure load-bearing element of the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the spectral and coding properties of Haar-random quantum codes subjected to independent per-qudit depolarizing noise of strength p. It claims that the spectrum of the decohered physical density matrix decomposes into Pauli-weight bands, that the coding threshold is the hashing bound H(p)=1-k/N (p_c≈0.189 for zero-rate qubit codes), and that the same bound holds for Haar-random codes as for random stabilizer codes. It further claims that for p above the hashing bound typical errors are uncorrectable but postselected error correction survives until a detection threshold p_d=(q-1)/q (1/2 for qubits), and that soft postselection by a Rényi-type reweighting produces transitions at H_alpha(p)=1-k/N. The primary threshold is supported by an explicit Weingarten replica calculation and by exact quantum MacWilliams identities for alpha=2,infinity, while the postselected coherent-information transition is supported only by numerics and a heuristic band-alignment assumption.","tokens_in":31358,"tokens_out":9355,"duration_ms":89692,"significance":"If the results hold, the paper establishes that the hashing bound is not special to stabilizer codes and gives an operational, information-theoretic meaning to Rényi-entropy singularities that have been observed in decohered quantum codes. The manuscript is unusually concrete: the Weingarten calculation is explicit, the MacWilliams-identity discussion is self-contained and potentially useful, no parameters are fitted, and the numerical data include scaling collapses. These strengths make the primary hashing-bound claim credible. The main weakness is that the advertised postselected-recovery transition is not derived analytically and rests on a band-alignment assumption that the authors themselves state is unproved.","major_comments":[{"comment":"The claim that soft-postselected error correction has a transition at p_c^(alpha), and that postselected recovery persists beyond the hashing bound, is not analytically established. The coherent information of the postselected state requires both S_vN(sigma_Q,alpha) and S_vN(sigma_RQ,alpha). The paper explicitly states in Sec. V of the main text 'We have not been able to derive an analytic ansatz for the coherent information of the post-selected state' and repeats in the Supplement 'We do not have an ansatz for this quantity.' The heuristic substitute, namely that each eigenvector in band w* of rho_RQ corresponds to q^k degenerate eigenvectors in band w* of rho_Q, is exactly the band-alignment/isolation assumption that needs testing; it does not follow from the Marchenko-Pastur or mean-shift analysis of the reduced spectra. The only evidence is Fig. 4(d) for N<=13. Since the abstract and","section":"Sec. V (main) and Sec. III (Supplement), Eqs. (12)-(16), (S33)"},{"comment":"The Weingarten derivation computes annealed averages log E_U tr rho^n and then invokes the replica limit; the claim that these equal the quenched entropies E_U S_vN(rho) is an unproved concentration assertion. Rare encodings can dominate E tr rho^n even when typical encodings behave differently, and the coding threshold is a property of typical codes. The footnote [33] states that the quantities are self-averaging 'in practice' and the N<=13 numerics support this, but the analytic statement that Haar-random codes saturate the hashing bound should be accompanied by a variance bound or by an explicit remark that the annealed average is the standard random-coding object. Please either prove or clearly state this annealed/quenched equivalence and its limitations.","section":"Supplement Sec. IV, Eqs. (S34), (S45)-(S46)"}],"minor_comments":[{"comment":"The caption is a leftover placeholder: 'FIG. 1. Your caption here.' This should be replaced with a proper caption.","section":"Fig. S3 caption"},{"comment":"Typographical issues: 'randomstabilizercodes' in the abstract and 'P AST THRESHOLD' in the Sec. V heading should be corrected.","section":"Abstract and Sec. V heading"},{"comment":"The quantity I_c^(infinity)(R>Q) is used but not defined in the main text; a one-line definition would improve readability.","section":"Supplement Sec. VII, Eq. (S73)"}],"recommendation":"major_revision","confidential_remarks":"The hashing-bound result appears solid and is the main publishable contribution. The postselected coherent-information claim is the weakest part and is currently advertised more strongly than the evidence supports. I would recommend major revision to either supply a genuine analytic argument for the postselected transition or carefully reword the abstract and introduction so that this part is presented as a numerical conjecture."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's central result—Haar-random codes saturate the hashing bound, so the depolarizing threshold is p_c≈0.189 for zero-rate qubit codes—looks right. The Weingarten calculation in the supplement is an independent derivation, and the numerics for N=5..13 are consistent. The spectral band picture is new and useful: it gives a mechanism for why the transition happens at the hashing bound, and it explains the finite-size scaling. The quantum MacWilliams identity treatment for p_d and the Rényi-2 threshold is also a nice addition, and the identity is general, not stabilizer-specific.\n\nThe soft spot is the secondary claim that postselected error correction survives past p_c up to p_d. That claim is about the coherent information of the postselected state σ_{RQ,α}, and the authors explicitly say they have no analytic ansatz for it. The heuristic band-alignment assumption—each eigenvector in band w* of ρ_RQ corresponds to q^k degenerate eigenvectors of ρ_Q—is exactly the kind of thing that can fail. If the eigenbases are not aligned across bands, or if reservoir mixing perturbs the low-weight blocks for p>p_c, the postselected coherent information could drop below k before p_c^{(α)}. The numerics are only N≤13, so the scaling collapse in Fig 4(d) is suggestive, not decisive. This is an admitted gap, not an invented one, and it is the least secure load-bearing element of the paper.\n\nThe hashing-bound threshold does not depend on that assumption, and I would not reject the paper for the gap. But the authors should be asked to either provide a direct calculation of the postselected coherent information (even in an annealed approximation) or to state clearly that the postselected-recovery claim is a conjecture supported by small-N numerics. It would also help if they released the simulation code and data.\n\nThis paper is for people working on mixed-state phase transitions and QEC thresholds. It deserves a serious referee: the main result is important enough, and the derivations are original. I would send it to review with a request that the postselected claim be addressed head-on.","headline":"Main threshold result is solid and independently derived; the postselected-recovery claim is the genuinely weak spot and should be either proven or labeled as a conjecture.","tokens_in":31851,"tokens_out":1901,"would_cite":true,"duration_ms":18864,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P73","81P45"],"pacs":["03.67.Pp","03.67.-a"],"model":"deepseek-v4-flash","headline":"Random quantum codes saturate the hashing bound","keywords":["Haar-random quantum codes","hashing bound","error-correction threshold","mixed-state phase transition","spectral bands","postselected error correction","Rényi entropy","quantum MacWilliams identity"],"falsifier":"Evaluate the exact or well-converged coherent information of a Haar-random code at N=20–30 under depolarizing noise for p between 0.19 and 0.5. If the low-weight bands fail to stay orthogonal—say the projected density matrix Π_w ρ Π_w loses logical information below p=1/2—the postselected-recovery claim is false; if the threshold drifts away from p_c≈0.189 as N grows, the hashing-bound claim for Haar-random codes is false.","tokens_in":30879,"feed_emoji":"🎲","tokens_out":6520,"duration_ms":55876,"temperature":0.7,"pith_summary":"Random quantum codes—codes that put logical information in a random subspace of many noisy qubits—lose the ability to correct errors at exactly the hashing bound, the same threshold as random stabilizer codes. For a single qubit encoded in many qubits, that threshold is p_c≈0.189. The paper establishes this by tracking the eigenvalue spectrum of the corrupted density matrix, which splits into bands labeled by how many qubits an error touched. Below threshold these bands stay separated; at threshold the typical error band fills all available space. Above threshold, typical errors are uncorrectable, but if one postselects on low-weight error bands, information remains safe until a much higher detection threshold p_d=1/2.","feed_headline":"Random quantum codes saturate the hashing bound","feed_subtitle":"Band structure of the decohered spectrum pins the qubit threshold at p≈0.189, with postselection surviving to p=1/2.","key_machinery":"The central object is the band spectrum of the decohered density matrix. The depolarizing channel is decomposed as a convex sum of fixed-weight error channels; each weight-w channel maps the logical subspace to a nearly orthogonal subspace, producing a 'band' of eigenvalues of size ~p^w with multiplicity (q^2−1)^w C(N,w). Each band's internal level density is taken to be Marchenko-Pastur (random-matrix), and inter-band overlaps are treated as small perturbations. The threshold calculation is independently confirmed by Weingarten calculus, which evaluates Haar-averaged purities; the detection and Rényi-2 thresholds are fixed by a quantum MacWilliams identity, a duality between the error pictu","core_discovery":"The paper's central claim is that a Haar-random encoding—a code whose logical space is a random subspace of the physical Hilbert space—has an error-correction transition at the hashing bound H(p_c)=1, matching random stabilizer codes. The argument runs through the spectrum of the decohered density matrix: each Pauli error of weight w creates an approximately orthogonal copy of the logical space, so the spectrum is a sum of bands of eigenvalues at scale p^w; the threshold is where the typical band's dimension exhausts the Hilbert space. The claim is backed two independent ways: a perturbative 'mean-shift' ansatz for the band positions, validated numerically for up to N=13 qubits, and an expli","pith_inferences":["A concrete test beyond the paper: at N≈20–30, simulate Haar-random codes and check whether the postselected coherent information on the w=0 and w=1 bands stays maximal up to p=1/2; no exact result at that size is currently at hand.","The derivation via weight enumerators suggests that the hashing-bound saturation is generic for nondegenerate subspace codes, not a special property of stabilizer structure; if true, other non-stabilizer code families should show the same threshold.","The paper's band picture indicates that the low-weight bands in local codes like the surface code will be modified by error degeneracy; whether a similar postselection window survives there is testable with existing statistical-mechanics mappings.","The finite-rate detection formula p_d=1−q^{r−1} implies a shrinking postselection window as the code rate grows, which could inform practical rate-versus-overhead tradeoffs in fault-tolerant designs."],"forward_implications":["For zero-rate qubit codes, the error threshold is p_c≈0.189, and the transition broadens as 1/√N at finite size.","For error rates between p_c and p_d=1/2, logical information is unrecoverable without postselection but is protected if one projects onto low-weight error bands; for finite-rate codes the detection threshold becomes p_d=1−q^{r−1}.","Rényi-α entropies of the decohered state are nonanalytic at H_α(p)=1, and these singularities are exactly the success/failure transition of a 'soft' postselection POVM.","The result generalizes to qudits: correction threshold at H(p)=1, detection at p=1−1/q.","The two postselection protocols—weight-projecting and Rényi-reweighting—are inequivalent, mirroring ensemble inequivalence in long-range statistical mechanics."],"fun_headline_variants":["Haar-random codes: error threshold at hashing bound","Random quantum codes cross at hashing bound","Hashing bound sets random code threshold","Band spectra reveal hashing bound threshold","Postselected errors survive past hashing bound"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that images of the logical subspace under distinct Pauli errors stay nearly orthogonal, and inter-band mixing stays negligible, for all weights below the hashing weight; the paper supports this by perturbation theory and numerics for up to 13 qubits, but does not prove it rigorously.","fun_headline_variants_meta":{"raw":{"variants":["Haar-random codes: error threshold at hashing bound","Random quantum codes cross at hashing bound","Hashing bound sets random code threshold","Band spectra reveal hashing bound threshold","Postselected errors survive past hashing bound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1203,"prompt_tokens":757,"completion_tokens":446,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":379}},"tokens_in":501,"tokens_out":446,"duration_ms":4327,"temperature":1.0,"reasoning_tokens":379,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T10:56:41.112811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the exact or well-converged coherent information of a Haar-random code at N=20–30 under depolarizing noise for p between 0.19 and 0.5. If the low-weight bands fail to stay orthogonal—say the projected density matrix Π_w ρ Π_w loses logical information below p=1/2—the postselected-recovery claim is false; if the threshold drifts away from p_c≈0.189 as N grows, the hashing-bound claim for Haar-random codes is false.","supporting_citations":[],"review_version":1}