{"id":"44ab59c9-93d9-48a6-9299-c57ca7ac8a23","arxiv_id":"2509.16439","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"For the maximally mixed state, aggressive truncation, Riemannian gauge optimization, and the analytic disentangler V = e^{-iφ}U each reduce a sub-optimal LPDO representation to the minimal χ = 1 bond dimension.","lead":"Noisy quantum operations push a quantum state to pure randomness, and a standard simulation representation wastes memory on fake entanglement there. This study gives three pruning tools that compress that representation to its minimum size, making noisy circuit simulations cheaper.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fidelity-preserving truncation is exact for full depolarization but the paper leaves it unproven; the claimed optimal κ=2 endpoint is not demonstrated.","rationale":"I agree with the reader that the fidelity-preservation premise is the load-bearing point, but on analysis it does not fail: the local channel at the stated error rates is exactly the full depolarizing channel, and the Schmidt-structure argument above makes truncation exact for any subset of singular values. Therefore the reader's worry about larger N, larger χmax, or instance dependence is theoretically settled, though the paper's failure to provide the argument remains a genuine gap. The more concrete overclaim is the κ=2 optimal endpoint: the numerical routines act only on the χ virtual indices; κ is inflated by the noise channel to dimension 4 and is never reduced. A simple post-hoc κ-SVD would restore κ=2, but that step is absent from the manuscript. The abstract additionally promises an interpolation in depolarization that the body explicitly defers. These issues are correctable with a proof and a few clarifying statements, so the reader's CONDITIONAL verdict stands unchanged. I would not reject the paper; the core truncation phenomenon is real and reproducible with the analytic argument added.","tokens_in":16629,"tokens_out":29932,"duration_ms":270705,"concrete_test":"Run the Sec. III A truncation protocol on one N=100, χmax=16 instance, but at a single bond keep only the smallest χ-singular value (rather than the largest); if the resulting physical state is still exactly I/2^N, the 'pure gauge' premise is confirmed and the empirical claim is exact. Separately, record the κ dimensions of the final χ=1 tensors from Fig. 4; if κ>2 persists, the optimal κ=2 endpoint is not reached by the demonstrated tools.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical result (Sec. III A, Figs. 3–4, 12a) rests on an empirical assertion: discarded χ-singular values of the sub-optimal LPMM are invisible to the physical state, so large-cutoff truncation plus renormalization preserves the state exactly. The manuscript provides no analytic argument and validates the claim on single random-MPS instances, so the reader cannot distinguish this from an artifact of the chosen instance. In fact the assertion is provable: with the bitflip and dephasing rates set to γ=0.5, the local channel is the full depolarizing channel E(X)=Tr[X] I/2. If the pre-noise state has Schmidt decomposition |ψ⟩=Σ_α λ_α |L_α⟩|R_α⟩, the post-noise purification has Schmidt vectors whose reduced physical states are E(|L_α⟩⟨L_β|)=δ_{αβ} I_L/2^{|L|}, so every cross term vanishes and every individual Schmidt component traces to the maximally mixed state. Hence any χ-truncation with renormalization preserves ρ=I/2^N exactly. The paper should state this proof. Separately, the conclusion 'completely solved' overstates the results: all numerical tools prune only χ, while κ remains at its post-noise dimension (4 for the bitflip+dephasing product), and no compression to the claimed optimal κ=2 is demonstrated. The abstract's promised interpolation in depolarization strength is also absent from the body, as the near-maximally-mixed regime is deferred to future work.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies locally purified density operator (LPDO) representations of the maximally mixed state (which the authors call LPMM). It constructs a sub-optimal LPMM representation by applying bitflip and dephasing noise to a random pure MPS, leaving the coherent bond dimension χ unchanged while inflating the mixture dimension κ. The central claim is that this sub-optimal representation can be mapped to the optimal χ=1, κ=2 representation by three tools: fidelity-preserving truncation with a large cutoff Λ, Riemannian optimization in the κ-subspace with entropy-based objectives, and an analytic disentangler derived from weak injectivity and the special form A(i)=1/√2. The analytic derivation in Eqs. (10)-(11) is simple and parameter-free. The numerical sections report that truncation preserves fidelity to machine precision, that χ can be pruned to one, and that χ_mean follows an exponential scaling in sweeps and cutoff, with fit parameters in App. C. The paper concludes that the problem of re-optimizing and pruning the LPMM representation is 'completely solved.'","tokens_in":16894,"tokens_out":4757,"duration_ms":42716,"significance":"If the central claim is fully supported, the paper would give a practical prescription for compressing LPDO representations of depolarized states, with direct relevance for classical simulation of noisy quantum circuits and for error-mitigation workflows. The analytic disentangler in Eqs. (10)-(11) is a genuine strength: it is derived, not fitted, and it explicitly identifies the κ-isometry that inverts a physical unitary on the optimal LPMM. The numerical work also has strong internal checks: norm conservation and machine-precision fidelity are reported for the truncation protocol, and App. B documents the behavior of the optimization variants. However, the significance is currently limited by three gaps: the exactness of fidelity-preserving truncation is only asserted empirically, the claimed mapping to the optimal κ=2 representation is not demonstrated numerically, and the abstract promises a depolarization interpolation that does not appear in the body.","major_comments":[{"comment":"The central claim that large-cutoff SVD truncation in the χ-subspace is fidelity-preserving for the LPMM rests on an empirical assertion ('more forgiving', 'empirical results presented will show') rather than on a proof. This is load-bearing because the pruning results in Fig. 4 and the exponential fits in App. C assume that the truncated state is still exactly ρ = I/2^N. For γ_d = γ_b = 0.5 the local channel is the full depolarizing channel, so the proof is elementary: if the initial purification has Schmidt decomposition |ψ⟩=Σ_α λ_α |L_α⟩|R_α⟩, then after the channel the reduced physical state of each Schmidt component is δ_{αβ} I/2^N, and every cross term vanishes; hence any χ-truncation followed by renormalization preserves ρ exactly. Please include this argument, and state clearly whether or not it extends to the finite-depolarization regime mentioned in the abstract.","section":"Sec. III A, Figs. 3-4, 12(a,b)"},{"comment":"The paper promises to map the sub-optimal representation to the optimal χ=1, κ=2 representation, but the numerical protocols only prune χ. The κ dimension remains at its post-noise value (4 for the product of bitflip and dephasing channels), and no numerical or analytical compression of κ is demonstrated. The analytic disentangler in Eqs. (10)-(11) is derived for the optimal tensor A(i)=1/√2 with κ=2 and does not apply directly to the κ=4 tensors generated by the noise protocol in Fig. 2. Consequently, the conclusion in Sec. IV that the problem is 'completely solved' overstates what is shown; at present the paper solves χ-pruning in the fully depolarized case and identifies the disentangler only for a restricted starting representation.","section":"Sec. II and Sec. III C; Fig. 2; Eqs. (10)-(11)"},{"comment":"The abstract states that 'away from the maximally mixed state, our simulations highlight how the truncation threshold smoothly interpolate, as a function of depolarization, between established matrix product results and our new results.' No such simulation or interpolation appears in the body; Sec. IV explicitly defers the near-maximally-mixed regime, ρ=(1-ε)ρ_1+εσ, to future work. Please add the missing data or remove this claim from the abstract. As written, the abstract promises a result the manuscript does not deliver.","section":"Abstract vs. Sec. IV"},{"comment":"The Riemannian optimization protocol is presented as a tool to prune χ while preserving the state, but App. B reports that its infidelity is three to four orders of magnitude larger than that of fidelity-preserving truncation for Λ≤0.3, reaching roughly 10^-5 in the worst case. This is not a negligible effect if the goal is an exact or near-exact representation of the LPMM, and it is not discussed in the main text beyond a brief remark. The paper should quantify the accuracy-versus-complexity trade-off of the optimizer, state whether the reported χ=1 endpoint is obtained at the cost of controlled approximation error, and reconcile this with the claim that the re-optimization problem is 'completely solved.'","section":"Sec. III B, Fig. 12(c,e); App. B"}],"minor_comments":[{"comment":"The phrase 'the approximation error (using fidelity as a metric) scaling is more forgiving' is vague; please define the fidelity metric explicitly and state the observed scaling (e.g., infidelity versus Λ) before discussing the numerical results.","section":"Sec. III A"},{"comment":"There are several typographical and grammatical errors: 'intution' in Sec. III A, 'emperical' in Sec. IV, 'guage' in the caption of Fig. 11, 'distenanglers' in Sec. IV, and 'the the' in Sec. II A. These should be corrected before publication.","section":"Throughout"},{"comment":"The numerical scaling results appear to be based on single random-MPS instances, but this is not stated explicitly. Please specify the number of instances used, and, if error bars in Fig. 13 reflect only fit covariance rather than instance-to-instance variability, say so.","section":"Figs. 4 and 13"},{"comment":"The step in the weak-injectivity discussion where the χ-space isometry M is set to the identity is asserted rather than justified; because the optimal LPMM has χ=1 this is plausible, but the reasoning should be stated explicitly for readers working with sub-optimal χ>1 representations.","section":"Sec. III C, Eq. (9)"},{"comment":"The caption says 'fidelity' without specifying the definition; Eq. (B1) defines F_P, but the main text and figure caption should connect them, especially because panels (c) and (e) show deviations of order 10^-5.","section":"Fig. 12"}],"recommendation":"major_revision","confidential_remarks":"The paper is a direct follow-up to the authors' Ref. [18], and the incremental novelty is substantial. The main concern for the editor is the gap between the abstract's strong claims (including the depolarization interpolation) and the body's content; the empirical truncation-exactness claim should be replaced by the simple proof sketched in my first major comment. The κ-compression gap should also be stated honestly in the conclusion; with those fixes the paper would be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing here is the numerical observation: a sub-optimal LPDO for the maximally mixed state (χ inherited from a random MPS purification) can be pruned to χ=1 by SVD truncation with a large cutoff Λ, with no loss of fidelity, and the cost scales like a simple exponential in Λ. The Riemannian optimization variants are also reasonable and sensibly compared. That part is worth knowing.\n\nWhat the paper misses is that the main numerical claim is provable in a few lines, and the proof would have made everything cleaner. Setting γ_b=γ_d=0.5 makes the local channel the full depolarizing channel E(X)=Tr[X]I/2. Take any bipartition of the original pure state with Schmidt decomposition Σ_α λ_α |L_α⟩|R_α⟩. Apply E to every physical site. The reduced physical state of each Schmidt component is E(|L_α⟩⟨L_β|)=δ_αβ I/2^{|L|}, so cross terms vanish and every kept component is maximally mixed. Any χ-truncation, after renormalization, therefore reproduces ρ=I/2^N exactly. The paper asserts this regularity empirically; it does not prove it, and there is no need to leave it empirical.\n\nThe soft spots are real but not fatal. The numerics appear to be single random-MPS instances with no disorder averaging, and no code or data are shipped, so the exponential fits in Fig. 13 are not independently checkable. The conclusion 'completely solved' is too strong: only χ is pruned; κ stays at its post-noise value of 4, and the paper never demonstrates the optimal κ=2 endpoint. The abstract's promise of smoothly interpolating between MPS and LPDO truncation thresholds as a function of depolarization is not present in the body; the near-maximally-mixed regime is explicitly deferred. The fidelity measure in Eq. B1 is non-standard—it is 1 trivially for two maximally mixed states—though it does detect deviations from the target, so it is defensible if explained. The SU(2) equivalence claim for optimized disentanglers is asserted without proof, minor.\n\nThe analytic disentangler (Eqs. 10–11) is correct but nearly trivial: since A(i)=1/√2, any unitary commutes with it. Fine as a check, not a load-bearing result.\n\nWho is this for? Anyone doing LPDO-based simulation of noisy circuits or zero-noise extrapolation, where the fully depolarized fixed point matters. It deserves a serious referee: the core observation is useful and the missing proof is short. I would send it to review, but ask for instance averaging, code/data release, a statement of the proof above, and toned-down claims about κ and the interpolation.\n\nRecommendation: major revision, then accept.","headline":"A useful empirical pruning result for the maximally mixed LPDO that misses a short proof it could easily have included, and overclaims on κ and the depolarization interpolation.","tokens_in":17532,"tokens_out":6303,"would_cite":true,"duration_ms":55693,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Depolarized quantum states can be compressed exactly to their minimal tensor representation.","keywords":["locally purified density operator","maximally mixed state","entanglement pruning","fidelity-preserving truncation","Riemannian optimization","weak injectivity","tensor network gauge freedom","depolarizing noise"],"falsifier":"Run the fidelity-preserving truncation sweeps on many random initial MPS instances at N=100, χmax=16, with Λ=0.5, and measure fidelity to the ideal maximally mixed state; visible infidelity beyond machine precision on any instance, or growth of χmean that does not flatten with sweeps, would falsify the pure-gauge assumption and the exponential pruning scaling.","tokens_in":16193,"feed_emoji":"✂️","tokens_out":4190,"duration_ms":32726,"temperature":0.7,"pith_summary":"The paper resolves the sub-optimality of Locally Purified Density Operator (LPDO) representations of the maximally mixed state, a fixed point reached by noise in near-term quantum computations. It shows that a representation whose coherent bond dimension is inherited from an initial random pure state can be mapped to the optimal form with unit bond dimension and purity index two, both numerically and analytically. If correct, fully depolarized LPDO simulations compress to the minimal memory cost of 2N coefficients, and the same gauge-fixing logic can guide simulations of states close to infinite temperature.","feed_headline":"Depolarized quantum states prune to minimal tensor form","feed_subtitle":"Fidelity stays exact under heavy truncation, and the disentangler is simply the inverse of the applied unitary.","key_machinery":"The central object is the Locally Purified Density Operator (LPDO), a tensor network for mixed states in which each site carries a mixture (κ) index and a coherent (χ) bond. The load-bearing identity is A(i)=1/√2 for the optimal tensor at every site, so U A(i)=A(i) U and the disentangling isometry in the κ-subspace is V=$e^{{-iφ}}$U; equivalently V† U=$e^{{iφ}}$1. This identity turns unitary action on physical indices into an isometry action on mixture indices, proving that spurious χ-correlations can be pruned exactly. Numerically, the mechanism is fidelity-preserving truncation with a large cutoff, which works because renormalization in the χ-subspace restores the separable state.","core_discovery":"The paper claims that a sub-optimal LPDO representation of the maximally mixed state, whose coherent bond dimension χ is inherited from an initial random pure state, is equivalent to the optimal representation with χ=1 and κ=2 at every site. This is demonstrated numerically through fidelity-preserving truncation with a large cutoff Λ and through Riemannian optimization over isometries in the mixture subspace, and analytically through the identity A(i)=1/√2, which commutes with any unitary. Because the noise-depolarized state is separable, the discarded χ singular values are pure gauge, so renormalization restores the exact physical state.","pith_inferences":["If the pure-gauge premise survives disorder averaging and larger system sizes, the same pruning strategy should apply to states near the fixed point, ρ=(1-ε)ρ1+εσ, with a smooth crossover in truncation threshold.","The analytic disentangler suggests a constructive recipe: for any unitary layer applied to a locally purified state, one can attempt to invert it in the mixture subspace rather than the physical space, which may speed up classical simulation of noisy circuits.","The connection between weak symmetry and weak injectivity implies that symmetry-protected mixed-state phases with weakly symmetric LPDOs may admit similar closed-form disentanglers, a direction the paper's methods make testable."],"forward_implications":["A fully depolarized LPDO simulated from a random initial state can be stored with 2N coefficients rather than a bond dimension that grows with χmax, an exponential memory saving.","Choosing a large truncation cutoff Λ does not damage fidelity for the separable maximally mixed state, contradicting the usual MPS intuition that error grows with cutoff.","Riemannian optimization over isometries in the κ-subspace achieves χ pruning, with objective functions based on second Renyi or von Neumann entropy, and is most useful when Λ is too small to prune alone.","The analytic disentangler V=e^{-iφ}U shows that unitaries acting on the optimal state induce exactly invertible gauge freedom, so representational entanglement can be removed without physical approximation.","Away from full depolarization, the truncation threshold interpolates between matrix product operator results and the maximally mixed results."],"supporting_citations":[{"why":"Prior finding that the LPDO representation of the maximally mixed state is sub-optimal; supplies the problem this paper solves.","marker":"[18]"},{"why":"Establishes that weak symmetry implies weak injectivity for LPDOs, the analytical basis for prunable spurious entanglement.","marker":"[3]"},{"why":"Defines the positive tensor network formalism including κ-subspace truncation and renormalization that the pruning protocol builds on.","marker":"[17]"},{"why":"Introduces entanglement pruning and fast tensor disentangling, providing the terminology and goal this paper adapts to LPDOs.","marker":"[19]"},{"why":"Provides Riemannian optimization of isometric tensor networks, used for the κ-subspace isometry search.","marker":"[24]"},{"why":"Shows how to compute a two-site reduced density matrix with a specific contraction sequence and notes tail growth in singular values, both used here.","marker":"[29]"}],"fun_headline_variants":["Maximally mixed states prune to minimal tensor form","Depolarized states exact at χ=1 via gauge freedom","Entanglement pruning restores optimal LPDO exactly","Quantum noise enables exact state compression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated around the truncation protocol, is that the discarded singular values in the χ-subspace of a maximally mixed LPDO are pure gauge, so a large cutoff plus renormalization recovers the exact state; this is demonstrated on single random instances without an analytic proof.","fun_headline_variants_meta":{"raw":{"variants":["Maximally mixed states prune to minimal tensor form","Depolarized states exact at χ=1 via gauge freedom","Entanglement pruning restores optimal LPDO exactly","Quantum noise enables exact state compression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1516,"prompt_tokens":930,"completion_tokens":586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":525}},"tokens_in":546,"tokens_out":586,"duration_ms":5780,"temperature":1.0,"reasoning_tokens":525,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:51:10.563410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the fidelity-preserving truncation sweeps on many random initial MPS instances at N=100, χmax=16, with Λ=0.5, and measure fidelity to the ideal maximally mixed state; visible infidelity beyond machine precision on any instance, or growth of χmean that does not flatten with sweeps, would falsify the pure-gauge assumption and the exponential pruning scaling.","supporting_citations":[{"cited_title":"Wanisch, N","cited_arxiv_id":null,"evidence_quote":"Prior finding that the LPDO representation of the maximally mixed state is sub-optimal; supplies the problem this paper solves."},{"cited_title":"This is due to the fact that given a representation, one could generate multiple equivalent representations by in- serting isometries into the virtual𝜒-indices","cited_arxiv_id":null,"evidence_quote":"Establishes that weak symmetry implies weak injectivity for LPDOs, the analytical basis for prunable spurious entanglement."}],"review_version":2}