{"id":"fb790932-3f5f-4c9c-bb5e-e928076551f4","arxiv_id":"2412.12812","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The distinct nonzero spectrum of any generating model's transfer operator bounds quantum generative memory by |Λ|^1/4 and classical memory by |Λ|^1/2, implying a quadratic quantum advantage.","lead":"This paper claims that the spectrum of a transfer operator built from any quantum hidden Markov model of a stochastic process provides strict lower bounds on the memory needed to generate the process, and that quantum models can need quadratically less memory than classical ones. It illustrates the claimed advantage with a three-state example that runs on a two-dimensional quantum memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's claimed process invariant is false: extra transfer-operator eigenvalues with zero coefficient α can be added via a decoupled memory sector, invalidating the |Λ| bounds in Theorems 2 and 3.","rationale":"The reader's weakest_assumption correctly identifies the α_λ=0 loophole in Theorem 1, and this is the single most load-bearing concern: the paper's advertised spectral invariants and the resulting memory lower bounds all rest on the claim that full nonzero spectra of any generating model coincide. The proof's own appendix states the exception, and a simple explicit counterexample (a decoupled memory sector added to an i.i.d. process) shows the theorem is false as stated. Because Theorem 2 and Theorem 3 use |Λ| from any valid generator, their bounds are not valid as written. I do not see a way around this without replacing Λ with the contributing spectrum {λ : α_λ≠0}, which would require the paper's definitions, theorem statements, and proofs to be reworked. The qualitative possibility of a quantum memory advantage may survive such a correction, and the Theorem 4 example is plausible, but the central claim as presented is not supported. The reader's REJECT verdict therefore stands; no verdict adjustment is needed. The critique is directed at the mathematical argument, not at the authors; the paper contains a real internal inconsistency that is directly testable.","tokens_in":14000,"tokens_out":6517,"duration_ms":63992,"concrete_test":"Implement the two-state classical HMM with T^(0)=diag(1/2, 0.3), T^(1)=diag(1/2, 0.7), initial π=(1,0). Verify that all finite-sequence probabilities match the fair-coin process (a one-state HMM with T^(0)=T^(1)=1/2). Compute the distinct nonzero eigenvalues of E = T^(0)⊗T^(0)+T^(1)⊗T^(1): the set is {0.5, 0.58}, whereas the one-state model gives {0.5}. This falsifies Theorem 1's Λ invariant. Then compute the Theorem 2 bound from the two-state model: c_Q ≥ log⌈2^{1/4}⌉ = log 2, while an explicit one-dimensional instrument generating the fair coin gives c_Q = 0, demonstrating the bound's failure as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is that Λ, the set of distinct nonzero eigenvalues of the transfer operator, is an invariant of the generated process. The proof of Theorem 1 in Appendix A does not establish this: the Vandermonde argument (Eqs. A9–A19) shows only that if an eigenvalue appears in one model's transfer operator but not the other's, then its coefficient α_λ = ⟨⟨1|Π_λ|ρ⟩⟩ must vanish. It does not show such eigenvalues cannot exist. The appendix itself concedes the equality of spectra 'unless α_λ = 0', which contradicts the theorem statement's unconditional Λ_R = Λ_Q. This gap is not hypothetical. Take the i.i.d. fair-coin process. A one-state classical HMM generates it with transfer-operator eigenvalue set {1/2}. Now take a two-state classical HMM with T^(0)=diag(1/2, 0.3), T^(1)=diag(1/2, 0.7), and initial distribution π=(1,0). The second state is never occupied, so the generated process is unchanged. But the transfer operator E = T^(0)⊗T^(0)+T^(1)⊗T^(1) has distinct nonzero eigenvalues {1/2, 0.58}; the eigenvalue 0.58 has α=0 because the initial vectorized state has no support on the decoupled sector. Thus Eq. (7) is false. Consequently Theorem 2's bound c_Q ≥ log⌈|Λ|^{1/4}⌉ is invalid when Λ is taken from an arbitrary generating model: for the two-state model |Λ|=2 gives c_Q ≥ log 2, whereas the fair-coin process has c_Q=0. The central invariant must be restricted to eigenvalues with α_λ≠0; as stated, the paper's bounds do not follow.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes spectral invariants for quantum hidden Markov models (QHMMs). It defines the vectorized transfer operator E_AB = Σ_x E_x ⊗ E_x, claims in Theorem 1 that the set Λ of distinct nonzero eigenvalues of this operator is invariant across all QHMMs generating the same stochastic process, and uses this claim to lower-bound the generative topological complexity: c_Q ≥ log⌈|Λ|^{1/4}⌉ in Theorem 2 and c_C ≥ log⌈|Λ|^{1/2}⌉ for classically generated (strictly incoherent) processes in Theorem 3. Theorem 4 then constructs a three-state classical HMM that admits a two-level quantum realization, claimed to show c_Q < c_C. The appendices contain the proofs and the explicit construction of the quantum model from a classical HMM.","tokens_in":14356,"tokens_out":11624,"duration_ms":121192,"significance":"If correct, the spectral-invariant approach would be a genuinely useful tool: minimality of generative models is generally hard to certify, and a computable invariant from any non-minimal model would give rigorous memory lower bounds. The tensor-network formulation of QHMMs, the resource-theoretic framing in terms of coherence, and the explicit constructive example of a quantum memory advantage are appealing features. The paper is clearly written and the main ideas are easy to identify. However, the central invariant as stated is false, and Theorem 2 and Theorem 3 inherit this flaw. The underlying idea appears salvageable by restricting to eigenvalues with nonzero coefficient α_λ, but the manuscript as written does not state, prove, or use that restriction. The significance of the contribution is therefore conditional on a substantial correction of the main theorem and its consequences.","major_comments":[{"comment":"Theorem 1 is false as stated. The Vandermonde argument in Appendix A does not show that the symmetric difference Λ_R Δ Λ_Q is empty; it shows only that any eigenvalue appearing in exactly one model must have zero coefficient α_λ. The appendix itself concedes this with the phrase \"unless α_λ_R = 0 or α_λ_Q = 0\". This is not a harmless edge case. For the i.i.d. fair-coin process, a one-state model with T^(0)=T^(1)=1/2 has transfer-operator eigenvalue set {1/2}, while the two-state model with T^(0)=diag(1/2,0.3), T^(1)=diag(1/2,0.7), and initial distribution π=(1,0) generates the same fair-coin process but has transfer operator E=Σ_x T^(x)⊗T^(x) with distinct nonzero eigenvalues {1/2,0.58}; the extra eigenvalue has α=0 because the initial state has no support on the second decoupled sector. Thus Eq. (7), the claimed process invariance of Λ, is false. The theorem also silently assumes diagonalizability of the transfer operator; footnote [60] asserting that non-diagonalizable matrices are measure zero is not a proof that all valid QHMM transfer operators are diagonalizable.","section":"Theorem 1 and Appendix A (Eq. (7), Eqs. (A9)–(A19))"},{"comment":"Because Theorem 1 is false, the bounds in Theorems 2 and 3 are not valid for an arbitrary generating model. In the fair-coin counterexample above, applying Theorem 2 to the two-state model would give c_Q ≥ log⌈2^{1/4}⌉ = log 2, while the fair-coin process has c_Q = 0. The source of the error is that Λ, as defined, includes eigenvalues that are in the spectrum of the transfer operator but do not contribute to any sequence probability because their coefficient α_λ = ⟨⟨1|Π_λ|ρ⟩⟩ vanishes. The proofs must be revised to use the \"active spectrum\" S = {λ ∈ Λ : α_λ ≠ 0}, with a proof that S is invariant across generating models. Without this correction, the claimed lower bounds in Eqs. (11) and (12) do not follow.","section":"Theorems 2 and 3 (Eqs. (11)–(12))"},{"comment":"The proof of Theorem 4 is incomplete as written. The three-state classical model is asserted to be \"classically irreducible\", but no argument is given that no two-state classical HMM generates the same process. The text says \"using Thm. 3 we can verify c_C = log 3\", but Theorem 3 is only a lower bound, and no value of |Λ| for the example is actually computed. Similarly, the construction gives a two-dimensional quantum memory, so c_Q ≤ log 2, but to conclude c_Q = log 2 one must exclude one-dimensional memories; this is not shown. The authors should supply the missing computation of the invariant (with the corrected α_λ ≠ 0 definition) and a direct minimality argument for the classical model.","section":"Theorem 4 and its proof (Eq. (14), Appendix E)"}],"minor_comments":[{"comment":"There is a duplicated word: \"any any valid presentation\" near the end of the introduction.","section":"Introduction"},{"comment":"There are several typos: \"memeory reduction\" and \"memeory states\" should be \"memory reduction\" and \"memory states\".","section":"Proof of Theorem 4 / Appendix E"},{"comment":"The definition \"¯⋆ := 1−⋆1−⋆2\" appears to be a typo; it should presumably be \"¯⋆ := 1−⋆\".","section":"Eq. (14)"},{"comment":"The displayed inequality \"c_C(−→X)^2 ≥ log[rank(E_RCmin)] ≥ log |Λ|\" appears dimensionally wrong: since c_C = log m and rank ≤ m^2, the correct inequality is 2 c_C ≥ log rank, not c_C^2 ≥ log rank. The final bound in Eq. (12) is correct, but the proof as written needs this corrected.","section":"Appendix D, Eq. (D8)"},{"comment":"The sentence \"any minimal QHMM (without the SIO constraint) cannot exceed the classical HMM's memory since this would contradict Thm. (2)\" is not a valid justification for c_C ≥ c_Q. The inequality c_C ≥ c_Q follows simply because every classical HMM is a QHMM, so the minimum in Eq. (2) can only be smaller.","section":"Theorem 3 proof (main text, after Eq. (13))"}],"recommendation":"major_revision","confidential_remarks":"I considered an outright rejection, since Theorem 1 is false as stated and the bounds in Theorems 2 and 3 rely directly on it. However, the flaw is identifiable and localizable, and the central approach appears salvageable by redefining the invariant as the set of eigenvalues with nonzero coefficient α_λ. I recommend major revision rather than rejection. If the authors cannot supply a correct invariant and complete the missing Theorem 4 minimality arguments, the paper should not be published in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, read this one if you want a clean example of a spectral invariant claim that fails at the boundary case. The paper claims the set of distinct nonzero eigenvalues of the transfer operator is an invariant of the generated process, and uses that to bound quantum and classical generative complexity. The genuinely new pieces are the spectral bound, the quadratic separation between SIO-based and general quantum bounds, and an explicit three-state example with a two-dimensional quantum memory. The example is plausible, and the coherence-resource framing is a clean way to state the advantage. The tensor-network reformulation of QHMMs is standard but applied cleanly, and the authors flag the diagonalisability caveat.\n\nThe soft spot is load-bearing. Theorem 1 is false as stated. The appendix proof itself concedes that two models can have different spectra 'unless α_λ = 0'. The Vandermonde argument only forces coefficients to match for eigenvalues present in both models; it does not rule out extra eigenvalues with zero coefficient. The concrete counterexample is the fair-coin process: a one-state HMM gives transfer-operator eigenvalues {1/2}, while a two-state model with a never-occupied second state adds a decoupled eigenvalue 0.58 with α=0, leaving the process unchanged. So Λ is not a process invariant, and the bounds in Theorems 2 and 3, taken from any generating model, are invalid. The fix is conceptually straightforward: restrict Λ to eigenvalues with α≠0, the contributing spectrum. The bounds and the quadratic advantage may survive that restriction, but the paper as written does not establish them.\n\nA smaller concern: the SIO rank argument in Theorem 3 depends on a specific Kraus representation of a classical HMM as SIOs; phase degrees of freedom mean the sparsity pattern is representation-dependent. That is minor next to the main flaw. The citation pattern is acceptable but misses classical realization theory (Ho-Kalman), which would have flagged the zero-coefficient issue.\n\nBottom line: promising but technically broken. The example and the contributing-spectrum idea are worth salvaging; researchers in quantum stochastic process simulation will want to see the corrected version. I would send it to peer review with a heavy-revision expectation, not desk reject it, but as is it should not be accepted.","headline":"The spectral invariant at the core of this paper is false as stated; the paper is worth a referee only if the authors restrict to contributing eigenvalues.","tokens_in":14868,"tokens_out":3030,"would_cite":false,"duration_ms":26477,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P45","60J10"],"pacs":["03.67.-a","02.50.Ga"],"model":"deepseek-v4-flash","headline":"The spectrum of a process's transfer operator is the same in every generating quantum hidden Markov model, and it fixes a lower bound on the memory needed to generate the process.","keywords":["quantum hidden Markov models","stochastic processes","generative complexity","transfer operator","spectral invariants","quantum coherence","strictly incoherent operations","memory-minimal models"],"falsifier":"Compute the distinct nonzero spectra of the vectorised transfer operators of two QHMMs that generate the same stochastic process, counting only eigenvalues whose $\\alpha_\\lambda$ is nonzero; if the two sets differ, Theorem 1 fails. A more direct test is to append a decoupled two-level classical sector to a known minimal QHMM: the visible process is unchanged, and checking whether the resulting $|\\Lambda|$ changes reveals whether the invariant needs the nonzero-coefficient caveat.","tokens_in":13781,"feed_emoji":"⚛️","tokens_out":10030,"duration_ms":82787,"temperature":0.7,"pith_summary":"The paper aims to solve part of the problem of telling how much hidden memory a stochastic process really needs: infinitely many quantum hidden Markov models can produce the same observed statistics, and before this paper there was no known feature of the process itself that all generating models must share. It identifies such a feature: the set of distinct nonzero eigenvalues of the transfer operator built from any generating model is the same for every model of that process. From this invariant it derives a lower bound on the quantum generative topological complexity, and shows that forbidding quantum coherence raises the bound quadratically, so any model that beats the classical bound must use coherence. This matters because it turns a seemingly uncomputable minimization over infinitely many models into a bound that can be read off from one model, and it gives a precise sense in which quantum memory advantages are real.","feed_headline":"Spectrum of a transfer operator sets the minimal memory","feed_subtitle":"Classical hidden Markov models need quadratically more memory unless quantum coherence is allowed","key_machinery":"The transfer operator $E_{AB}=\\sum_x \\mathcal{E}^x_A\\otimes\\mathcal{E}^x_B$ and its vectorised form carry the argument. Contracting the outputs of two copies of the instrument and taking the $L$-th power makes the probability of every length-$L$ word equal to $\\langle\\!\\langle 1|E_{AB}^L|\\rho_{AB}\\rangle\\!\\rangle$, so two models generate the same process exactly when these traces agree for all $L$. The distinct nonzero eigenvalues of $E_{AB}$ then become a fingerprint of the process rather than of any particular model: equality for all $L$ forces the spectra of any two generating models to coincide. The dimension counting that produces the bounds is simply that an $n$-dimensional memory makes the vectorised transfer operator an $n^4\\times n^4$ matrix, while a classical strictly incoherent model makes that matrix sparse with rank at most $n^2$.","core_discovery":"The paper's central claim is that every quantum hidden Markov model that generates a given stochastic process shares one spectral invariant: the set of distinct nonzero eigenvalues $\\Lambda$ of the vectorised transfer operator $E_{AB}=\\sum_x \\mathcal{E}^x_A\\otimes \\mathcal{E}^x_B$, together with the coefficients $\\alpha_\\lambda=\\langle\\!\\langle 1|\\Pi_\\lambda|\\rho\\rangle\\!\\rangle$ for shared eigenvalues. Because the transfer operator is an $n^4\\times n^4$ matrix when the memory has dimension $n$, a process whose invariant spectrum has $|\\Lambda|$ distinct nonzero eigenvalues cannot be generated in fewer than $\\log\\lceil |\\Lambda|^{1/4}\\rceil$ memory units, giving Theorem 2. When the generating operations are restricted to strictly incoherent operations, the quantum description of a classical hidden Markov model, the same invariant requires $\\log\\lceil |\\Lambda|^{1/2}\\rceil$ memory, a quadratically larger bound. Since the quantum bound is weaker, the gap between the two bounds is a resource-theoretic signature of coherence, and Theorem 4 exhibits a three-state classical process whose emissions fit in a two-dimensional quantum memory, proving $c_Q