{"id":"c61d73ab-169e-4a29-94d2-836fa7f43375","arxiv_id":"2507.21337","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors propose quantum and quantum-inspired hidden Markov model estimators for stochastic volatility and claim a quadratic reduction in hidden states, but the theoretical proof and likelihood formulas contain serious gaps.","lead":"This paper tries to estimate stock market volatility models using quantum and quantum-inspired hidden Markov models, claiming that a quantum model needs far fewer hidden states to match a classical one. A generalist might read it to see whether quantum computing offers any real advantage in finance beyond the current hype.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1 is not established: the proof equates KL divergence after arbitrary invertible Hankel transformations, which does not preserve KL, and rank equality alone does not imply KL-equidivergence.","rationale":"The reader rejected the paper with high correctness risk, and the reader's rationale includes the statement that the proof of Theorem 4.1 relies on an unjustified assertion about KL-equidivergence via Hankel matrix transformations. I agree with that criticism, but I would locate the load-bearing failure one step earlier and more sharply: even if the likelihood functions in equations (2.7) and (2.8) were correctly specified, the algebra in Appendix A.3.3 does not prove that a QHMM with sqrt(n_c) hidden states is KL-equidivergent to a classical HMM with n_c states. The proof only compares identity blocks after model-dependent invertible transformations; that is not a statement about the distributions the Hankel matrices encode. Since Theorem 4.1 is the basis for the claimed quadratic speedup and for the KL-equidivergent model set used in Theorem 6.1, the central claim is unsupported. The Assumption A.14 concern raised by the reader is real but secondary: it only matters after the existence of the KL-equidivergent quantum model is established, and the proof of that existence is the weak point. For these reasons the rejection verdict is appropriate and no adjustment is needed.","tokens_in":35074,"tokens_out":5679,"duration_ms":65587,"concrete_test":"Take the 4-state nonparametric classical HMM and the 2-state QHMM used in Figure 4(a). Enumerate all length-4 observation strings, and compute the exact KL divergence between the classical model's string distribution and the QHMM's string distribution, using equation (3.3) for the QHMM probabilities. If the KL divergence is positive, the two models are not KL-equidivergent and Theorem 4.1's construction fails. Alternatively, check equation (A.8) symbolically: because C and D are arbitrary invertible matrices applied separately to H_c and H_q, the left-hand side vanishes even when H_c != H_q, confirming that the proof reduces to a rank comparison rather than a distributional equivalence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim of quadratic state reduction with equal KL divergence rests on Theorem 4.1, whose proof in Appendix A.3.3 is invalid. Theorem A.2 only gives rank(H_Q) <= (n_q)^2 for any QHMM; it does not state that every classical HMM of rank n_c = n^2 can be realized by a QHMM with n_q = n hidden states generating the same process. The proof then takes finite Hankel matrices H_c(d) and H_q(d), writes canonical decompositions H_c = C_c diag(I_{n_c},0) D_c and H_q = C_q diag(I_{n_q^2},0) D_q, and claims D_KL((C_c)^{-1} H_c (D_c)^{-1} || (C_q)^{-1} H_q (D_q)^{-1}) = 0. Since (n_q)^2 = n_c, both transformed matrices are the same identity-block matrix, so the equality is a tautology. The KL divergence in equation (A.7) is not invariant under arbitrary invertible row/column transformations, and the transformed matrices are not probability matrices. Equality of rank therefore does not imply equality of the induced stochastic processes or of D*_KL. Absent a proof that the constructed QHMM has the same finite-dimensional distributions, Theorem 4.1, Theorem 4.2, and the non-asymptotic comparison built on the same KL-equidivergent set are unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes classical and quantum hidden Markov model approximations for stochastic volatility diffusions. Its central theoretical claims are (i) a classical HMM with n_c = n^2 hidden states can be replaced, without loss of KL divergence, by a quantum HMM with n_q = n hidden states (Theorem 4.1), yielding a quadratic state reduction; (ii) this implies an asymptotic KL comparison favoring the quantum model (Theorem 4.2); and (iii) non-asymptotic bounds from Lehéricy (2021) are tighter for the KL-equidivergent quantum HMM (Theorem 6.1). The paper also reports likelihood-based estimation, filtering, and simulation comparisons on S&P 500 data. The manuscript contains a number of self-identified limitations, including the lack of a fully developed quantum likelihood-ratio theory and the provisional nature of Assumption A.14, but these are presented as caveats rather than as obstacles to the main claims.","tokens_in":35410,"tokens_out":4769,"duration_ms":63230,"significance":"If the central claims were valid, the quadratic reduction in hidden-state dimension and the tighter non-asymptotic bounds would be of substantial interest to both quantum computing and financial econometrics. The paper also has genuine strengths: it provides a concrete ansatz-circuit QHMM construction, a reproducible Python implementation, and a clear focus on classical statistical inference rather than Bayesian MCMC. However, the central theoretical results are not established. The likelihood functions are mis-specified, the proof of the main dimensionality-reduction theorem relies on an invalid invariance argument, and the non-asymptotic comparison is carried by an unproved assumption that transfers classical concentration constants to the quantum model. The empirical simulations cannot compensate for these defects because they use the same incorrect likelihoods.","major_comments":[{"comment":"The log-likelihood is written with the logarithm inside the double summation: ℓ_t = Σ_i Σ_j log[P_y(...) g(...) X_{t-1}^i]. This is not the log of the marginal likelihood of an HMM, which would be log Σ_i Σ_j P_y(...) g(...) X_{t-1}^i. The displayed expression is a weighted average of log terms, not a log-likelihood. Since the paper's KL-divergence measures, MLE, and likelihood-ratio simulations in Sections 4 and 5 all use this object, the error is load-bearing and invalidates the estimation and comparison results that follow.","section":"Section 2.4, Eqs. (2.7) and (2.8)"},{"comment":"The proof of Theorem 4.1 does not establish KL-equidivergence. Theorem A.2 gives only an upper bound, rank(H_Q) ≤ (n_q)^2, for any QHMM; it does not assert that for every classical HMM of rank n_c = n^2 there exists a QHMM with n_q = n hidden states generating the same finite-dimensional distributions. The proof then defines canonical decompositions and claims D_KL((C_c)^{-1}H_c(D_c)^{-1} || (C_q)^{-1}H_q(D_q)^{-1}) = 0 because both transformed matrices have the same identity-block form. This is a tautology about rank, not a statement about the stochastic processes. KL divergence is not invariant under arbitrary invertible row/column transformations, the transformed matrices are not nonnegative probability matrices, and equality of rank does not imply equality of the Hankel entries or of D*_KL. Consequently, Theorems 4.1, 4.2, and the non-asymptotic comparison in Section 6 are unsupported.","section":"Appendix A.3.3, Theorem 4.1 and Theorem A.2"},{"comment":"Assumption A.14 asserts that the constants A, C_Λ, and T0 from Lehéricy's non-asymptotic bounds for the classical HMM also apply to the KL-equidivergent quantum HMM. This assumption is load-bearing for Theorem 6.1 and Corollary 6.1, which claim tighter quantum bounds, but it is stated without proof. The accompanying remark that the assumption is 'reasonable' in light of Theorems A.2 and 4.1 is not a derivation, and Theorem 4.1 is itself unproved. The tighter quantum bound is therefore obtained by postulating exactly the concentration behavior that needs to be demonstrated.","section":"Appendix A.1.2, Assumption A.14 and Theorem 6.1"},{"comment":"Lemma A.3 claims that if two models M1 and M2 are KL-equidivergent, then E*[log L(M1)/L(M2)] = 0, citing the Pythagorean identity. Equality of D*_KL(M1) and D*_KL(M2) does not, by itself, imply that the cross-entropy or cross-KL between M1 and M2 vanishes; the Pythagorean identity holds only for specific projection geometries, not for arbitrary pairs with equal divergence to a third distribution. The proof of Theorem 4.2 relies on this step, so the asymptotic comparison D*_KL(M_nq) ≤ D*_KL(˜M_nc) is not established even conditional on Theorem 4.1.","section":"Appendix A.3.2, Lemma A.3 and proof of Theorem 4.2"}],"minor_comments":[{"comment":"There are numerous typographical errors and inconsistent notations, e.g., 'Obsersable states' in Definition 2.2, 'entires' in Definition 2.4, 'distribtuions' in Section 5.2, and the title 'V olatility' on the first page. These should be corrected carefully.","section":"Throughout"},{"comment":"The caption and text refer to 'Log Likelihood Ratio' but the horizontal axis is labeled log10(L_i/L_j); please clarify whether the plotted quantity is the natural-log likelihood ratio or its base-10 logarithm, and state this consistently in the text.","section":"Section 5.4, Figure 4"},{"comment":"Assumption A.14 is introduced in a regularity-conditions section without any discussion of how the quantum HMM satisfies the underlying mixing or forgetting conditions that the classical constants encode; even as a provisional assumption, this deserves explicit justification or a reference.","section":"Appendix A.1.2, Assumption A.14"}],"recommendation":"reject","confidential_remarks":"The manuscript's central theoretical contribution rests on an invalid proof of Theorem 4.1, a mis-specified likelihood function, and an unproved transfer-of-constants assumption for the non-asymptotic bounds. These are load-bearing errors that cannot be fixed by local revision; the paper would need a fundamentally different theoretical foundation. A rejection is recommended."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the application is genuinely new: the authors map stochastic volatility diffusions onto classical and quantum HMMs, run simulation-based comparisons, ship code, and are honest about hardware limits. Second, the central theoretical claim—quadratic reduction in hidden states with equal KL divergence—is not established. Theorem 4.1's proof uses invertible canonical decompositions of finite Hankel matrices and calls the transformed matrices KL-equidivergent because they become the same identity-block matrix. That is a tautology. KL divergence is not preserved by arbitrary invertible row/column transformations, and rank equality alone does not imply the two models generate the same stochastic process. The authors cite Markov et al.'s bound rank(H_Q) ≤ (n_q)^2, but that is an upper bound, not a realization theorem showing every classical HMM with n_c states has a QHMM with sqrt(n_c) states producing the same finite-dimensional distributions. Without that, Theorem 4.1, Theorem 4.2, and the non-asymptotic comparison resting on the same KL-equidivergent set are unsupported.\n\nThe likelihood in equations (2.7) and (2.8) also places the logarithm inside the sum over hidden states. That is not the log of the marginal likelihood of an HMM; log of a mixture is not a mixture of logs. If the intent is an approximate or quasi-likelihood, the paper needs to say so explicitly and re-derive the consistency and bound arguments. Assumption A.14 then transfers the constants from Lehéricy's classical non-asymptotic bounds to the equidivergent quantum model without proof, and since the equidivergence is not established, the comparison in Theorem 6.1 is resting on an assumption that effectively enforces the conclusion.\n\nWhat is worth keeping: the literature engagement is serious, the caveats about quantum likelihood ratios and missing q-QMLE theory are stated plainly, and the empirical work is a reasonable proof of concept. The empirical comparison is not a test of the theory though—different parameter counts between the QHMM and classical HMM make the likelihood ratios hard to interpret.\n\nThe paper deserves a serious referee because the application is real and the repair may be possible, but the likelihood and the proof of Theorem 4.1 need to be reworked before the main claims can stand.","headline":"The paper's genuinely new application—stochastic volatility estimation via quantum hidden Markov models—comes with a misspecified likelihood and a central theorem whose proof relies on a tautology, so the main claims are not yet supported.","tokens_in":35930,"tokens_out":3588,"would_cite":false,"duration_ms":47692,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M05","81P68","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A quantum hidden Markov model can match a classical one using only the square root of the classical number of hidden states.","keywords":["quantum hidden Markov models","stochastic volatility","maximum likelihood estimation","Kullback-Leibler divergence","dimensionality reduction","non-asymptotic bounds","volatility filtering","CIR diffusion"],"falsifier":"Run a long simulation ($T=10{,}000$) from a known CIR stochastic volatility process, fit a classical HMM with 16 hidden states and a KL-equidivergent quantum HMM with 4 hidden states, and estimate each model's KL divergence from the true density. Theorem 4.2 predicts the quantum divergence is no larger; repeated samples in which the quantum model is farther from the truth would refute the central claim. A direct check of whether the concentration constants in Assumption A.14 can actually be derived for the quantum channel would settle the tighter-bound claim.","tokens_in":34857,"feed_emoji":"📈","tokens_out":7129,"duration_ms":80678,"temperature":0.7,"pith_summary":"This paper tries to establish that stochastic volatility, a latent process that makes maximum likelihood estimation hard, can be estimated and filtered through hidden Markov models built for quantum hardware, and that the quantum version needs far fewer hidden states to reach the same accuracy. The central result is a dimension reduction: a classical HMM with $n^2$ hidden volatility states can be matched, in Kullback-Leibler divergence to the true data-generating process, by a quantum HMM with only $n$ hidden quantum states. The paper also claims that non-asymptotic error bounds for the quantum model are tighter than for the classical model, and supports this with simulations on S&P 500 returns. If true, this gives econometricians a quasi-closed-form likelihood and volatility filter for stochastic volatility, with a quadratic state reduction that becomes a runtime advantage whenever the number of states dominates the sample length.","feed_headline":"Quantum model matches 16-state volatility HMM with 4 states","feed_subtitle":"A quadratic state reduction plus tighter error bounds could make volatility estimation and filtering practical on quantum hardware.","key_machinery":"The machinery is the generalized Hankel matrix of a stochastic process and the rank bound for unitary quantum hidden Markov models. A generalized Hankel matrix records probabilities of finite observation strings, and its rank controls the minimal number of hidden states a hidden Markov model needs. For a classical HMM of order $n_c$, Anderson's realization theory identifies the model with a bi-infinite Hankel matrix of rank $n_c$, and a minimal realization of that rank exists. A unitary quantum HMM with $n_q$ hidden states has Hankel rank at most $(n_q)^2$, which delivers the quadratic reduction when $(n_q)^2 = n_c$. KL-equidivergence is then obtained by comparing canonical decompositions of the two Hankel matrices and invoking a Pythagorean identity for KL divergence. The non-asymptotic comparison runs through a cited penalized-MLE bound for misspecified HMMs, with the quantum model evaluated at the reduced state count.","core_discovery":"The paper's central claim is that quantum hidden Markov models dominate classical ones in state count without sacrificing approximation quality. For every classical HMM of order $n_c = n^2$, Theorem 4.1 asserts there exists at least one KL-equidivergent unitary quantum HMM with $n_q = n$ hidden states: the two models have the same minimal Kullback-Leibler divergence from the true stochastic volatility data-generating process. Theorem 4.2 strengthens this asymptotically, stating that for $n_q = \\sqrt{n_c}$ the quantum model's KL divergence is no larger than the classical model's, so the expected log-likelihood ratio favors the quantum model. Theorem 6.1 then claims that, under transferability of the constants in the non-asymptotic bounds, the KL-equidivergent quantum HMM has a tighter probability bound on parameter estimation and filtered volatility error than the classical penalized MLE. The empirical section reports that a 4-state quantum HMM beats or matches 16-state classical HMMs in almost all of 1000 simulated trials, even though the classical parametric model is matched to the data-generating process.","pith_inferences":["The quadratic state reduction is proved through Hankel-rank arguments that use no detail specific to stochastic volatility, so the same KL-equidivergence claim should transfer to any latent-state time series model admitting a finite-rank HMM approximation; testing this on regime-switching or discrete-time stochastic volatility models would be a natural extension.","Because the numerical comparison shows the quantum model winning most often at $T=500$, the asymptotic result may hold well before the large-sample limit; a finite-sample version of Theorem 4.2 giving a rate for the expected log-likelihood ratio would be a testable strengthening the paper leaves implicit.","The runtime analysis shows an exponential speedup in hidden state count offset by an exponential slowdown in sample length, so the practical window for the quantum advantage is short samples with many states; shrinking the $T$-dependence, for example by quantum amplitude estimation of the likelihood, would widen that window."],"forward_implications":["A practitioner who needs $n^2$ classical volatility states can instead use $n$ quantum states and keep the same KL accuracy, cutting the state dimension quadratically.","The likelihood and volatility filter become quasi-closed-form expressions in Kraus operators and partial traces, avoiding numerical integration and MCMC accept-reject loops.","In large samples the quantum model is expected to fit the true DGP at least as well as the classical model with the square of its state count, and the simulations at $T=500$ show this mostly already holds.","The quantum filter inherits a tighter non-asymptotic high-probability bound on the error of estimated parameters and filtered integrated volatility, provided the concentration constants transfer.","On quantum hardware the training complexity can become polylogarithmic in the hidden state count, though it grows exponentially in the number of time steps."],"supporting_citations":[{"why":"Supplies the representation of a continuous-time stochastic volatility diffusion as a hidden Markov model, which the paper's discretization builds on.","marker":"Genon-Catalot, Jeantheau, and Larédo (2000)"},{"why":"Defines the rank of an HMM through its generalized Hankel matrix, the quantity Theorem 4.1 exploits.","marker":"Anderson (1999)"},{"why":"Shows a minimal HMM can have rank equal to its order, justifying the equality used in the dimension-reduction argument.","marker":"Huang, Ge, Kakade, and Dahleh (2014)"},{"why":"Provides the uniqueness of the minimal Moore representation and the KL-divergence/Hankel framework used in the proof of Theorem 4.1.","marker":"Vanluyten, Willems, and De Moor (2008)"},{"why":"States that every classical HMM has a same-order QHMM generating the same stochastic process, a base ingredient for the quantum equivalence.","marker":"Monras, Beige, and Wiesner (2011)"},{"why":"Defines the unitary quantum HMM and proves the Hankel rank bound $\\mathrm{rank}(H_Q) \\leq (n_q^L)^2$ that carries the quadratic reduction.","marker":"Markov, Rastunkov, Deshmukh, Fry, and Stefanski (2022)"},{"why":"Provides the non-asymptotic MLE bounds for misspecified HMMs that Theorem 6.1 transfers to the quantum setting.","marker":"Lehéricy (2021)"},{"why":"Supplies the Pythagorean identity for KL divergence used to prove KL-equidivergence and the likelihood-ratio comparison in Theorem 4.2.","marker":"Csiszár (1975)"},{"why":"Supports the asymptotic convergence of misspecified HMM estimators behind Assumption 4.1.","marker":"Mevel and Finesso (2004)"}],"fun_headline_variants":["Quantum HMM: 4 states match 16 classical for volatility","4-state quantum HMM beats 16-state classical in volatility","Volatility: quantum HMM cuts states 16 to 4 with tighter bounds","Quantum HMM square-root state reduction for volatility MLE","Tighter bounds with 4-state quantum HMM for volatility"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Assumption A.14 asserts, without proof, that the concentration constants for the classical HMM non-asymptotic bounds also apply to the KL-equidivergent quantum HMM; if that transfer fails, Theorem 6.1's tighter quantum bound does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Quantum HMM: 4 states match 16 classical for volatility","4-state quantum HMM beats 16-state classical in volatility","Volatility: quantum HMM cuts states 16 to 4 with tighter bounds","Quantum HMM square-root state reduction for volatility MLE","Tighter bounds with 4-state quantum HMM for volatility"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2823,"prompt_tokens":854,"completion_tokens":1969,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":1889}},"tokens_in":470,"tokens_out":1969,"duration_ms":16600,"temperature":1.0,"reasoning_tokens":1889,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:52:12.642004+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a long simulation ($T=10{,}000$) from a known CIR stochastic volatility process, fit a classical HMM with 16 hidden states and a KL-equidivergent quantum HMM with 4 hidden states, and estimate each model's KL divergence from the true density. Theorem 4.2 predicts the quantum divergence is no larger; repeated samples in which the quantum model is farther from the truth would refute the central claim. A direct check of whether the concentration constants in Assumption A.14 can actually be derived for the quantum channel would settle the tighter-bound claim.","supporting_citations":[],"review_version":1}