{"id":"3770689a-82e0-49fa-956c-c57aa5bfc298","arxiv_id":"1909.01044","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper adapts slow feature analysis to gate parameters, claiming it stabilizes an optimal quantum state, but the claim lacks a proof and the evaluation is disconnected from the algorithm.","lead":"This paper proposes an unsupervised learning method to stabilize the optimal quantum state of a gate-model quantum computer across multiple runs. It also offers a stability-classification scheme, but the derivation does not show how the proposed smoothing preserves the objective function, and the numerical examples do not test the method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (9) is asserted, not derived: minimizing F* in Eq. (22) never involves the cost operator C, so beta = S^T alpha is not shown to preserve f(theta*); a generic orthogonal mixing of optimal parameter vectors yields non-optimal states.","rationale":"The reader's weakest assumption identified exactly the missing link: F* does not contain C or f, so nothing enforces f(phi)=f(theta*). My stress test confirms this is not just a missing detail but a false implication: there exist orthogonal S with the same F* value that map an optimal parameter vector to a non-optimal one. The numerical evaluation in Section 5 does not test Algorithm 1 or Theorem 1; it assumes f_D is sinusoidal and computes delta=1/N, so it provides no independent support. Because the central claim is unsupported and in fact contradicted by a simple two-parameter counterexample, the REJECT verdict stands. No change to the reader's verdict is needed.","tokens_in":14393,"tokens_out":20601,"duration_ms":207102,"concrete_test":"Analytic check: verify the counterexample above. Set L=2, R=2, U_i(theta)=exp(-i theta Y), input |00>, C=|++><++|, f(theta)=((1+sin theta_1)(1+sin theta_2))/4. Choose alpha = [(pi/2,pi/2), (pi/2,pi/2+delta)] with delta=0.1, and S = 1/sqrt(2) [[1,-1],[1,1]]. Compute beta=S^T alpha and evaluate f(beta_1). The result is f(beta_1)=(1+sin(pi/sqrt(2)))/4 ~ 0.45 while f(theta*_1)=1; simultaneously, Delta-beta and Delta-alpha have identical squared norms and the tau term vanishes for R=2, so the C-independent quantity minimized in Eq. (22) is the same for S and for the identity. This directly falsifies the claimed implication; a corrected theorem would need an additional constraint involving C.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion is Eq. (9): the stabilized parameters phi must satisfy f(phi)=f(theta*). The proof of Theorem 1 never derives this. The S matrix is found by minimizing F* (Eq. 22), which depends only on Delta-alpha and the weight coefficients omega_rs; neither F* nor the generalized eigenvalue equation (28) contains the cost operator C or the function f. Because beta = S^T alpha, each stabilized column is an orthogonal linear combination of the alpha columns, and orthogonal combinations of optimal gate parameters are not generally optimal. A concrete instance: with L=2, U_i(theta)=exp(-i theta Y), input |00>, C=|++><++|, |+>=(|0>+|1>)/sqrt(2), we get f(theta)=((1+sin theta_1)(1+sin theta_2))/4, whose unique maximum on [0,pi]^2 is theta*=(pi/2,pi/2), f=1. Let alpha have columns theta*_1=(pi/2,pi/2) and theta*_2=(pi/2,pi/2+delta), and let S be the 45-degree rotation. S is orthogonal, beta_1=S^T theta*_1=(pi/sqrt(2),0), and f(beta_1)=(1+sin(pi/sqrt(2)))/4 ~ 0.45, not 1. Yet Delta-beta=S^T Delta-alpha has the same squared norm as Delta-alpha, and the tau term contributes nothing for R=2, so no term in F* distinguishes S from the identity. The minimization cannot rule out an S that destroys optimality. Separately, the S from Eq. (28) is not shown to obey S^T S=I; the generalized eigenproblem imposes S^T(Delta-alpha eta Delta-alpha^T)S=I, not Eq. (13). The numerical section does not exercise Algorithm 1. The theorem therefore rests on an unproven invariance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method to stabilize the optimal state of a gate-model quantum computer. Gate parameter vectors θ*_r for R running sequences are collected into a matrix α, and the stabilized parameters are obtained as β = Sᵀα, where S is claimed to be orthogonal. Theorem 1 asserts that S can be found by minimizing an objective function F* (Eq. 22) that penalizes temporal variation of the gate parameters. Theorem 2 proposes a kernel-based classification of the stabilized sequences into stability classes. Section 5 evaluates the relative entropy between stabilized and target parameters under an assumed sinusoidal model and computes a correlation coefficient between toy gate-parameter functions. The central claim is that the stabilized state |φ⟩ satisfies f(φ)=f(θ*), preserving optimality across all running sequences.","tokens_in":14914,"tokens_out":3256,"duration_ms":33993,"significance":"If the central claim were established, the idea of preserving optimality while stabilizing gate parameters through an orthogonal transformation would be a valuable contribution to near-term gate-model quantum computation, particularly combined with unsupervised learning. However, the manuscript does not prove the central invariance, and its numerical evaluation does not exercise the proposed algorithms. The paper provides no machine-checked proofs or reproducible code, and the only quantitative results are derived from an assumed sinusoidal model whose parameters are chosen ad hoc. Consequently, the work, as presented, does not establish any falsifiable prediction or validated method, and its significance is currently more suggestive than substantive.","major_comments":[{"comment":"The central claim of Theorem 1, that stabilizing via β=Sᵀα preserves f(φ)=f(θ*) as required by Eq. (9), is never proven and is in fact false in general. The objective function F* in Eq. (22) contains only the temporal-difference term χ and the weighted term cτ; neither term involves the cost operator C or the objective function f. An orthogonal linear combination of optimal parameter vectors need not be optimal. A concrete counterexample is given by L=2, U_i(θ)=exp(-iθY), input |00>, and C=|++⟩⟨++| with |+⟩=(|0⟩+|1⟩)/√2. Then f(θ)=((1+sin θ1)(1+sin θ2))/4, whose unique maximum on [0,π]² is θ*=(π/2,π/2). Taking α columns θ*_1=(π/2,π/2) and θ*_2=(π/2,π/2+δ), the 45-degree rotation S gives β_1=Sᵀθ*_1=(π/√2,0), and f(β_1)=(1+sin(π/√2))/4≈0.45, not 1. Since F* depends only on ∆α and the weights ω_rs, this S and the identity are indistinguishable to F*, so the minimization cannot rule out a stabilizer that destroys optimality.","section":"§3, Eq. (22)"},{"comment":"The S matrix obtained from the generalized eigenvalue problem in Eq. (28) is not shown to satisfy the orthogonality condition SᵀS=I imposed in Eq. (13). The constraint that is actually imposed is (∆β)ᵀη∆β=I, i.e., Sᵀ(∆αη∆αᵀ)S=I, and η itself depends on α through the weight coefficients (20). Even if a solution to Eq. (28) exists, it is an orthogonal transformation with respect to the matrix ∆αη∆αᵀ, not a Euclidean orthogonal matrix, and therefore β=Sᵀα does not necessarily satisfy the stated stabilization relation. The proof of Theorem 1 needs to establish that the eigenvector matrix can be chosen so that SᵀS=I, and this is not done.","section":"§3, Eqs. (24), (28)"},{"comment":"The numerical evaluation in Section 5 does not implement or simulate Algorithm 1 (AS) or Algorithm 2 (AC). Instead, the stability parameter δ(r) is derived entirely from the assumed sinusoidal model (60) for f_D(β‖β*), with parameters γ, λ, N, c, and E(D) chosen ad hoc. Equation (64) then yields δ(r)=1/N, which is an inverse of the number of oscillations of the assumed sinusoidal function; it is a property of that assumed model, not of the proposed stabilization algorithm. The correlation coefficient µ(β,β*) in §5.2 is computed for arbitrary functions f(φ_r) defined in Eqs. (67) and (69) and is never connected to the output β of Algorithm 1. Consequently, Section 5 provides no numerical evidence that the proposed method achieves state stabilization or preserves optimality.","section":"§5, Eqs. (60)–(64)"},{"comment":"The classification procedure in Theorem 2 and Algorithm 2 is not operationally defined. The classifier f_C^k in Eq. (39) is asserted to output probabilities, but no training procedure, loss function, or connection to the stabilized data β from Algorithm 1 is given. Moreover, Step 7 defines C(φ_r)=ξ(φ_r)C_p+ℓ_k(φ_r)C_q, but the addition of stability classes C_p and C_q, and multiplication by scalar parameters, is not defined. It is unclear whether the output is a class label, a vector, or a fuzzy membership vector. This makes the classification claim unfalsifiable as stated.","section":"§4, Algorithm 2, Step 7"}],"minor_comments":[{"comment":"The text says 'where C is the classical value of the objective function', but Eq. (3) uses C as an operator in the expectation value ⟨θ*|C|θ*⟩. This is inconsistent and should be clarified.","section":"§2, Eq. (3)"},{"comment":"The weight coefficient ω_rs depends on the squared norm of differences of ∆θ* vectors, but the manuscript does not specify the norm or explain how the kernel width ζ is chosen. The same symbol ζ is also used later; please disambiguate.","section":"§3, Eq. (20)"},{"comment":"The constant c in Eq. (60) is reused for the amplitude of the sinusoidal model, while c was already used as the regularization constant in Eq. (22) and as the kernel bandwidth in Eq. (47). This triplication of the symbol c is confusing.","section":"§5, Eq. (60)"},{"comment":"There are several typographical errors, including 'uniaries' and 'quatum', and some equations have inconsistent bracket formatting (e.g., Eq. (2) and Eq. (7)). These should be corrected.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The manuscript's core claim is not established: the objective function used to determine the stabilizer matrix does not involve the cost operator, and the orthogonality of S is not proven. The numerical section never runs the proposed algorithms. These are load-bearing issues that cannot be fixed by modest revision; they require a fundamental reworking or substantial new theoretical results and numerical experiments. In addition, the novelty relative to existing slow feature analysis literature is not clearly delineated, though I do not base the recommendation on this concern alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the central theorem is unproven and false as stated. The paper takes slow feature analysis (minimize temporal variance under an orthonormal constraint, refs [32,33]) and calls it state stabilization, but it never shows the stabilized parameters preserve the objective function. A concrete L=2 counterexample (rotations about Y, C=|++><++|) gives f(theta*)=1 at (pi/2,pi/2), while a 45-degree orthogonal mixing yields f≈0.45; the F* objective cannot distinguish the bad rotation from the identity, so the claimed invariance fails. The proof never touches the cost operator C, and the generalized eigenproblem does not enforce S^T S=I either.\n\nWhat the paper does well: it identifies a legitimate engineering problem—keeping a near-term gate-model computation near an optimal state across runs—and it cites the SFA literature rather than hiding the connection. That is honest, but it also means the technical core is not new.\n\nSoft spots, in order of severity: (1) the numerical section never runs Algorithm 1 or Algorithm 2; it fits a sinusoidal model and concludes δ=1/N, which is an artifact of the chosen formula, not a property of the method. (2) The classification procedure is underdefined: no training labels, no concrete classifier, circular definitions of classes. (3) The reference list is padded with self-citations and generic surveys, which is secondary but doesn't help.\n\nThe citation pattern aside, the flaws are load-bearing. A referee would find the counterexample in under an hour. There is no code, no data, no formal proof, and no external check. I would not cite this, and I wouldn't bring it to reading group except as a case study in how a plausible-sounding theorem can fail. My recommendation: desk reject rather than spend a full review cycle; if you want a documented rejection, one referee with a two-page counterexample is enough.","headline":"Central theorem is asserted, not proven, and fails on a two-qubit example; the paper repackages slow feature analysis without a valid result.","tokens_in":15362,"tokens_out":4158,"would_cite":false,"duration_ms":42649,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68"],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"The paper claims that an orthogonal stabilizer matrix, found by minimizing temporal variation in gate parameters, can hold a gate-model quantum computer in its optimal state across arbitrarily many running sequences.","keywords":["quantum state stabilization","gate-model quantum computers","gate parameter optimization","unsupervised learning","stability classes","slow feature analysis","orthogonal transformation","near-term quantum computing"],"falsifier":"Take a one-qubit circuit $U(\\theta)=e^{-i\\theta X}$ with objective $C=Z$ and two runs whose optimal parameters make the minimizer of $F^*$ nontrivial; compute the stabilized parameters $\\phi=S^T\\theta$, then compare $\\langle\\phi_r|Z|\\phi_r\\rangle$ with $\\langle\\theta_r|Z|\\theta_r\\rangle$ for each $r$. A single mismatch would show that the stabilized state is not optimal, contradicting the paper's central claim.","tokens_in":14177,"feed_emoji":"⚛️","tokens_out":9015,"duration_ms":78429,"temperature":0.7,"pith_summary":"This paper tries to establish that the optimal state of a gate-model quantum computer can be stabilized across an arbitrary number of running sequences by an orthogonal transformation of the gate parameters. The authors define a stabilizer matrix $S$ that maps the sequence of optimal parameter vectors $\\alpha$ to a stabilized sequence $\\beta=S^T\\alpha$, and they propose finding $S$ by minimizing the temporal variation between consecutive parameter vectors. They further claim that the stabilized state preserves the objective value of the optimal state, $f(\\phi)=f(\\theta^*)$, and that the stabilized sequences can be sorted into stability classes by unsupervised learning. A reader might care because holding a near-term quantum computer near its optimal operating point over many runs is a practical obstacle that this procedure addresses purely classically.","feed_headline":"Stabilizer matrix keeps optimal quantum state fixed across runs","feed_subtitle":"Orthogonal gate-parameter transform, learned without supervision, tames run-to-run drift in near-term quantum computers.","key_machinery":"The central object is the stabilizer matrix $S$, an orthogonal transformation that is supposed to convert the matrix of optimal gate parameters $\\alpha$ into the matrix of stabilized parameters $\\beta=S^T\\alpha$ without changing the objective function value. The argument is carried by the objective $F^*$, which measures temporal variation: the term $\\chi$ penalizes squared differences between neighboring stabilized parameter vectors, and the term $c\\tau$ penalizes differences between more distant runs with weights that decay with the distance in the original parameter sequence. Minimizing $F^*$ is reduced to a generalized eigenvalue problem, and the eigenvectors supply $S$. The classification half of the paper is carried by the kernel machine $K$ and the quantities $\\varphi_k(\\phi_r)=(\\nu_k(\\phi_r))^T f^C_k(\\phi_r)$, which compare stabilized sequences in a reproducing kernel Hilbert space and assign each sequence to a stability class.","core_discovery":"The central claim is Theorem 1: for $R$ running sequences, a stabilizer matrix $S$ with $S^TS=I$ can be found by minimizing $F^* = \\arg\\min_S \\left(\\operatorname{Tr}\\left(S^T \\Delta\\alpha (\\Delta\\alpha)^T S\\right) + c\\tau\\right)$, where $\\Delta\\alpha$ collects the differences between consecutive optimal parameter vectors, $\\tau$ penalizes longer-range differences through weighted coefficients $\\omega_{rs}$, and $c$ is a regularization constant. The minimization is rewritten as the generalized eigenvalue problem $\\left(\\Delta\\alpha\\,\\sigma\\,(\\Delta\\alpha)^T\\right)S = \\lambda\\left(\\Delta\\alpha\\,\\eta\\,(\\Delta\\alpha)^T\\right)S$, and the solution yields $\\beta=S^T\\alpha$. The paper asserts that the resulting stabilized states $|\\phi_r\\rangle$ achieve $f(\\phi)=f(\\theta^*)$, i.e. the stabilization does not cost any objective-function value. Theorem 2 then assigns each stabilized sequence to one of $K$ stability classes through kernel-based correlation analysis in a high-dimensional Hilbert space, with the probabilistic classifier $f^C_k(\\phi_{r,i})$ and weights $\\nu_k(\\phi_{r,i})=\\phi_{r,i}/\\pi$.","pith_inferences":["A direct extension, not tested in the paper, is to close the loop: recompute $S$ after each stabilization step and re-optimize the objective on the stabilized manifold, turning a one-shot correction into adaptive calibration.","The same machinery would transfer to variational and approximate-optimization parameter schedules, where drift between objective evaluations is a known obstacle, since the construction only uses the parameter sequences, not the specific circuit.","The invariance claim is the point most worth testing: a numerical sweep over random unitaries and cost operators would show whether $S$ must be restricted to transformations that commute with $C$, a stronger condition than orthogonality."],"forward_implications":["Any set of optimal parameter vectors $\\alpha$ yields a stabilized set $\\beta=S^T\\alpha$ with reduced run-to-run temporal variation, because the objective $F^*$ is explicitly minimized for that purpose.","The stabilized state is claimed to preserve the objective value, $f(\\phi)=f(\\theta^*)$, so solution quality of the original problem is not degraded.","The stabilization is a classical computation on the gate parameters, so it can be applied to an existing circuit without altering the hardware unitaries.","For any number of runs $R$, each run receives a stability-class label, allowing drift to be monitored over time.","Because the learning procedures are unsupervised, no pre-labeled stable states are needed to find $S$ or the stability classes."],"supporting_citations":[{"why":"defines the parameterized quantum circuit family whose unitaries $U_i(\\theta_i)=e^{-i\\theta_i P}$ are the objects being stabilized.","marker":"[12]"},{"why":"supplies the form of the unitaries as exponentials of generalized Pauli operators, the starting point of the gate-parameter model.","marker":"[13]"},{"why":"appears as the reference for the regularization conventions used in the $F^*$ objective.","marker":"[31]"},{"why":"provides the manifold-regularized slow feature analysis objective that $F^*$ adapts for stabilization.","marker":"[32]"},{"why":"is the slow feature analysis source for the idea of minimizing temporal variation in the gate parameters.","marker":"[33]"},{"why":"supplies kernel PCA methodology behind the kernel machine $K$ used for stability classification.","marker":"[34]"},{"why":"supports the second-order temporal pooling classifier that the stability-class algorithm adapts.","marker":"[37]"}],"fun_headline_variants":["Stabilizer matrix locks quantum state with zero objective loss","No-cost stabilization for gate-model quantum computers","Orthogonal stabilizer matrix ends quantum drift, preserves state","State stabilization that doesn't cost you the objective"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that changing the gate parameters from the optimal values in the particular way the stabilizer matrix prescribes never changes how well the problem is solved, even though the recipe for choosing that change never looks at the objective function or the cost operator.","fun_headline_variants_meta":{"raw":{"variants":["Stabilizer matrix locks quantum state with zero objective loss","No-cost stabilization for gate-model quantum computers","Orthogonal stabilizer matrix ends quantum drift, preserves state","State stabilization that doesn't cost you the objective"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1340,"prompt_tokens":896,"completion_tokens":444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":382}},"tokens_in":512,"tokens_out":444,"duration_ms":5047,"temperature":1.0,"reasoning_tokens":382,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:28:00.207533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a one-qubit circuit $U(\\theta)=e^{-i\\theta X}$ with objective $C=Z$ and two runs whose optimal parameters make the minimizer of $F^*$ nontrivial; compute the stabilized parameters $\\phi=S^T\\theta$, then compare $\\langle\\phi_r|Z|\\phi_r\\rangle$ with $\\langle\\theta_r|Z|\\theta_r\\rangle$ for each $r$. A single mismatch would show that the stabilized state is not optimal, contradicting the paper's central claim.","supporting_citations":[{"cited_title":"S., Elons, A","cited_arxiv_id":null,"evidence_quote":"appears as the reference for the regularization conventions used in the $F^*$ objective."},{"cited_title":"Manifold Regularized Slow Feature Analysis for Dynamic Texture Recognition","cited_arxiv_id":"1706.03015","evidence_quote":"provides the manifold-regularized slow feature analysis objective that $F^*$ adapts for stabilization."},{"cited_title":"and Sejnowski, T","cited_arxiv_id":null,"evidence_quote":"is the slow feature analysis source for the idea of minimizing temporal variation in the gate parameters."},{"cited_title":"R., Scholz, M","cited_arxiv_id":null,"evidence_quote":"supplies kernel PCA methodology behind the kernel machine $K$ used for stability classification."}],"review_version":1}