{"id":"ae3187ee-d2fb-4671-aee1-6979c64f8fc8","arxiv_id":"2506.17883","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors prove a benign optimization landscape for a Pauli-based diagonalization cost, but the claimed new family of efficiently diagonalizable Hamiltonians is invalid as stated because its 'diagonal' D includes Y operators, and the experiments are warm-started from exact eigensolutions.","lead":"The paper proposes classical optimization algorithms that diagonalize Hamiltonians by minimizing a cost function over Pauli strings, and it proves that every nonzero stationary point of that cost is a global minimum. It also claims a new family of fast-forwardable Hamiltonians, but that family's central example calls a non-diagonal matrix diagonal, and the numerical tests start from the known answer.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Example 1's D=I+Σ d_jY_j contradicts Eq. (2), so the proposed family of quantum-diagonalizable Hamiltonians is not shown to exist; the central breadth claim collapses.","rationale":"The reader's overall rejection is justified, but my primary load-bearing concern is not the weakest_assumption listed (ansatz availability). It is the internal contradiction in Example 1: the paper's own Eq. (2) defines diagonal matrices via I/Z Pauli factors, while Eq. (35) defines D with Y_j factors. Since Appendix C's proof of exponential Lie algebra dimension depends on those Y_j factors, the error is not cosmetic. The landscape theorems (Theorems 2 and 3) may be salvageable and are not the focus of my objection. Separately, the algorithm's polynomial-time claim is qualified in the paper's introduction as holding 'given a known unitary decomposition,' and the numerical experiments initialize from exact eigensolutions, so the ansatz-discovery gap is real but secondary. Together these issues mean the central claims are not currently supported, so REJECT remains the appropriate verdict.","tokens_in":18325,"tokens_out":7325,"duration_ms":78439,"concrete_test":"Take n=2 in Eq. (35), set d1=d2=1, and evaluate D=I+Y1+Y2 in the computational basis: the matrix element ⟨00|D|11⟩ = ⟨0|Y|1⟩ = -i ≠ 0. Thus D has a nonzero off-diagonal entry, contradicting Eq. (2), which permits only I/Z factors in a diagonal matrix. Equivalently, numerically form H=UDU† and verify that U is not a spectral unitary because U†HU=D is not diagonal; this single check settles the internal inconsistency without any Lie-algebra computation.","verdict_should_be":"REJECT","load_bearing_attack":"Example 1 (Eq. 35) defines D=I+Σ_{j=1}^n d_j Y_j. Eq. (2) explicitly classifies any Pauli string containing an X or Y factor as off-diagonal, and states that a diagonal matrix has a Pauli expansion using only I and Z factors. Hence D is not a diagonal matrix in the computational basis, so H=UDU† is not an eigendecomposition and U is not a diagonalizing unitary. Definition 2, which the new family is claimed to satisfy, requires a genuinely diagonal D that is sparse in the Pauli basis. Appendix C then uses the Y_j factors in D to generate X_1, Z_1, Z_1Z_2, etc., so this is not a simple typo: replacing Y_j with Z_j would break the Lie-algebra argument. The paper's headline broadening of fast-forwardable systems to exponentially large Lie algebras therefore rests on an invalid example, and the claimed new family of quantum-diagonalizable Hamiltonians is unproven.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents classical optimization algorithms for Hamiltonian diagonalization. Given a Hamiltonian H expressed as a sum of M Pauli strings, the authors parameterize a candidate unitary K(r,θ) = Σ_j r_j e^{iθ_j} P_j and define a cost F(r,θ) consisting of the squared off-diagonal Pauli coefficients of K†HK plus a penalty enforcing K†K ∝ I. They prove (Theorems 2 and 3) that every nonzero stationary point of F is a global minimum and corresponds to a KHK decomposition of H, and they give a posteriori error bounds linking approximate cost minimization to approximate diagonalization. A deterministic gradient-descent-with-normalization algorithm is analyzed under a uniform KL condition (Assumption 1), and a randomized-coordinate variant is proposed with reduced per-iteration cost. The paper also claims a new family of quantum-diagonalizable Hamiltonians with exponentially large Lie algebras but polynomially sparse Pauli decompositions (Examples 1 and 2). Numerical experiments on random sparse Hamiltonians, the XXZ model, and a Hubbard model are reported.","tokens_in":18551,"tokens_out":12707,"duration_ms":125867,"significance":"If all claims were valid, the paper would make a useful contribution: a classically implementable optimization landscape with no spurious stationary points, polynomial per-iteration cost for Hamiltonians with sparse Pauli diagonalizations, and error certificates would complement existing Lie-algebraic and variational diagonalization methods. The algebraic core—the stationary-point argument in Theorems 2 and 3 and the error bound in Lemma 2—appears sound and is the strongest part of the paper. The code is publicly available and the numerical experiments are clearly described. However, the claimed new family of quantum-diagonalizable Hamiltonians is not established as presented, and the polynomial-efficiency claim presupposes a known ansatz containing the diagonalizing unitary; these issues affect the paper's central advertised contributions. The convergence theorem also has a regime where the proof does not support the stated rate.","major_comments":[{"comment":"The matrix D = I + Σ_{j=1}^n d_j Y_j is not diagonal in the computational basis under the paper's own classification in Eq. (2), which places every Pauli string containing an X or Y factor in the off-diagonal part. Consequently H = U D U† is not an eigen-decomposition and U is not a diagonalizing unitary, so the example does not satisfy Definition 2. This is not a typo: Appendix C uses the Y_j factors in D to obtain X_1, Z_1, Z_1Z_2 and the commutators that generate the set (C1); replacing Y_j by Z_j would break the Lie-algebra argument. The claimed new family of quantum-diagonalizable Hamiltonians with exponentially large Lie algebras, and the discussion in Section VI that this family strictly contains the PLA Hamiltonians, are therefore unsupported.","section":"§IV, Example 1 (Eq. (35)) and Definition 2"},{"comment":"The optimization problem is only defined relative to a user-supplied Pauli ansatz {P_j}, and the correctness and efficiency claims assume that this ansatz spans a diagonalizing unitary of H with d = poly(n). No procedure is given to construct such an ansatz from H alone; in the Section V experiments the initial K is obtained from an eigensolver and then expanded in Pauli strings, so the ansatz and initialization are derived from the known eigendecomposition. The abstract's claim of polynomial-time efficiency for diagonalizing unknown H is therefore not established; the paper demonstrates polynomial per-iteration cost for a subproblem whose input includes an oracle-like ansatz.","section":"§II–§V, Eq. (4) and Algorithm 1"},{"comment":"The convergence-rate derivation is not valid for the claimed range α ∈ [0,1]. In Eqs. (A9)–(A10), the factor 1 − μa F(x_t)^{α−1} is used; when α < 1 and F(x_t) is small, F^{α−1} = 1/F^{1−α} is large and the factor can become negative, so the inequality ε ≤ (1 − μa ε^{α−1})^T F(x_0) used to extract T is meaningless. The stated bound T = O(ε^{−max{α−1,0}} log(1/ε)), which sets the exponent to zero for α ∈ [0,1], does not follow from the proof; a separate argument is needed for the linear-convergence claim in that regime.","section":"Theorem 4 and Appendix A"}],"minor_comments":[{"comment":"The text contains the typo 'learing rate schedule'; it should read 'learning rate schedule'.","section":"Algorithms 1 and 2"},{"comment":"As written, G1 is defined through K(r,θ) but K depends on the optimization parameters; the definition should clarify that G1 is the fixed set of all off-diagonal Pauli strings that can appear in the symbolic expansion of K†HK, with zero coefficients allowed at particular parameter values.","section":"§II, Eq. (3) and definition of G1"},{"comment":"The phrase 'quadratically reduces the per-iteration cost from O(d^4 M^2) to O(d^2 M)' is imprecise, since the ratio is O(d^2 M); it would be clearer to say the reduction factor is O(d^2 M).","section":"§III.D"},{"comment":"The randomized-coordinate algorithm is presented as a practical contribution, but Section III.D states that no convergence guarantee is available for it ('deriving a lower bound of the norm of r_{t+1} and a convergence guarantee is still an open issue'); the abstract and conclusion should state this limitation explicitly rather than presenting the variant as fully supported.","section":"Abstract and §III.D"}],"recommendation":"reject","confidential_remarks":"The stationary-point analysis and the a posteriori error bounds are the strongest parts of the manuscript. If the authors are willing to reframe the paper without the invalid new-family claim and to state the ansatz assumption explicitly as an oracle-like input, a revised manuscript focused on the optimization landscape could be of interest. In its current form, the central breadth claim is not established and the efficiency claim is conditional on information that is not shown to be available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a clean, self-contained result buried under an invalid headline claim. Theorems 2 and 3 — any nonzero stationary point of the Pauli cost with orthogonality penalty is a global minimum and yields a KHK decomposition — are algebraically coherent and appear correct. That is a real contribution to the landscape literature for this kind of cost function. The posterior error bound in Lemma 2 is also useful.\n\nBut the advertised broadening of fast-forwardable systems to exponentially large Lie algebras does not hold as written. Example 1 defines D = I + Σ d_j Y_j and calls it diagonal. By the paper's own Eq. (2), any Pauli string with an X or Y is off-diagonal; diagonal matrices use only I and Z. So H = U D U† is not an eigendecomposition and U is not a diagonalizing unitary. This is not a typo: the proof in Appendix C uses the Y_j factors to generate X_1, Z_1, Z_1Z_2, etc. Replacing Y_j with Z_j would collapse the Lie-algebra argument. So the claimed new family of quantum-diagonalizable Hamiltonians with exponentially large dynamical Lie algebras is unproven. The diagram in Fig. 3 overclaims.\n\nThe second soft spot is the algorithm's efficiency claim. The ansatz K(r,θ) is a fixed Pauli set; the user must supply a set containing a diagonalizing unitary, with d polynomial, and initial r,θ. In every numerical demonstration the ansatz and initial parameters come from the known eigendecomposition (Section V). No procedure finds the ansatz from H alone. So polynomial-time classical diagonalization of unknown H is not demonstrated. The randomized variant's convergence is explicitly left open.\n\nThe convergence theory depends on Assumption 1, a uniform KL condition that is stated but not proven for any concrete family; numerical α values are reported, but that is validation, not proof.\n\nNet: the landscape theorem deserves a serious referee and could be published with the examples corrected or removed. As is, the central breadth claim fails. I'd send it to review — a good referee can separate the salvageable core from the overreach — but not accept it now. The authors should be told to either fix Example 1 with a genuine diagonal D or drop the exponential-Lie-algebra claim.","headline":"A clean landscape theorem is buried under an invalid headline example: Example 1's D contains Y Paulis, so it is not diagonal by the paper's own Eq. (2), and the claimed exponential-Lie-algebra family is unproven.","tokens_in":19073,"tokens_out":2011,"would_cite":false,"duration_ms":19920,"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":"A classical optimization algorithm can diagonalize a Hamiltonian without ever getting stuck in a spurious minimum.","keywords":["Hamiltonian diagonalization","Pauli basis","cost landscape","global minima","fast-forwarding","Lie algebra","gradient descent","quantum simulation"],"falsifier":"A concrete falsifier is a numerical scan on a small Hamiltonian, such as the four-qubit XXZ model, with a fixed Pauli ansatz known to contain the exact diagonalizing unitary: run Algorithm 1 from many random initial points and check whether any nonzero stationary point with $F>0$ is reached, since the theorem predicts none. A complementary test would construct a Hamiltonian whose diagonalizing unitary provably lies outside every polynomial-size Pauli ansatz and show the algorithm's output is not a diagonalization, exposing the practical scope limitation.","tokens_in":18056,"feed_emoji":"⚛️","tokens_out":6935,"duration_ms":65463,"temperature":0.7,"pith_summary":"This paper aims to establish that Hamiltonian diagonalization, the key primitive for long-time quantum simulation, can be recast as a classical optimization problem with a benign landscape: every nonzero stationary point of the cost function is a global minimum and yields a valid diagonalization. The cost function is built from Pauli-string coefficients, penalizing off-diagonal terms and enforcing unitarity through an orthogonality constraint. If true, this provides a route to polynomial-time diagonalization for a broader class of fast-forwardable Hamiltonians, including cases whose Lie algebras are exponentially large. It also supplies a posteriori error bounds that convert optimization error into diagonalization accuracy, plus a randomized-coordinate variant with cheaper per-iteration cost.","feed_headline":"Optimization diagonalizes Hamiltonians with no spurious minima","feed_subtitle":"A Pauli cost function turns Hamiltonian diagonalization into a benign landscape with polynomial iteration cost.","key_machinery":"The central object is a parameterized Pauli-basis ansatz $K(r,\\theta)=\\sum_{j=1}^d r_j e^{i\\theta_j} P_j$, whose coefficients $r_j e^{i\\theta_j}$ are exactly the Pauli-basis coefficients of $K$ in polar form. The cost $F$ combines $f(r,\\theta)$, the squared off-diagonal Pauli coefficients of $K^\\dagger H K$, with an orthogonality penalty built from coefficients $\\phi_P$ of $K^\\dagger K$. The identity that carries the argument is $\\sum_j r_j \\partial F/\\partial r_j = 4F(r,\\theta)$, which, together with the explicit gradient formula, forces $F=0$ at any nonzero stationary point; the orthogonality part then forces $K^\\dagger K = \\|r\\|^2 I$, making the stationary $K$ unitary up to scaling.","core_discovery":"The central claim is Theorem 3: any nonzero stationary point $(r_c,\\theta_c)$ of the total cost $F(r,\\theta)=\\sum_{P\\in G_1} \\mathrm{tr}(K(r,\\theta)^\\dagger H K(r,\\theta) P)^2 + \\sum_{P\\in G_2} \\phi_P(r,\\theta)^2$ is a global minimum, and $K(r_c,\\theta_c)=\\sum_{j=1}^d r_j e^{i\\theta_j} P_j$ is unitary up to scaling, giving $H = \\|r\\|^{-4} K h K^\\dagger$ with $h=K^\\dagger H K$ diagonal. Hence every nontrivial fixed point of gradient descent with normalization is a genuine diagonalization. The paper also proves convergence under a uniform Kurdyka-Lojasiewicz condition and derives an a posteriori bound that maps small $F$ to closeness of $H$ to $\\widetilde{H}=K h_0 K^\\dagger$ and of the corresponding eigenspace projectors.","pith_inferences":["A natural testable extension is a bootstrap that grows the Pauli ansatz from the terms of $H$ until the stationary-point condition $F=0$ is met, since the paper gives no autonomous procedure for discovering the ansatz.","The construction of Example 1 suggests a general recipe for generating hard-to-diagonalize-looking Hamiltonians: conjugate a diagonal Pauli-sparse $D$ by products of a few anticommuting Pauli rotations, which may inflate the dynamical Lie algebra to $su(2^n)$ while keeping the eigendecomposition Pauli-sparse.","The continuation observation that optimal parameters for one Hamiltonian initialize well for nearby parameters points toward a practical parameter-sweep workflow for tunable or time-dependent models, though the paper demonstrates this only numerically."],"forward_implications":["Any nontrivial stationary point of the optimization is a correct diagonalization, so gradient descent with normalization cannot get stuck at spurious minima; the only failure mode is convergence to $r=0$, which the normalization step is designed to avoid.","The convergence theorem gives sublinear or linear rates under a uniform KL condition, with iteration count $O\\left(\\epsilon^{-\\max\\{\\alpha-1,0\\}} \\log(1/\\epsilon)\\right)$, translating numerical optimization effort into a diagonalization guarantee.","The a posteriori error bound means a user can certify approximate diagonalization and eigenspace accuracy directly from the final cost function value.","For Hamiltonians whose diagonalizing unitary is sparse in the Pauli basis, the per-iteration cost is polynomial: $O(d^4 M^2)$ deterministic and $O(d^2 M)$ for the randomized-coordinate variant.","The constructed family of Hamiltonians shows that quantum-diagonalizable Hamiltonians strictly contain those with polynomially sized Lie algebras, so exponential Lie-algebra dimension alone does not preclude efficient diagonalization."],"supporting_citations":[{"why":"Defines fast-forwarding and the no-fast-forwarding bound that motivates diagonalization.","marker":"[2]"},{"why":"Supplies the maximum number of anticommuting Pauli strings used to build unitaries with exponentially large Lie algebras.","marker":"[4]"},{"why":"Baseline variational fast-forwarding method whose landscape may contain spurious optima; the paper compares against it.","marker":"[8]"},{"why":"Baseline double-bracket diagonalization whose circuit depth grows exponentially with the number of iterations.","marker":"[14]"},{"why":"Lie-algebraic fast-forwarding framework whose per-iteration cost becomes exponential for the constructed Hamiltonians.","marker":"[16]"},{"why":"Cartan-decomposition optimizer that the paper's cost function generalizes and whose code is used in the numerical experiments.","marker":"[21]"},{"why":"Defines quantum diagonalizable Hamiltonians and raises the comparison with fast-forwarding.","marker":"[26]"},{"why":"Provides the Pauli generating set used to prove that the Lie algebras in Examples 1 and 2 are $su(2^n)$.","marker":"[31]"},{"why":"Lie-diagonalization algorithm with growing circuit depth, used as a comparison baseline.","marker":"[32]"}],"fun_headline_variants":["Classical algorithm diagonalizes Hamiltonians without spurious minima","Hamiltonian diagonalization without spurious minima, now polynomial","Classical optimizer finds true diagonalization, no faulty minima","Polynomial-time classical diagonalization for quantum Hamiltonians","Broadening fast-forwardable systems via benign optimization landscape"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The algorithm must be handed a Pauli-string ansatz $K(r,\\theta)$ whose span contains a unitary that diagonalizes $H$, and in every numerical experiment that ansatz and its initialization are taken from the known eigendecomposition of the target Hamiltonian; the paper gives no procedure to discover such an ansatz from $H$ alone.","fun_headline_variants_meta":{"raw":{"variants":["Classical algorithm diagonalizes Hamiltonians without spurious minima","Hamiltonian diagonalization without spurious minima, now polynomial","Classical optimizer finds true diagonalization, no faulty minima","Polynomial-time classical diagonalization for quantum Hamiltonians","Broadening fast-forwardable systems via benign optimization landscape"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000823,"raw_usage":{"total_tokens":3608,"prompt_tokens":961,"completion_tokens":2647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":2568}},"tokens_in":577,"tokens_out":2647,"duration_ms":20216,"temperature":1.0,"reasoning_tokens":2568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:59:31.706029+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete falsifier is a numerical scan on a small Hamiltonian, such as the four-qubit XXZ model, with a fixed Pauli ansatz known to contain the exact diagonalizing unitary: run Algorithm 1 from many random initial points and check whether any nonzero stationary point with $F>0$ is reached, since the theorem predicts none. A complementary test would construct a Hamiltonian whose diagonalizing unitary provably lies outside every polynomial-size Pauli ansatz and show the algorithm's output is not a diagonalization, exposing the practical scope limitation.","supporting_citations":[{"cited_title":"Fast-forwarding of hamiltonians and exponentially precise measure- ments.Nature communications, 8(1):1572, 2017","cited_arxiv_id":null,"evidence_quote":"Defines fast-forwarding and the no-fast-forwarding bound that motivates diagonalization."},{"cited_title":"Nearly optimal measurement schedul- ing for partial tomography of quantum states.Physical Review X, 10(3):031064, 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum number of anticommuting Pauli strings used to build unitaries with exponentially large Lie algebras."},{"cited_title":"Variational fast forwarding for quantum simulation beyond the coherence time.npj Quantum Infor- mation, 6(1):82, 2020","cited_arxiv_id":null,"evidence_quote":"Baseline variational fast-forwarding method whose landscape may contain spurious optima; the paper compares against it."},{"cited_title":"fast-forward","cited_arxiv_id":null,"evidence_quote":"Baseline double-bracket diagonalization whose circuit depth grows exponentially with the number of iterations."},{"cited_title":"Adaptive, problem-tailored variational quantum eigensolver mitigates rough parameter landscapes and barren plateaus.npj Quantum Information, 9(1):19, 2023","cited_arxiv_id":null,"evidence_quote":"Lie-algebraic fast-forwarding framework whose per-iteration cost becomes exponential for the constructed Hamiltonians."},{"cited_title":"Quantum random power method for ground state computation","cited_arxiv_id":"2408.08556","evidence_quote":"Cartan-decomposition optimizer that the paper's cost function generalizes and whose code is used in the numerical experiments."},{"cited_title":"Hamiltonian simulation in the low-energy subspace.npj Quantum Information, 7(1):119, 2021","cited_arxiv_id":null,"evidence_quote":"Provides the Pauli generating set used to prove that the Lie algebras in Examples 1 and 2 are $su(2^n)$."},{"cited_title":"Optimally generating $\\mathfrak{su}(2^N)$ using Pauli strings","cited_arxiv_id":"2408.03294","evidence_quote":"Lie-diagonalization algorithm with growing circuit depth, used as a comparison baseline."}],"review_version":2}