{"id":"6cdd6f24-62bd-42cf-afb3-4be1ea6e8ff5","arxiv_id":"2608.09661","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The authors show that a Walsh-Hadamard transform cuts encoding memory by up to 1298x and an SVD-based two-term encoding gives a >10,000x per-iteration speedup for VQLS on CFD problems, while expressibility metrics fail to predict ansatz convergence.","lead":"This paper benchmarks ways to encode the matrix in a variational quantum linear solver, reporting large memory and per-iteration speedups for two encoding tricks on fluid-flow test problems. It also finds that common circuit expressibility scores do not predict how well an ansatz converges, and demonstrates the pipeline on a 15-qubit system on a supercomputer.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 10,000x per-iteration speedup in §IV-E rests on PennyLane applying the dense SVD unitaries U± from Eq. (6) as QubitUnitary; without gate synthesis or convergence data, the headline advantage is a simulator primitive, not a demonstrated circuit or end-to-end speedup.","rationale":"The reader correctly locates the load-bearing assumption in §III-C: the two-term SVD LCU is mathematically exact but operationally unspecified without circuits for U±. I read the per-iteration timing claim carefully; Figure 11 uses PennyLane QubitUnitary, so the 10,000x factor measures how many circuit evaluations are avoided, not gate-level cost. This is especially important because the paper's own §IV-E reports out-of-memory for SVD coherent VQLS at larger sizes, and the 15-qubit success uses Pauli coherent encoding instead. I do not think the SVD identity is wrong, and the FWHT memory result and ansatz observations are plausible; therefore this is a conditional-accept issue rather than a rejection. The concrete synthesis benchmark would settle whether the speedup survives a realistic gate model. The reader and I identify the same primary weakness, so I mark agreement as agree; I would keep the CONDITIONAL verdict unchanged.","tokens_in":14316,"tokens_out":11073,"duration_ms":106766,"concrete_test":"Compile U+ and U- from Eq. (6) for the exact 8-qubit tridiagonal Toeplitz instance behind Fig. 11 into elementary gates using a CNOT-count-oriented unitary synthesis (e.g., Qiskit QSD or KAK recursion), then rerun the SVD coherent VQLS timing in lightning.kokkos with that compiled circuit in place of QubitUnitary. If the CNOT count exceeds the total CNOT count of one pruned-Pauli VQLS iteration, or if the measured per-iteration advantage over the Pauli baseline falls below 10,000x, the headline speedup is an artifact of the dense-matrix primitive. As a secondary check, report the number of iterations to reach the same cost threshold for both encodings so the comparison can be expressed as end-to-end solve time.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (6)–(10) give an exact two-term LCU, A=(||A||2/2)(U+ + U-), but the paper never decomposes the orthogonal factors U and V into elementary gates. The §IV-E timings call PennyLane's QubitUnitary for these arbitrary matrices; a generic n-qubit unitary needs exponentially many (Θ(4^n)) two-qubit gates for exact synthesis, and the dense representation itself costs O(2^{2n}) memory, matching the reported out-of-memory failure of SVD coherent VQLS in §IV-E. Thus the >10,000x number at 8 qubits is mostly the number of circuit evaluations saved by a two-term LCU, not a reduction in circuit cost: one SVD coherent evaluation is a single dense block-encoding unitary, while one Pauli evaluation is thousands of shallow controlled-Pauli circuits. The manuscript also omits the classical SVD preprocessing and gives no convergence trajectory for the SVD coherent variant at the timed sizes, so the per-iteration advantage cannot be converted into wall-clock solve time or a hardware-realistic comparison. The 'circuit-efficient' title therefore rests on the simulator's dense-matrix primitive rather than on demonstrated circuit implementations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper benchmarks four LCU matrix-encoding strategies for VQLS on CFD-style linear systems—naive Pauli decomposition, PennyLane-integrated decomposition, FWHT-based parallel Pauli decomposition, and an SVD-based two-term LCU—and tests 11 ansatz families on a 4x4 Hele-Shaw problem. The main reported results are: FWHT reduces peak memory by up to 1298x for an 11x11 Hele-Shaw Jacobian; the SVD-based two-term LCU yields over 10,000x per-iteration speedup over Pauli-based VQLS at 8 qubits; expressibility and entanglement metrics correlate only weakly with VQLS convergence; and the Pauli coherent VQLS simulates a 15-qubit tridiagonal Toeplitz system on Frontier within the wall-time limit. The SVD derivation in Eqs. (6)-(10) is a valid exact two-term LCU, and the paper is candid in Section V that the coherent variant requires deep block-encoding circuits and that SVD preprocessing carries non-negligible classical overhead.","tokens_in":14560,"tokens_out":6067,"duration_ms":53866,"significance":"If the headline claims held in an end-to-end sense, this would be a useful engineering baseline: the FWHT decomposition is a clear scalability improvement, the exact SVD two-term LCU is mathematically elegant and reduces the number of circuit evaluations per VQLS iteration, and the negative ansatz result is practically relevant for algorithm selection. Strengths include the honest parameter sweep over truncation tolerance, the explicit decomposition algorithms, and the Frontier deployment data at up to 15 qubits. The significance is substantially limited, however, by the fact that the 10,000x per-iteration speedup is measured with PennyLane's dense QubitUnitary primitive for the SVD unitaries, for which no gate decomposition is supplied; the circuit-efficiency claim therefore applies to a simulator-specific primitive, not to a realistic quantum circuit. The missing end-to-end accounting (SVD preprocessing cost, convergence iterations, and memory footprints) also prevents converting the per-iteration numbers into wall-clock solve-time or hardware-realistic comparisons.","major_comments":[{"comment":"The abstract and Section IV-E claim 'over 10,000x per-iteration speedup' for SVD coherent VQLS, but the timings apply U+ and U- as PennyLane dense QubitUnitary operations. For generic orthogonal matrices U and V from Eq. (6), exact synthesis requires exponentially many two-qubit gates, and the dense representation costs O(2^{2n}) memory. Because no gate decomposition or depth estimate is supplied, the measured speedup reflects the number of circuit evaluations avoided by a two-term LCU, not a circuit-cost or hardware-realistic speedup; the 'circuit-efficient' characterization in the title is therefore unsupported. The paper's own Section V concedes that the coherent variant 'requires deep block-encoding circuits,' but this cost is neither quantified nor included in the headline comparison.","section":"§III-C, Eqs. (6)-(10); §IV-E, Fig. 11"},{"comment":"The SVD coherent variant fails with an out-of-memory error on GPU simulation before reaching 15 qubits, so the most favorable scaling curve in Figure 11 is not demonstrated at the largest sizes. The paper should report the memory footprint per qubit count, the size at which each encoding becomes infeasible, and the feasible range over which the claimed advantage actually holds. Without this, the statement that SVD coherent VQLS has 'the best performance among all mentioned VQLS variants' cannot be assessed as a scalability result.","section":"§IV-E, Fig. 11; §V"},{"comment":"The per-iteration timings exclude the classical SVD preprocessing of Eq. (6) and the construction of D± in Eq. (8), which is O(2^{3n}) for a dense 2^n x 2^n matrix. They also exclude the number of optimizer iterations required by the SVD coherent variant to converge; no convergence trajectories are shown for the timed runs. To support the practical claim of a '10,000x speedup' or 'best per-iteration timing,' the paper needs end-to-end wall-clock comparisons including preprocessing and iteration counts, or it must explicitly restrict the claim to the per-iteration simulator cost.","section":"§IV-E; §V"},{"comment":"The claim that expressibility and entanglement metrics 'correlate only weakly' with VQLS convergence is presented without a quantitative correlation measure. Table I gives KL divergence and Meyer-Wallach values for n=4, while Figure 9 shows only learning curves for one 4x4 Hele-Shaw instance with a 100-step budget. The paper should report a correlation coefficient or equivalent over the 11 ansatze and, ideally, over multiple problem instances; as written, the conclusion is an informal reading of two figures.","section":"§IV-D, Table I and Fig. 9"}],"minor_comments":[{"comment":"The objective statement 'minimizes 1−U(A)|V(w)⟩ · |b⟩ †' is malformed; it should be written as a scalar fidelity expression such as 1 − |⟨b| U(A) V(w)|0⟩|^2.","section":"§III-E"},{"comment":"The text contains several typos and grammar errors, including 'an singular value decomposition,' 'pennylanelightning.kokkosto,' 'Toplitz,' 'computation fluid dynamic,' and 'demonstrating a increasing performance.' A careful copyedit is needed.","section":"Abstract and throughout"},{"comment":"The formula '2L2nb circuit executions' is ambiguous; it should be written with explicit notation, likely 2 L 2^{n_b} circuit executions per function evaluation.","section":"§IV-C"},{"comment":"The caption says 'Scipy-free optimizers,' but the text and legend refer to 'Scipy-based' optimization; the caption contradicts the body and should be corrected.","section":"Fig. 11 caption"},{"comment":"Algorithm 2 does not include the tolerance parameter or the truncation step that Section IV-A describes for the approximate Pauli decomposition; adding the thresholding step to the pseudocode would improve reproducibility.","section":"Algorithm 2 and §IV-A"}],"recommendation":"major_revision","confidential_remarks":"The benchmark-heavy nature of the paper makes data and code availability important; the manuscript currently has no data availability statement. I would encourage the editor to request release of the timing tables, memory logs, and benchmarking scripts alongside the revision. The core SVD-LCU identity is correct and the FWHT memory results are useful, but the headline speedup needs to be re-scoped or supplemented with gate-decomposition and end-to-end costs before the paper's title-level claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real value here is the systematic benchmark of matrix encodings for VQLS on CFD-type matrices. The FWHT-based Pauli decomposition numbers (1298x peak memory reduction on an 11x11 Hele-Shaw grid, O(n^2 log n) scaling) are concrete and useful, and the truncation sweep in Figure 6 is a practical contribution. The two-term SVD LCU derivation (Eqs. 6-10) is mathematically correct, and the ansatz comparison sensibly tests whether expressibility and entanglement metrics predict convergence; the finding that they don't is worth reporting, even though it comes from a single 4x4 instance. I also credit the authors for actually running on Frontier and for admitting in the conclusions that SVD preprocessing has non-negligible overhead and that no single encoding dominates. The soft spots are exactly where the reader and the stress-test put them. The 10,000x per-iteration speedup at 8 qubits comes from calling PennyLane's QubitUnitary with the dense orthogonal factors U and V from Eq. (6). Since no gate decomposition is given for these arbitrary unitaries, the speedup is a measure of how many circuit evaluations a two-term LCU saves in a simulator, not a reduction in circuit cost on hardware or in a scalable state-vector simulator. The paper also omits the classical SVD preprocessing time, and provides no convergence trajectory for the SVD coherent variant at the timed sizes - so a reader cannot tell whether the per-iteration win survives wall-clock or yields a correct solution. The title's \"circuit-efficient\" is therefore an overclaim. These are fixable in revision: give a synthesis method for U+- (or state clearly that none is known and benchmark accordingly), report end-to-end times including SVD, and show convergence at 15 qubits. One thing I disagree with in the skeptic's note: calling the ansatz correlation \"weak\" is fine, but the paper doesn't overstate it - they conclude only that generic metrics are not a reliable proxy. Fair. Also, the self-citation to their earlier Hele-Shaw work is legitimate given the shared problem setup. Who is this for? Someone building VQLS pipelines and choosing between Pauli and block-encoding strategies will get practical guidance from the encoding benchmarks and the truncation trade-off. It deserves a serious referee, but the referee should insist on the missing gate synthesis and convergence data before publication. My own verdict would be conditional accept after major revision.","headline":"Useful engineering benchmarks for VQLS matrix encodings, but the headline 10,000x speedup is a simulator primitive, not a demonstrated circuit or end-to-end advantage; the paper needs revision before it should appear.","tokens_in":702,"tokens_out":719,"would_cite":true,"duration_ms":21621,"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":"Matrix encoding dominates VQLS runtime: an SVD two-term encoding delivers over 10,000x per-iteration speedup at 8 qubits, while FWHT-based Pauli decomposition cuts peak memory up to 1,298x, and expressibility metrics fail to predict…","keywords":["VQLS","linear combination of unitaries","Fast Walsh-Hadamard transform","Pauli decomposition","singular value decomposition","block encoding","Hele-Shaw flow","quantum-HPC"],"falsifier":"Compile the SVD-derived unitaries $U_+$ and $U_-$ for an 8-qubit tridiagonal Toeplitz system into elementary gates (single-qubit rotations and CNOTs) and measure the resulting circuit depth and per-iteration time: if the depth grows exponentially with qubit count, the 10,000x per-iteration speedup disappears outside the dense-matrix simulator. A complementary test is to run the coherent SVD block-encoded VQLS on real quantum hardware at 8 qubits and compare wall-clock time per iteration with the Pauli-encoded VQLS; the speedup claim stands only if the hardware times still favor the SVD encoding.","tokens_in":14088,"feed_emoji":"⚛️","tokens_out":14358,"duration_ms":104093,"temperature":0.7,"pith_summary":"This paper tries to show that, for the Variational Quantum Linear Solver (VQLS) applied to fluid-dynamics linear systems, the way the system matrix is encoded is the main lever on runtime and memory. It derives an exact two-term linear-combination-of-unitaries (LCU) representation from the singular value decomposition, $A = \\frac{\\|A\\|_2}{2}(U_+ + U_-)$, and reports that combining it with a coherent, block-encoded VQLS gives over 10,000x per-iteration speedup over standard Pauli-based VQLS at 8 qubits, with the advantage widening through 15 qubits. The paper also shows that a Fast Walsh-Hadamard Transform (FWHT)-based Pauli decomposition reduces peak classical memory by up to 1,298x on an 11x11 Hele-Shaw grid, and that tolerance-based truncation can cut a 15-qubit Toeplitz matrix to four Pauli terms at 0.045 relative Frobenius error. On the ansatz side, benchmarking 11 circuit families, it finds expressibility and entanglement metrics correlate only weakly with convergence, arguing for problem-aware ansatz design. The motivation is to give an end-to-end, reproducible baseline for VQLS on hybrid quantum-HPC systems, where CFD workloads require many repeated linear solves.","feed_headline":"SVD encoding makes VQLS 10,000x faster per step","feed_subtitle":"For fluid-flow linear systems, how the matrix is encoded matters more than ansatz choice; memory also drops up to 1,298x.","key_machinery":"The load-bearing identity is the SVD two-term LCU: $A = \\frac{\\|A\\|_2}{2}(U D_+ V^T + U D_- V^T)$, with $D_\\pm = \\frac{\\Sigma}{\\|A\\|_2} \\pm i\\sqrt{I - (\\Sigma/\\|A\\|_2)^2}$, where $A = U\\Sigma V^T$ and $\\|A\\|_2$ is the spectral norm. It maps each normalized singular value onto the unit circle, making $D_\\pm$ unitary and therefore $U_\\pm = U D_\\pm V^T$ unitary, so every real square matrix becomes a sum of exactly two unitaries regardless of size. The other central mechanism is the Fast Walsh-Hadamard Transform computation of Pauli coefficients, which replaces the naive enumeration of all $4^n$ Pauli strings with an $O(n^2 \\log n)$ transform using fixed auxiliary memory. Ansatz quality is measured by expressibility, the Kullback-Leibler divergence of the circuit's fidelity distribution from the Haar-uniform distribution, and by mean Meyer-Wallach entanglement capability.","core_discovery":"On its own terms, the paper's central claim is that matrix-encoding strategy, not ansatz choice, dominates per-iteration cost in VQLS. The key move is an exact two-term LCU from the singular value decomposition: writing $A = U\\Sigma V^T$ and normalizing $\\Sigma$ by the spectral norm, the diagonal matrices $D_\\pm = \\frac{\\Sigma}{\\|A\\|_2} \\pm i\\sqrt{I - (\\Sigma/\\|A\\|_2)^2}$ are unitary, so $A = \\frac{\\|A\\|_2}{2}(U D_+ V^T + U D_- V^T)$ expresses any real square matrix as half the sum of two unitaries. With a coherent VQLS that encodes $A$ as a block in a single circuit, the cost evaluated per iteration drops from $2L^2 n_b$ circuit executions in the Pauli approach to a single circuit, giving the measured 10,000x speedup at 8 qubits and best timing up to 15 qubits. The paper further claims that the FWHT-based Pauli decomposition computes all coefficients with $O(n^2 \\log n)$ work and fixed auxiliary memory, cutting peak memory by up to 1,298x on an 11x11 Hele-Shaw Jacobian, and that an approximate, tolerance-controlled version compresses a 15-qubit tridiagonal Toeplitz matrix to four Pauli terms with relative Frobenius error 0.045. Finally, across 11 ansatz families, expressibility (KL divergence from the Haar-uniform fidelity distribution) and entanglement capability show only weak correlation with convergence, so generic circuit metrics are unreliable guides to VQLS performance.","pith_inferences":["The 10,000x speedup is measured in a simulator that applies the SVD-derived unitaries as dense matrices; since no gate decomposition for $U$ and $V$ is provided, the speedup is likely to shrink or vanish on real hardware where arbitrary orthogonal matrices require exponential-depth circuits.","The FWHT memory reduction applies to the classical preprocessing step only, not to the number of circuits executed per VQLS iteration, so it helps matrix initialization scale while leaving Pauli-encoded per-iteration runtime unchanged.","If structure-aware decompositions of $U$ and $V$, such as Givens rotations for tridiagonal systems, can be compiled efficiently, the SVD two-term encoding could become hardware-portable; testing per-iteration time with such compiled circuits would settle whether the speedup survives outside the simulator.","The weak correlation between expressibility and convergence suggests cost-landscape features such as gradient variance or the density of local minima, rather than expressibility, govern VQLS trainability; measuring those features across the 11 ansatz families would be a direct next step."],"forward_implications":["FWHT-based Pauli decomposition scales classical matrix preprocessing to $O(n^2 \\log n)$ with fixed auxiliary memory, so VQLS initialization can reach larger grid sizes before memory runs out.","The SVD two-term LCU reduces the number of unitaries in the cost circuit to exactly two, and with coherent VQLS it yields the lowest per-iteration time of every strategy tested, with the advantage growing from 8 to 15 qubits.","Tolerance-controlled approximate Pauli decomposition can compress a 15-qubit tridiagonal Toeplitz matrix to four Pauli terms at 0.045 relative Frobenius error, substantially reducing circuit-count pressure with negligible loss of solution fidelity.","Expressibility and entanglement metrics are not reliable proxies for VQLS convergence, so ansatz selection should be driven by problem structure rather than generic circuit statistics.","The complete workflow runs on a production HPC system and simulates a 15-qubit tridiagonal Toeplitz linear system on a single node within a two-hour wall-time budget."],"supporting_citations":[{"why":"Defines the VQLS algorithm and cost function that the entire workflow optimizes.","marker":"[18]"},{"why":"Supplies the Fast Walsh-Hadamard Transform method for Pauli decomposition coefficients that the paper parallelizes and benchmarks.","marker":"[25]"},{"why":"Defines the expressibility and entanglement-capability metrics used to compare the 11 ansatz families.","marker":"[26]"},{"why":"Introduces the coherent VQLS formulation with block encoding that the SVD-based encoding is combined with.","marker":"[27]"},{"why":"Provides the Hele-Shaw flow discretization and its linear system used as the canonical CFD test problem.","marker":"[23]"},{"why":"The simulation software used to build and run all VQLS circuits whose runtime and memory are reported in the benchmarks.","marker":"[33]"}],"fun_headline_variants":["SVD encoding accelerates VQLS 10,000x per iteration","Matrix encoding, not ansatz, drives VQLS speedup","FWHT slashes VQLS memory 1,298x on CFD grids","Coherent VQLS beats Pauli encoding by 10,000x","SVD-based VQLS: 10,000x faster, 1,298x less memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The timing comparisons assume the two SVD-derived unitaries $U_+$ and $U_-$ can be applied at the cost of a dense-matrix unitary operation in the simulation software, but the paper gives no gate-level decomposition of the orthogonal matrices $U$ and $V$, which generally require exponentially deep circuits on real hardware.","fun_headline_variants_meta":{"raw":{"variants":["SVD encoding accelerates VQLS 10,000x per iteration","Matrix encoding, not ansatz, drives VQLS speedup","FWHT slashes VQLS memory 1,298x on CFD grids","Coherent VQLS beats Pauli encoding by 10,000x","SVD-based VQLS: 10,000x faster, 1,298x less memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1608,"prompt_tokens":1288,"completion_tokens":320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":904,"completion_tokens_details":{"reasoning_tokens":215}},"tokens_in":904,"tokens_out":320,"duration_ms":3026,"temperature":1.0,"reasoning_tokens":215,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:09:21.452711+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compile the SVD-derived unitaries $U_+$ and $U_-$ for an 8-qubit tridiagonal Toeplitz system into elementary gates (single-qubit rotations and CNOTs) and measure the resulting circuit depth and per-iteration time: if the depth grows exponentially with qubit count, the 10,000x per-iteration speedup disappears outside the dense-matrix simulator. A complementary test is to run the coherent SVD block-encoded VQLS on real quantum hardware at 8 qubits and compare wall-clock time per iteration with the Pauli-encoded VQLS; the speedup claim stands only if the hardware times still favor the SVD encoding.","supporting_citations":[{"cited_title":"Variational quantum linear solver,","cited_arxiv_id":null,"evidence_quote":"Defines the VQLS algorithm and cost function that the entire workflow optimizes."},{"cited_title":"Pauli decomposition via the fast walsh-hadamard transform,","cited_arxiv_id":null,"evidence_quote":"Supplies the Fast Walsh-Hadamard Transform method for Pauli decomposition coefficients that the paper parallelizes and benchmarks."},{"cited_title":"Coherent Variational Quantum Linear Solver,","cited_arxiv_id":null,"evidence_quote":"Introduces the coherent VQLS formulation with block encoding that the SVD-based encoding is combined with."},{"cited_title":"Solving the Hele–Shaw flow using the Harrow– Hassidim–Lloyd algorithm on superconducting devices: A study of efficiency and challenges,","cited_arxiv_id":null,"evidence_quote":"Provides the Hele-Shaw flow discretization and its linear system used as the canonical CFD test problem."}],"review_version":1}