{"id":"09c3b8c9-82fa-46af-9784-5983b8d07f1c","arxiv_id":"2501.08794","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A variational quantum algorithm with custom readout mitigation solves simplified TARGET2-Securities settlement instances of up to 40 trades, but real hardware runs remain noise-limited and do not beat random sampling.","lead":"This paper tests a variational quantum algorithm on small instances of the securities settlement problem used by TARGET2-Securities, running up to 40 transactions on IBM superconducting hardware. The hardware runs are noise-limited and often no better than random sampling, while a classical simulation of the same ansatz solves all tested instances.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 can return a collateral vector y that violates the securities-balance constraint (2b), so Eq. (8) may not evaluate true NTSP feasibility of candidate solutions x.","rationale":"The reader's weakest_assumption is that Algorithm 1 computes a correct (feasible and near-optimal) y for each proposed x. I agree and sharpen the concern: the greedy heuristic can over-pledge, producing an infeasible y with respect to Eq. (2b) even when a feasible y exists. This is more directly load-bearing than missing a higher-value collateral assignment, because it corrupts the feasibility evaluation of candidate solutions, not merely their objective value. If the heuristic rejects valid x, then the optimizer's search is biased and the comparison against CPLEX (which solves the true problem) is not a fair measure of the solver's performance on the actual NTSP. The paper is otherwise transparent about hardware noise and the random-sampling behavior of QTSA, which the reader already weighed into the CONDITIONAL verdict. My concern reinforces that conditional status: it asks for a concrete verification that the benchmark instances are not adversely affected by a suboptimal or infeasible greedy y. Since the reader already identified this as the weakest assumption and recommended a conditional verdict, the verdict should remain unchanged. The proposed test—an exact CPLEX feasibility check for y on the same x samples—would settle whether the concern actually lands, distinguishing a merely theoretical flaw from one that affects the reported results.","tokens_in":22515,"tokens_out":9001,"duration_ms":95586,"concrete_test":"For each of the five instances in Table 2, collect every distinct x evaluated during the reported runs of QTSA, Q-INSP, and SAMPLER (the data should be regenerated or logged by the authors). For each such x, solve the exact feasibility subproblem for y using CPLEX: fix x and optimize (or just check feasibility of) the integer variables y subject to Eqs. (2), (4), and the domain constraints, with the same parameters used in the paper. Compare against Algorithm 1's y: (i) count how many x have a feasible exact y but an infeasible greedy y (i.e., greedy violates Eq. (2b) or fails to raise enough credit), and (ii) for x where both are feasible, compare the maximum achievable credit. If any instance has even one such x among the sampled set, the objective in Eq. (8) is demonstrably corrupted; if none, the concern is empirically refuted for these benchmarks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that QTSA solved NTSP instances up to 35 transactions depends on the objective C(x, y_x) in Eq. (8) faithfully representing the NTSP of Sec. 2.2. The vector y_x is computed by the greedy heuristic in Algorithm 1 (App. A). This heuristic fills each SPL link to the maximum allowed by the cash limit alim_tmp and the on-flow quantity limit qlim_l, but it never caps the pledged lots by the securities actually available in the position after settling x. The variable valp is read in the loop but is not used to upper-bound lotmax. Consequently, the output y can pledge more securities than remain in the security position, violating the non-negativity constraint in Eq. (2b). Since Eq. (2b) is one of the penalized constraints in Eq. (7) (see Table 3), such an x receives a positive penalty even though a smaller y would make it feasible. The optimizer is then discarding valid settlement proposals, and the function being maximized is not the true NTSP objective. The Non-Shared Collateral assumption (Table 1) does not prevent this, because multiple SPL links can share the same cash limit alim_m and over-pledging can still drain a single security position below zero. The paper provides no proof, and no numerical check, that for the benchmark instances in Table 2 the greedy y is feasible whenever any feasible y exists. This affects QTSA, Q-INSP, and SAMPLER equally, but it undermines both the comparison with CPLEX and the interpretation of the reported results as 'solving the NTSP'.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formalizes a simplified securities transaction settlement problem (NTSP) as a linear integer program, reformulates it as an unconstrained objective with nonlinear activation-function penalties, and benchmarks a variational quantum algorithm (QTSA) on IBM superconducting devices against a quantum-inspired solver (Q-INSP) and a random sampler (SAMPLER) on real-world-inspired instances of 20 to 40 transactions. It reports average payoff ratios relative to CPLEX, analyzes convergence with and without the iHAMMER readout-mitigation technique, and concludes that NTSP instances with up to 35 transactions were solved by QTSA on hardware.","tokens_in":22920,"tokens_out":7417,"duration_ms":72818,"significance":"If the claims are correct, the work would provide the largest demonstration of a quantum solver for a practical securities-settlement problem on current hardware, together with a penalty-encoding scheme and a heuristic readout mitigation. The experimental section is transparent about configurations and hyperparameters, and the paper explicitly acknowledges that QTSA behaves like a random sampler on smaller instances. However, the load-bearing feasibility of the greedy collateral heuristic is unproven and appears violated by the algorithm as written, which compromises the interpretation of the benchmark and the 'solved' claims.","major_comments":[{"comment":"The greedy heuristic for computing y reads the variable valp (final balance of security position p after settling x) but never uses it to bound lotmax; lotmax is capped only by the remaining credit limit alim_tmp and the on-flow quantity limit qlim_l. As a result, the returned y can pledge more securities than are available in position p, violating Eq. (2b). The Non-Shared Collateral assumption does not prevent this, since a single link can over-pledge its own position. Because Eq. (2b) is penalized in the objective Eq. (7) (Table 3), the optimizer can discard a feasible x due to an inflated penalty, or report C(x,yx)>0 for an x whose y is actually infeasible. The paper provides no proof and no numerical check that for the Table 2 instances the greedy y is feasible whenever a feasible y exists. This affects QTSA, Q-INSP, and SAMPLER equally and invalidates the comparison with CPLEX, which solves the true constraint set. Please fix Algorithm 1 (e.g., cap lotmax by valp), re-run or re-analyze the benchmarks, and add a feasibility verification for all reported solutions.","section":"Appendix A, Algorithm 1; Eqs. (7)-(8)"},{"comment":"The paper's own data show that QTSA and SAMPLER are statistically compatible for instances with up to 35 transactions (average payoff ratios and standard deviations overlap), and Sec. 4.3 concludes that QTSA behaves similarly to a random sampler. Yet the conclusion states that 'NTSP instances with up to 35 transactions were solved by QTSA on gate-based superconducting quantum hardware.' This is an overstatement: the reported evidence supports at best that QTSA matches random sampling on these instances, not that it provides a systematic solution. The claim should be qualified (e.g., 'found feasible solutions at rates comparable to a random sampler') or supported by a formal comparison showing superiority over SAMPLER.","section":"Sec. 4.2, Fig. 5; Sec. 5"}],"minor_comments":[{"comment":"The variable lotmax is used inconsistently with Eq. (4b): qmin_l is a quantity of securities, while lotmax appears to be a number of lots. Please clarify the units and conversion between lots and quantities.","section":"Appendix A, Algorithm 1"},{"comment":"The text reads 'Q-ISNP can approximate it well'; this appears to be a typo for 'Q-INSP'.","section":"Sec. 4.2, paragraph after Fig. 5"},{"comment":"The statement that the total number of possible solutions 2^n is 'sufficiently small to guarantee' that QTSA and SAMPLER outperform Q-INSP for n<=35 is questionable: 2^35 is about 3.4e10, and no guarantee follows from the size of the search space alone. Please rephrase to 'small enough that random sampling can be competitive'.","section":"Sec. 4.2, paragraph on n<=35"},{"comment":"The notation P_{n,sigma}(x <= ceil(log2 delta)) uses x both as the binomial random variable and as the settlement decision vector used throughout the paper; please use a different symbol for the random variable.","section":"App. D.2, Eq. (12)"},{"comment":"The fidelity is plotted as an average over 10 runs, but the text mentions that 5 of the 30-qubit runs were taken at different times with different noise conditions; indicating per-run noise conditions or separate markers would help interpret the bump at n=30.","section":"Sec. 4.1, Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is an honest engineering benchmark, but the central claim needs strengthening. The feasibility issue in Algorithm 1 is concrete and fixable, yet the re-evaluation may change the numerical results. The conclusion also overstates the hardware result relative to the paper's own random-sampler comparison. I recommend major revision with the expectation that the authors either fix the heuristic and re-run the benchmarks or significantly temper the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is better than the usual quantum-finance hype: it reports real hardware runs, openly says QTSA matches random sampling for up to 35 trades, and gives a fairly complete picture of the noise limitations. The new pieces are the nonlinear activation penalty for inequality constraints, the parameter mapping G, and the iterative iHAMMER readout mitigation. Those are worth a look. The experiments on IBM devices with up to 40 qubits are also a reasonable stress test of current NISQ hardware.\n\nThe main soft spots are two. First, the greedy heuristic in Algorithm 1 is not constrained by the securities actually available in the position after settling x. It reads valp but never uses it to cap lotmax, so it can over-pledge and violate Eq. (2b). Since Eq. (2b) is a penalized constraint, the optimizer can throw away an x that would be feasible with a smaller y. That means the objective in Eq. (8) is not always the true NTSP objective. The stress-test note is correct on this, and the Non-Shared Collateral assumption does not save it. The paper gives no proof or numerical check that the greedy y is feasible for the benchmark instances, so the CPLEX comparison is not on solid ground.\n\nSecond, the conclusion says QTSA 'solved' instances up to 35 transactions, but the data show QTSA and SAMPLER are statistically indistinguishable there. That is a stretch. What QTSA actually demonstrated is that the variational pipeline can find valid solutions roughly as well as random sampling on small instances. The honest language in Sec. 4.3 is much better than the conclusion.\n\nThe lack of code and data also hurts reproducibility. The activation hyperparameters are hand-chosen, but that is normal engineering and not a flaw by itself.\n\nWho is this for? People working on NISQ optimization for financial settlement, or on benchmark studies of VQAs on real hardware. It is not a breakthrough, but it is a solid empirical data point. I would send it to peer review, but require a major revision: fix or rigorously bound the greedy y, revisit the 'solved' claim, and ideally publish the code and benchmark instances.\n\nMy own verdict would be conditional acceptance after that work.","headline":"Honest empirical VQA study with a genuinely new NTSP formulation and iHAMMER mitigation, but the 'solved on hardware' claim is overstated and the greedy collateral heuristic can break the very feasibility constraint it is supposed to enforce.","tokens_in":23444,"tokens_out":2097,"would_cite":false,"duration_ms":22379,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","90C10"],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"A variational quantum algorithm can find valid settlements for batches of up to 35 real transactions on today's superconducting hardware, and up to 40 in noiseless simulation.","keywords":["variational quantum algorithms","securities settlement","TARGET2-Securities","NTSP","mixed integer programming","constraint encoding","readout error mitigation","Bayesian optimization"],"falsifier":"Construct a small NTSP instance where the greedy collateral heuristic can be compared against exact enumeration: compute y* by solving the full mixed-integer problem and y_g by Algorithm 1 for every possible x. If any x admits a feasible y* with C(x, y*) > C(x, y_g), then the cost function in Eq. (8) does not match the true NTSP objective, and the reported payoff ratios are measuring a different problem.","tokens_in":22346,"feed_emoji":"⚛️","tokens_out":6907,"duration_ms":69788,"temperature":0.7,"pith_summary":"The paper claims that the Night-time Transaction Settlement Problem (NTSP) used by TARGET2-Securities can be recast as an unconstrained optimization and attacked with a variational quantum algorithm on today's superconducting hardware. The key move is to fold cash, securities, collateral, and after-link constraints into the cost function using smooth non-linear penalty functions, so the objective is a diagonal observable over bitstrings representing which transactions settle. On real quantum machines the algorithm QTSA found valid settlements for instances with up to 35 transactions, which the authors state is the largest NTSP instance solved on quantum hardware to date; its noiseless quantum-inspired counterpart Q-INSP handled 40. The authors also introduce an iterative readout-error mitigation routine, iHAMMER, and show that with it a 20-transaction instance can exhibit convergent optimization on actual hardware. The motivating claim is that, despite current noise, small realistic settlement batches can be tackled by variational quantum optimization, and the same pipeline transfers to other mixed-integer programs.","feed_headline":"Quantum method solves 35-trade settlement on real hardware","feed_subtitle":"A variational pipeline with custom noise mitigation also settles 40 trades classically, beating earlier quantum demos.","key_machinery":"The load-bearing object is the cost function C(θ) = Σ_x p̃_x(θ) C(x, y_x), built from the payoff F(x) plus penalty terms h_v(g_v(x,y)+b_v) for each constraint. It is a diagonal Ising-type observable, so no Pauli decomposition is needed; the quantum computer outputs bitstrings and a classical loop evaluates the cost per bitstring. Two ancillary mechanisms make this work in practice: a Gaussian parameter mapping G(θ_i) that biases the hardware-efficient ansatz toward sparse states so that p̃_x is close to p_x, and an iterative Hamming Reconstruction mitigation (iHAMMER) that redistributes probability from isolated noisy outcomes based on a calibration run. The classical optimizer is Bayesian optimization, and for each sampled x the integer collateral variables y_x are produced by a greedy heuristic rather than by the quantum circuit.","core_discovery":"The central discovery is a complete variational pipeline for the NTSP. The model uses binary variables x for settled transactions and integer variables y for collateral lots, with constraints encoded through parameterized non-linear activation functions. Because the resulting cost is diagonal in the computational basis, the quantum computer only needs to sample candidate bitstrings, and the expected cost is estimated from reconstructed probabilities. A greedy heuristic computes the collateral lot count for each sampled settlement, and a Gaussian parameter mapping makes the output state sparse so that the probabilities can be estimated with a polynomial number of shots. The paper demonstrates that on current gate-based superconducting hardware, QTSA yields valid solutions for instances up to 35 transactions, and that iHAMMER readout mitigation enables convergent optimization for 20-transaction instances, while the noiseless Q-INSP solves all tested batches up to 40 transactions.","pith_inferences":["Beyond the paper: a natural next step not pursued here is to embed this variational solver inside T2S's existing night-time heuristic, either as a warm-start or as a fallback for batches where the classical heuristic is known to be suboptimal; the 45-minute run 4 window is at least compatible with the iteration budgets used in the experiments.","The activation-function penalty encoding is not limited to finance: the same recipe could encode capacity, precedence, and collateral-style inequalities in scheduling or logistics problems whose objective is diagonal over bitstrings and therefore amenable to this sampling-based evaluation.","iHAMMER's calibration-driven iterative filtering could be tested as a generic readout mitigation for any variational algorithm that estimates a diagonal observable from bitstring samples, independent of the NTSP.","A clean diagnostic for whether hardware noise or the greedy collateral heuristic is the true bottleneck would be to run Q-INSP on the 40-transaction instance while replacing the greedy y computation with an exact mixed-integer solve; if performance still degrades, the heuristic is the limiting component rather than shot noise."],"forward_implications":["The same variational pipeline can be applied to other optimization problems modelable as mixed-integer programs, since inequality constraints are handled by generic non-linear penalty functions rather than problem-specific Pauli decompositions.","On current hardware, QTSA solves NTSP instances up to 35 transactions, and the authors state this surpasses the size of NTSP instances addressed in previous quantum works.","The noiseless Q-INSP variant solves all tested batches including 40 transactions, so if hardware noise and mitigation quality improve, the same architecture could plausibly scale to 40 and beyond on a device.","The demonstration that iHAMMER enables convergent Bayesian optimization for 20-transaction instances on real hardware suggests that tailored readout mitigation can extend the reach of sampling-based variational algorithms beyond what raw device fidelity allows.","Hardware noise currently limits scalability: for instances with up to 35 transactions, QTSA's payoff is statistically compatible with random sampling, so the method's practical value on larger T2S batches depends on further reductions in noise or more effective mitigation."],"supporting_citations":[{"why":"Provides the NTSP formalization and securities-and-cash settlement framework on which the model in Sec. 2.2 is built.","marker":"[17]"},{"why":"Defines the T2S functional constraints (run 4, after links, auto-collateralization, 45-minute window) that the model encodes.","marker":"[18]"},{"why":"Supplies the T2S settlement objectives (volume, amounts, priorities) used to construct the payoff function F(x).","marker":"[33]"},{"why":"Describes the earlier quantum algorithm for mixed-binary transaction settlement, which this work extends and surpasses in instance size.","marker":"[13]"},{"why":"Introduces the qubit-efficient encoding for NTSP that this paper compares against and whose sampling-overhead limitation motivates the present approach.","marker":"[14]"},{"why":"Establishes that shot noise and hardware noise make some VQAs behave like random samplers, the interpretation used for QTSA's payoff statistics.","marker":"[25]"},{"why":"Introduces the Hamming Reconstruction method that iHAMMER extends with an iterative, calibration-based stopping rule.","marker":"[62]"},{"why":"Provides the Bayesian optimization algorithm used as the classical optimizer for the variational loop.","marker":"[65]"},{"why":"The mixed-integer solver used to compute optimal payoffs, against which the reported ρ ratios are measured.","marker":"[73]"}],"fun_headline_variants":["Quantum variational pipeline settles 35 trades on real hardware","Variational algorithm clears 35 trades on IBM quantum hardware","Quantum solver tackles 35-trade settlement with noise mitigation","Custom noise mitigation lets quantum pipeline settle 35 trades","Superconducting quantum method optimizes settlement for 35 trades"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For any proposed settlement x, the greedy collateral routine is assumed to find the best feasible collateral assignment y; if it returns a suboptimal y, the cost function misranks candidate solutions and the entire optimization is being driven by a corrupted objective.","fun_headline_variants_meta":{"raw":{"variants":["Quantum variational pipeline settles 35 trades on real hardware","Variational algorithm clears 35 trades on IBM quantum hardware","Quantum solver tackles 35-trade settlement with noise mitigation","Custom noise mitigation lets quantum pipeline settle 35 trades","Superconducting quantum method optimizes settlement for 35 trades"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000527,"raw_usage":{"total_tokens":2474,"prompt_tokens":804,"completion_tokens":1670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":420,"completion_tokens_details":{"reasoning_tokens":1590}},"tokens_in":420,"tokens_out":1670,"duration_ms":12201,"temperature":1.0,"reasoning_tokens":1590,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:17:11.601162+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a small NTSP instance where the greedy collateral heuristic can be compared against exact enumeration: compute y* by solving the full mixed-integer problem and y_g by Algorithm 1 for every possible x. If any x admits a feasible y* with C(x, y*) > C(x, y_g), then the cost function in Eq. (8) does not match the true NTSP objective, and the reported payoff ratios are measuring a different problem.","supporting_citations":[{"cited_title":"Securities and Cash Settlement Framework","cited_arxiv_id":null,"evidence_quote":"Provides the NTSP formalization and securities-and-cash settlement framework on which the model in Sec. 2.2 is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the T2S functional constraints (run 4, after links, auto-collateralization, 45-minute window) that the model encodes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the T2S settlement objectives (volume, amounts, priorities) used to construct the payoff function F(x)."},{"cited_title":"Quantum algorithms for mixed binary optimization applied to transaction settlement","cited_arxiv_id":null,"evidence_quote":"Describes the earlier quantum algorithm for mixed-binary transaction settlement, which this work extends and surpasses in instance size."},{"cited_title":"Ex- ponential qubit reduction in optimization for financial transaction settlement","cited_arxiv_id":null,"evidence_quote":"Introduces the qubit-efficient encoding for NTSP that this paper compares against and whose sampling-overhead limitation motivates the present approach."},{"cited_title":"Challenges of variational quantum opti- mization with measurement shot noise","cited_arxiv_id":null,"evidence_quote":"Establishes that shot noise and hardware noise make some VQAs behave like random samplers, the interpretation used for QTSA's payoff statistics."}],"review_version":1}