{"id":"a687d31f-83c6-4f4e-ab1b-07f7c3de2f9b","arxiv_id":"2510.00145","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"CircuitTree uses tree-based Bayesian optimization with layerwise splitting to prepare quantum states with lower TVD and fewer gates, but its convergence proof is not valid as stated.","lead":"This paper introduces CircuitTree, a Bayesian-optimization method for preparing near-term quantum states by matching output distributions with fewer gates. It reports large gate/depth savings, but the accompanying convergence proof relies on assumptions that are effectively the result being proven.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.7 is not established: the Appendix proof contains an invalid inequality and analyzes full-space UCB, not Algorithm 1's layerwise EI; the claimed convergence guarantee is unsupported.","rationale":"The reader's weakest assumption correctly identified that the proof does not analyze Algorithm 1's layerwise distributed EI and that Assumption 4.3 already assumes the exploration needed for convergence. I go further and point to a concrete algebraic error in the inner proof: the chain in B.3 equates an arbitrary previously sampled point's value with the global minimum. This is not a matter of missing details; it is a false step in the central derivation. Therefore the central claim—Theorem 4.7, and the abstract's 'theoretical convergence guarantees'—is unsupported. I maintain the reader's REJECT verdict because the paper's advertised theoretical contribution, rather than merely its empirical heuristic, is the load-bearing claim that fails under scrutiny. I do not allege any dishonesty; the open-source code and empirical results are real evidence for the method's practical behavior, but they do not repair the proof. A corrected proof and fair baselines could make the paper conditional or acceptable, but as written the central claim is not established.","tokens_in":18869,"tokens_out":8731,"duration_ms":80299,"concrete_test":"Attempt to repair the proof of Theorem 4.7 by formalizing Eq. (3) with explicit quantifiers. Specifically, determine whether i(t) is chosen as the sampled point nearest to θ_t or nearest to a global optimizer θ*. If nearest to θ_t, exhibit a simple Lipschitz function (e.g., f(θ)=|θ−1| on [0,2]) and a query history where the nearest previously sampled point to a later query has f(θ_{i(t)}) > f*, demonstrating the inequality f(θ_{i(t)}) ≤ f* is false. If nearest to θ*, show how Lipschitz continuity then bounds f(θ_t), which is not possible when θ_t is arbitrarily far from θ* under the same ρ_{t−1}. Alternatively, instrument the released CircuitTree code with a 1-D or 2-D target, log the full query sequence, and verify on each iteration whether the proof's claimed chain holds; any iteration with f(θ_{i(t)}) > f* while ρ_t → 0 would settle that the theorem's proof is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is Theorem 4.7, which asserts convergence of the sequence {θ_t} produced by Algorithm 1. Even granting Assumptions 4.1–4.3 and the full-space UCB analysis in Appendix B, the proof has a concrete false step. In B.3, after defining simple regret r_t = f(θ_t) − f*, the proof picks i(t) ≤ t−1 with ‖θ_t − θ_{i(t)}‖ ≤ ρ_{t−1} and then writes:\n\nf(θ_t) ≤ f(θ_{i(t)}) + Lρ_{t−1} ≤ f* + Lρ_{t−1}.\n\nThe second inequality asserts f(θ_{i(t)}) ≤ f*, i.e., that an already-sampled point attains the global minimum. This is not guaranteed. The covering radius only implies that some sampled point near the optimum has value ≤ f* + Lρ_{t−1}; it says nothing about the value of the particular previously sampled point nearest to the current query θ_t. Thus the claimed O(t^{−1/d}) instantaneous regret bound, and hence lim E[f(θ_t)] = f*, does not follow from the covering-radius argument.\n\nThis is compounded by the proof's mismatch with the algorithm: Appendix B analyzes the full-space UCB rule of Eq. (1), while Algorithm 1 uses layerwise distributed Expected Improvement with per-layer surrogates. No argument connects these. Assumption 4.3/Lemma B.2 is also effectively an exploration assumption that already entails dense sampling, so the proof assumes much of what it claims to show. The method may still work as an engineering heuristic, but the stated convergence guarantee is not proven.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CircuitTree, a Bayesian-optimization framework for approximate quantum state preparation that minimizes total variation distance between the output distribution of a parameterized circuit and that of a target state. The method uses gradient-boosted regression trees as surrogates, a layerwise decomposition of the parameter space into independently optimized subspaces, and expected-improvement acquisition. The paper claims a convergence guarantee with an O(t^{-1/d}) rate (Theorem 4.7), and reports empirical improvements over BQSKit on IBM hardware in terms of TVD, gate count, and circuit depth. Appendices provide the convergence proof and experimental methodology.","tokens_in":19308,"tokens_out":5806,"duration_ms":51693,"significance":"If the theoretical and empirical claims were sound, CircuitTree would be a useful, scalable alternative to Gaussian-process BO for non-smooth quantum objectives, and the layerwise decomposition would be a practical way to exploit circuit structure. The paper has genuine strengths: it is open-sourced, evaluated on real hardware, and addresses a relevant problem. However, the central convergence theorem is not established as stated, and the main experimental comparison uses a baseline that solves a different optimization problem. As it stands, the contribution is best understood as an engineering heuristic with promising but inconclusive empirical support.","major_comments":[{"comment":"The inequality f(θ_t) ≤ f(θ_{i(t)}) + Lρ_{t−1} ≤ f* + Lρ_{t−1} is invalid. The covering radius only guarantees that some previously sampled point is within ρ_{t−1} of the current query θ_t; it does not guarantee that this particular point attains the global minimum. A previously sampled point can be close to θ_t in parameter space while being arbitrarily far from the global minimum in function value. Consequently, E[r_t] ≤ L (C_d D^d)^{1/d} t^{−1/d} + σ does not follow, and Theorem 4.7 and the O(t^{−1/d}) rate are unsupported.","section":"Appendix B.3, Eq. (3)"},{"comment":"The proof analyzes the full-space acquisition rule θ_{t+1} = argmin μ_t(θ) − κ_t s_t(θ) in Eq. (1), but Algorithm 1 uses layerwise distributed expected improvement: per-layer surrogates are trained, all other layers are fixed, and layer-local EI is optimized before synchronization. The remark that EI is equivalent to Eq. (1) for κ=1 is asserted without proof and is not generally correct for the quantile/ensemble uncertainty described in §3.1. No argument connects the layerwise distributed EI loop to the full-space UCB analysis. Therefore the proof does not establish convergence of the algorithm whose guarantee Theorem 4.7 claims.","section":"Appendix B vs. Algorithm 1"},{"comment":"The variance-floor assumption effectively assumes the exploration that the proof then derives. Lemma B.2 is essentially a restatement of Assumption 4.3, and its proof contains unsupported steps: s_t^2(θ̃) ≥ ν^2/M_t s_res^2 does not follow from a single empty-leaf tree, and the claim that residuals cannot all be identical is not a consequence of Assumption 4.2 (zero noise and constant f give identical y_i). Lemma B.3 assumes s_{i−1}(θ_{i−1})→0, which is not proved, and invoking a Borel–Cantelli argument for a deterministic query sequence is not meaningful. Since Lemma B.3 is the step that converts the variance floor into query density, the proof assumes much of what it claims to show.","section":"Assumption 4.3, Lemma B.2, Lemma B.3"},{"comment":"The baseline comparison is not matched. BQSKit performs approximate general unitary synthesis, whereas CircuitTree optimizes a fixed-depth ansatz against the output distribution of a target state. These are different optimization problems, so comparisons of TVD, gate count, and depth conflate the synthesis objective with the approximation objective. The reported claims — 'up to 59% lower TVD', '61% fewer CX gates', '78% shorter depth' — do not by themselves establish an advantage for approximate state preparation. A matched baseline (e.g., a variational ansatz trained by a standard optimizer, or a synthesis method constrained to the same ansatz) is needed.","section":"§6.4, Fig. 5"}],"minor_comments":[{"comment":"Line 10 updates D_t with (θ_t, y_t^(ℓ)), but θ_t is not updated after the layer-local optimization on line 8. The notation should make clear that the stored parameter vector is the full vector with the ℓ-th block replaced by θ_t^(ℓ), or the algorithm should define the new full θ_t before evaluation.","section":"Algorithm 1, line 10"},{"comment":"Eq. (1) is written as argmin of μ_t − κ_t s_t, which is a lower confidence bound (LCB), not an upper confidence bound (UCB). The text should use consistent terminology.","section":"Appendix B, Eq. (1)"},{"comment":"The text reports that GP surrogates failed to finish within five days, but no GP curve or runtime comparison appears in the figures. Please provide quantitative support or state the runtime separately.","section":"§6.1, Fig. 2"},{"comment":"The conclusion that 'layerwise optimization achieved a 2.4× reduction in convergence time and 50% lower final TVD' is stated without visible error bars or seed-level statistics in Fig. 3; the reader cannot assess variability from the figure as shown.","section":"§6.2, Fig. 3"}],"recommendation":"reject","confidential_remarks":"The theoretical appendix was not updated to match the implemented algorithm, and the key proof step is invalid. The empirical baseline mismatch compounds the problem. A revision that merely patches local details would not suffice; the authors would need to either remove the formal convergence claim and reframe the paper as an empirical heuristic study, or redesign both the algorithm-to-proof correspondence and the experimental comparison. As submitted, the central claims are not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Readable paper with a useful engineering method, but the convergence theorem does not hold up. The empirical core—GBRT surrogates with layerwise subspace splitting for approximate state preparation—is a reasonable new application, and the hardware results look credible. But the proof of Theorem 4.7 is not sound, and I would not trust the theoretical guarantees as stated.\n\nWhat's actually new: the combination of tree-based Bayesian optimization with a distributed layerwise decomposition aligned to the ansatz structure. Others have used tree-based BO and block-coordinate decomposition, but this particular packaging for approximate state preparation is new, and the open-sourced code is a plus. The experiments are mostly well executed, with real-hardware runs on ibm_nazca and a sensible fixed-depth ansatz. The comparison to BQSKit is not apples-to-apples, since BQSKit targets full unitary synthesis, so the headline TVD/gate numbers should be read with that caveat; but the comparison is still informative.\n\nWhere the paper falls short is the theory. Three problems. First, Appendix B analyzes full-space UCB (Eq. 1), not the layerwise EI algorithm actually run in Algorithm 1. No argument connects the two. Second, Assumption 4.3 is basically an exploration assumption: it stipulates a variance floor at every unqueried point, which already entails that queries spread densely. That is assuming much of what the theorem needs to show. Third, and most concretely, the proof of B.3 contains an invalid inequality. From the covering radius, for the current query θ_t you can find a previously sampled point i(t) with ||θ_t - θ_{i(t)}|| ≤ ρ_{t-1}. The proof then writes f(θ_t) ≤ f(θ_{i(t)}) + Lρ_{t-1} ≤ f^* + Lρ_{t-1}. The second inequality would require f(θ_{i(t)}) ≤ f^*, which is not guaranteed; f^* is an infimum, so sampled values are generally above it. The covering radius argument only implies that some sampled point near the optimum has value close to f^*, not the particular nearest point. So the O(t^{-1/d}) regret bound does not follow.\n\nThere is also a metadata issue: the arXiv title/abstract say one thing, the full text says another. That should be fixed.\n\nWho is this for: researchers in variational quantum algorithms or quantum compilation who want a practical, black-box optimizer for state preparation. The method may work as an engineering heuristic, and the empirical results are worth a look. But the theory as written overclaims. I'd send this to peer review—it deserves referee time—but with a clear note that the convergence claim needs either a corrected proof or a reframing as a heuristic. I would not cite the convergence theorem in my own work until the proof is fixed.","headline":"Useful empirical method for approximate state preparation, but the convergence theorem is unproven; treat CircuitTree as a heuristic, not a guaranteed optimizer.","tokens_in":19760,"tokens_out":3584,"would_cite":false,"duration_ms":28741,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"CircuitTree shows that approximate quantum state preparation can be solved with tree-based Bayesian optimization over a layerwise parameter decomposition, converging to the optimal distribution at rate O(t^(−1/d)) and empirically reducing g","keywords":["approximate quantum state preparation","Bayesian optimization","gradient-boosted regression trees","layerwise parameter decomposition","total variation distance","variational quantum circuits","hybrid quantum-classical algorithm","convergence guarantees"],"falsifier":"Run the layerwise algorithm while logging, at each round, the covering radius and the ensemble variance at points in unexplored regions. If after t iterations there exists any open ball of radius r in the parameter space containing no queried point, while t exceeds (C_d D^d / r)^d, then the geometric covering-radius bound (and the resulting O(t^(−1/d)) regret rate) is falsified. More directly: if a point assigned to an empty leaf of every tree can be found after training, the variance floor η is zero and Lemma B.2 fails.","tokens_in":18764,"feed_emoji":"⚛️","tokens_out":3211,"duration_ms":29550,"temperature":0.7,"pith_summary":"CircuitTree tackles approximate quantum state preparation: finding a short parameterized circuit whose measurement distribution matches a target state, without full state reconstruction or analytic gradients. The paper argues this is naturally a black-box optimization problem over a non-smooth, high-dimensional loss (total variation distance), and that gradient-boosted regression trees are a better surrogate model than Gaussian processes, which assume smoothness and scale badly. Its central theoretical claim is Theorem 4.7: under Lipschitzness, bounded noise, and a variance-floor condition, the optimization converges to the optimal distribution, with regret shrinking at rate O(t^(−1/d)). Empirical results on random circuits, amplitude-encoded states, and VQE workloads show 2–3× lower total variation distance while using 40–60% fewer two-qubit gates than existing synthesis approaches.","feed_headline":"Tree-surrogate optimizer cuts quantum gate counts by 60%","feed_subtitle":"Bayesian optimization with regression trees matches target distributions with fewer, shallower circuits and a provable convergence rate.","key_machinery":"The central mechanism is the combination of a gradient-boosted regression tree (GBRT) surrogate, which captures non-smooth discontinuities and scales linearly with samples, and a layerwise parameter decomposition that splits the circuit parameters into blocks aligned with the hardware-native layered ansatz. Each layer is optimized in parallel by its own local surrogate, with periodic synchronization of a shared global parameter vector. The theoretical engine is a variance-floor argument: unqueried points must fall in an empty leaf of at least one tree, guaranteeing ensemble variance remains positive and driving the covering radius of the query set to zero; Lipschitz continuity then yields th","core_discovery":"The paper's central claim is that approximate quantum state preparation can be reliably performed by a surrogate-guided, tree-based Bayesian optimization framework that exploits the layered architecture of variational circuits. The key theoretical result states that, under three assumptions — Lipschitz continuity of the loss, bounded centered noise, and a variance floor at all unqueried points — the sequence of parameters produced by the layerwise distributed algorithm satisfies lim E[f(θ_t)] = f* in the noise-free case, with a convergence rate of O(t^(−1/d)). The proof works by showing that the variance floor forces the query sequence to become dense in the parameter space, after which Lips","pith_inferences":["Inference: The convergence proof analyzes a full-space UCB acquisition rule, while the algorithm uses layerwise expected improvement; if the layerwise loop introduces correlations across layers that break the variance-floor assumption, the O(t^(−1/d)) rate may not hold for the implemented method — a direct empirical check is to monitor the per-layer covering radius.","Inference: The O(t^(−1/d)) rate is slow in high dimensions, so the practical advantage of layerwise decomposition is not just parallelism but a drastic reduction of effective dimensionality; this suggests further gains from hierarchical or additive structure in the circuit parameter space.","Inference: The same surrogate-guided, structure-aware loop could be extended beyond total variation distance to other distributional objectives, such as Wasserstein distances or quantum fidelity, and to continuous-variable circuits whose parameter spaces have a similar layered structure.","Inference: If the variance-floor assumption is intentionally violated, for example by forcing all trees to share similar splits so that some regions have zero variance, the exploration argument collapses and queries will cluster, yielding a concrete experimental falsifier of the convergence claim."],"forward_implications":["Approximate state preparation can be performed without access to analytic gradients or full unitary descriptions, relying only on measurement statistics, which is directly applicable to noisy near-term hardware.","The layerwise distributed optimization strategy gives a principled form of block-coordinate descent that is more stable than random subspace splitting and preserves global convergence guarantees.","Modest shot budgets (around 250 measurements) suffice for stable convergence, meaning the approach is practical under hardware measurement constraints.","In noise-free settings, the convergence guarantee implies the algorithm is no-regret, with cumulative regret growing as O(T^(1−1/d)), and the rate sharpens to O(log T) in one dimension.","The framework provides the first stated provable convergence guarantees for approximate state preparation using non-Gaussian surrogates, opening a path for other structured black-box quantum optimization problems."],"fun_headline_variants":["Quantum circuit optimizer cuts gates 60% with Bayesian trees","Tree-based Bayesian optimization shrinks quantum circuits","CircuitTree: 60% fewer gates for quantum sampling","Surrogate-guided quantum sampling uses 60% fewer gates","Bayesian tree optimizer speeds quantum distribution matching"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The convergence proof depends on the assumption that every parameter point the optimizer has not yet queried falls into an empty leaf of at least one tree, so the surrogate's variance never collapses to zero in unexplored regions; the paper does not prove that its layerwise distributed algorithm actually enforces this condition.","fun_headline_variants_meta":{"raw":{"variants":["Quantum circuit optimizer cuts gates 60% with Bayesian trees","Tree-based Bayesian optimization shrinks quantum circuits","CircuitTree: 60% fewer gates for quantum sampling","Surrogate-guided quantum sampling uses 60% fewer gates","Bayesian tree optimizer speeds quantum distribution matching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000144,"raw_usage":{"total_tokens":983,"prompt_tokens":688,"completion_tokens":295,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":220}},"tokens_in":432,"tokens_out":295,"duration_ms":3072,"temperature":1.0,"reasoning_tokens":220,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T13:27:42.520252+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the layerwise algorithm while logging, at each round, the covering radius and the ensemble variance at points in unexplored regions. If after t iterations there exists any open ball of radius r in the parameter space containing no queried point, while t exceeds (C_d D^d / r)^d, then the geometric covering-radius bound (and the resulting O(t^(−1/d)) regret rate) is falsified. More directly: if a point assigned to an empty leaf of every tree can be found after training, the variance floor η is zero and Lemma B.2 fails.","supporting_citations":[],"review_version":1}