{"id":"50f137c3-f835-4fc2-94fa-7593b13cdc6d","arxiv_id":"2505.09729","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A state-space gradient descent algorithm with ancilla-based non-unitary updates prepares ground or metastable states of quantum Hamiltonians and is claimed to avoid barren plateaus.","lead":"The paper proposes a variational quantum algorithm that descends the energy landscape using local system and ancilla operations, and that at convergence returns either the ground state or a physically meaningful metastable state. A generalist might read it as a concrete recipe for getting useful output from near-term quantum computers even when exact ground-state preparation fails.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The no-barren-plateau claim rests on equating the SSGD brickwall gates e^{-iθP} (fixed Pauli P, random θ) with Haar-random two-qubit unitaries; that equivalence is asserted without proof and is not true for the raw gate distribution.","rationale":"The reader and I converge on the same weak point: the transfer from [33] to the SSGD circuit depends on an unverified Haar equivalence for Pauli rotations. I regard this as the single most load-bearing issue because the abstract's 'we show that our algorithm does not suffer from the barren plateau problem' is one of only two sharp theoretical claims; Lemma 2 and Definition 1 are local and correct, and the numerics are consistent with the mechanism but do not independently establish scaling. The check is cheap and decisive: if the theorem's assumptions are not met by Pauli rotations, the proof collapses; if they are met, the paper needs only to replace the misleading Haar sentence with the correct hypothesis. I therefore do not move the verdict: the paper remains CONDITIONAL, pending the check of [33] and/or the variance simulation. The lack of code/data and the absence of a convergence proof are real but secondary; they do not by themselves overturn the algorithmic proposal.","tokens_in":12446,"tokens_out":12530,"duration_ms":132719,"concrete_test":"Open Ref. [33], Definition 3 and Theorem 1, and determine whether they assume each two-qubit gate is Haar-distributed (or forms a 2-design) or only that each gate is e^{-iθP} for a Pauli P with θ drawn from a continuous distribution. Independently of that reading, simulate the SSGD brickwall circuit for the 1D TFIM/XXXX Hamiltonian at N=6,8,10,12 qubits with θ uniform and compute Var_θ[Tr(Hρ(θ))] for (i) gates e^{-iθP} from the specified G_A/G_S and (ii) gates drawn independently from Haar over SU(4). If the Pauli-rotation variance decays with N while the Haar variance stays constant, the no-barren-plateau claim is refuted; if the two variances agree to within statistical error, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The advertised theoretical guarantee beyond local optimality is the absence of barren plateaus (Abstract and Sec. II.A). The entire argument is the sentence 'all parameters ... chosen randomly, such that each two-qubit unitary is effectively sampled from the Haar measure over SU(2)', followed by an application of [33, Theorem 1]. For a fixed Pauli P, U(θ)=e^{-iθP} is supported on a one-parameter subgroup of SU(4), so it is not Haar distributed; its second moments do not match Haar averages. The word 'effectively' carries all the weight, but no approximation error, moment-matching condition, or t-design property is stated. If [33, Theorem 1] requires Haar or 2-design moments, the variance lower bound does not follow. If it only requires continuous Pauli rotations, the paper should cite that condition instead of claiming Haar equivalence. A secondary gap: even a valid random-parameter variance bound would concern randomly initialized circuits, while SSGD's parameters are adaptively chosen by gradient/Hessian updates, so the link between the bound and the algorithm's trainability is not made.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes State-Space Gradient Descent (SSGD), a variational algorithm that iteratively reduces the energy of a quantum state by applying local unitary and non-unitary (via an ancilla) operations chosen from gradient and Hessian information. The central theoretical claims are: (i) a local-minimum state in the sense of Definition 1 satisfies both first- and second-order optimality conditions and, by Lemma 2, is also stable against local Lindbladian perturbations; (ii) the algorithm is provably free from barren plateaus because a brickwall circuit variant has energy variance lower bounded by a constant (Sec. II.A); and (iii) numerical simulations on the 1D transverse-field Ising model and a Rydberg atom chain show that the algorithm converges to either the ground state or a physically meaningful metastable state. The paper includes pseudocode for Algorithm 1, two technical lemmas in Appendix A, and numerical comparisons of dissipative versus purely unitary updates.","tokens_in":12688,"tokens_out":2918,"duration_ms":30568,"significance":"If the central claims hold, the algorithm would be a useful contribution to variational quantum optimization: it targets physically meaningful local minima rather than arbitrary ansatz artifacts, and it offers a concrete, near-term-friendly procedure for preparing metastable states. The use of ancilla-mediated operations to escape spurious local minima is well motivated by the cited Haar-random state result. The numerical demonstrations for TFIM and Rydberg chains provide initial evidence for the metastable-state claim. The paper also benefits from a clean separation: Lemma 1 and Lemma 2 are elementary and appear correct, and the algorithm's fixed points satisfy the stated optimality conditions by construction. However, the advertised barren-plateau guarantee rests on an unproved and questionable equivalence between the actual SSGD gates and Haar-random two-qubit unitaries, and the convergence of Algorithm 1 is asserted without proof; these are load-bearing gaps for the paper's headline claims.","major_comments":[{"comment":"Algorithm 1 does not specify how the Hessian K is measured or how many samples are used. The paper states that the gradient noise variance is chosen as σ_j^2 = δt_S, but no similar prescription is given for the Hessian noise, which is important because the ancilla direction uses the sign of the Hessian eigenvalues. The threshold Etol in Eq. (6) is a new free parameter, and the paper does not discuss how to set it in experiments or how sensitive the algorithm is to this parameter.","section":"Sec. II.B / Algorithm 1"}],"minor_comments":[{"comment":"There are several typos and grammatical issues (e.g., 'oftentimes' used repeatedly, missing articles, and inconsistent use of “the algorithm” vs “the SSGD algorithm”). A careful proofreading pass is recommended.","section":"Global"}],"recommendation":"major_revision","confidential_remarks":"The paper has a sound core: Lemma 1 and Lemma 2 are correct, the algorithm's fixed-point conditions are well defined, and the numerical experiments are a useful first demonstration. However, the barren-plateau claim is the headline theoretical result and it currently rests on an unverified equivalence to Haar-random gates. The authors should either prove the variance bound directly for Pauli rotations or restrict the claim to a clearly stated assumption. Additionally, the convergence of Algorithm 1 is not analyzed, which is a gap for a paper that claims a convergence guarantee. The numerical evidence for metastability is suggestive but would benefit from quantitative lifetimes and fidelity checks. With these revisions the paper could be suitable for publication, but as written the central theoretical guarantees are overstated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me skip the small talk. The paper describes a variational algorithm (SSGD) that does something genuinely useful: it uses ancilla-assisted non-unitary updates chosen from Hessian information, and at convergence it produces a state that is a local minimum under both unitary and Lindbladian perturbations. Lemma 2 is the real clean result: Hessian PSD implies first-order Lindbladian stability, so you don't need to simulate dissipative dynamics. The numerics on TFIM and Rydberg chains are honest, including cases where the ancilla method is slower than purely unitary descent, and the final states do cluster around known ground and metastable energies. Credit where due: the algorithm design is new in its specifics, and the paper does not oversell the metastability link—it explicitly says the connection is numerical.\n\nThe soft spot is in the no-barren-plateau proof. The sentence 'such that each two-qubit unitary is effectively sampled from the Haar measure over SU(2)' carries the entire argument. For a fixed Pauli P, e^{-iθP} is a one-parameter subgroup of SU(4), not a Haar-random gate; its moments do not match Haar. Maybe the cited theorem in [33] only needs continuous Pauli rotations, but the paper does not say that. As written, the variance lower bound does not follow. And even if it did, it would be a statement about random initialization, while SSGD's parameters are adaptively chosen by gradient and Hessian updates. The link between the bound and the algorithm's trainability is missing. This is a load-bearing flaw in the headline claim, but it is fixable: either prove a moment condition for the actual gate distribution or cite a theorem that covers it.\n\nTwo smaller gaps: Algorithm 1 has no convergence proof, and the numerics are small-scale with no code or data released, so reproducibility is limited. These are minor relative to the BP issue.\n\nUnless I am misreading [33], the central contribution—the algorithm and Lemma 2—holds up; the no-BP claim as stated does not. The paper deserves a serious referee, and I would send it out. I would ask for a major revision on the BP argument and a gentle request for convergence statements and code/data. I would not cite the no-BP result in my own work until it is fixed, but I would cite the algorithm and the Lindbladian lemma.","headline":"The algorithm and the Lindbladian lemma are the real content; the no-barren-plateau proof rests on an unsupported Haar-random assumption and should be softened or fixed.","tokens_in":13197,"tokens_out":2702,"would_cite":false,"duration_ms":27936,"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":"This paper introduces state-space gradient descent (SSGD), a variational quantum algorithm that iteratively lowers the energy of a quantum state using ancilla-assisted local operations, and claims that at convergence it reaches a…","keywords":["variational quantum algorithm","state-space gradient descent","metastable states","barren plateaus","Lindbladian perturbations","local-minimum states","transverse-field Ising model","Rydberg atom arrays"],"falsifier":"Compute the energy variance of the SSGD circuit for increasing system size N using the actual gates $e^{{-iθP}}$ with θ drawn from the same distribution used in training; if the variance decreases exponentially with N, the claimed barren-plateau avoidance fails. Alternatively, run SSGD on a 1D TFIM with a larger system and check whether the final-state energies continue to cluster at ground/metastable energies rather than at generic local minima.","tokens_in":12206,"feed_emoji":"⚛️","tokens_out":6503,"duration_ms":55411,"temperature":0.7,"pith_summary":"This paper introduces state-space gradient descent (SSGD), a variational quantum algorithm that iteratively lowers the energy of a quantum state using ancilla-assisted local operations. The central claim is that at convergence the algorithm reaches a local-minimum state: no local operation can reduce the energy to first order, and the Hessian is positive semidefinite. Moreover, such a state is provably stable against small Lindbladian (dissipative) perturbations, so it is either the true ground state or a physically meaningful metastable state. The paper also argues that a brickwall version of the circuit has energy variance bounded below by a constant, so it does not suffer from barren plateaus. Numerical simulations on the 1D transverse-field Ising model and a Rydberg atom chain show that SSGD consistently terminates at either the ground state or a known long-lived metastable state.","feed_headline":"New energy-descent method lands on ground or metastable states","feed_subtitle":"No barren plateaus; when it misses the ground state, it still lands on a long-lived metastable state.","key_machinery":"The central object is the local-minimum state defined by vanishing first-order energy gradients and a positive semidefinite Hessian with respect to the generator set G. The load-bearing identity is Lemma 2's equivalence between this second-order condition and first-order Lindbladian optimality, proved by writing a local Lindbladian L = A + iB and embedding it as a Hermitian G on one ancilla plus k−1 system qubits, so that Tr(L(ρ)H) = (1/2) α† K α. The no-barren-plateau mechanism is the brickwall circuit with alternating GA and GS layers and ancilla reset, which makes the circuit a dynamically parameterized circuit with small lightcone, giving a constant lower bound on the energy variance.","core_discovery":"The paper's core discovery is that a variational algorithm can be designed so that its failure mode is physically informative rather than arbitrary. SSGD maintains a state on the system register plus a single ancilla; each iteration applies a unitary generated by local Pauli operators from a set G = GA ∪ GS, then resets the ancilla, giving an effective non-unitary operation on the system. The algorithm chooses update directions from the energy gradient (for system generators) and from the low-lying directions of the Hessian (for ancilla generators). A state is a local-minimum state (Definition 1) if the first-order gradient vanishes and the Hessian is positive semidefinite; Lemma 2 shows this second-order condition implies the state is also first-order stable against any local Lindbladian generator, so the local minima are robust to physically realizable dissipative perturbations. The authors further prove that a brickwall-structured version of SSGD has energy variance lower bounded by a constant (via [33, Theorem 1]), ruling out barren plateaus. In numerics, the final states cluster around the ground state energy or the metastable state energy in both the TFIM and the Rydberg chain.","pith_inferences":["The paper's Lemma 2 may extend beyond the first-order Lindbladian check: the same embedding argument could yield a rigorous lower bound on the lifetime of SSGD's local-minimum states under weak dissipation, effectively converting the numerical metastability evidence into a proof for certain models.","The Haar-random assumption in the barren-plateau argument could be tested and possibly replaced by a weaker assumption using Weingarten calculus for one-parameter subgroups; if the variance bound holds for the actual gate distribution, the result becomes fully rigorous for SSGD rather than for a neighboring circuit family.","Since SSGD only requires gradient and Hessian estimates, a classical tensor-network implementation of the same state-space descent might locate metastable states in larger systems, giving a quantum-inspired classical algorithm for metastability.","The choice of Hessian eigenvector direction for ancilla generators resembles second-order optimization; one could connect it to quantum natural gradient and possibly show improved convergence rates."],"forward_implications":["If SSGD works as claimed, near-term quantum devices can prepare not only ground states but also metastable states by choosing the initial state, which is useful for studying false-vacuum decay, prethermalization, and quantum memory.","The algorithm's convergence to a Lindbladian-stable local minimum means the final state is robust against weak dissipation, a property that matters for any state preparation routine on noisy hardware.","The no-barren-plateau guarantee for the brickwall circuit suggests that SSGD can be scaled to larger system sizes without vanishing gradients, addressing a key bottleneck for variational quantum algorithms.","Since the algorithm only needs gradient and Hessian measurements that are efficient, it can be implemented with modest circuit depth, making it a candidate for near-term quantum experiments.","The connection between local-minimum states and long-lived metastable states opens a path toward a rigorous lifetime bound, as the authors note."],"supporting_citations":[{"why":"Shows that Haar-random states have exponentially vanishing energy gradients, motivating the need for non-unitary operations via ancilla, and supplies Lemma C.1 used to justify the ancilla's role.","marker":"[19]"},{"why":"Supplies Theorem 1 giving the lower bound on energy variance that SSGD's brickwall circuit invokes to prove absence of barren plateaus.","marker":"[33]"},{"why":"Provides the mathematical theory of metastable states and the connection between energy barriers and long lifetimes that SSGD's local-minimum states are meant to instantiate.","marker":"[23]"},{"why":"The algorithm is inspired by ADAPT-VQE, which is the state-space approach that SSGD extends with ancilla and full gradient/Hessian updates.","marker":"[10]"},{"why":"The numerical simulations are performed with the QuTiP package, producing the clustering of final states at ground and metastable energies.","marker":"[34]"}],"fun_headline_variants":["Quantum descent finds ground or metastable states","No barren plateaus; yields ground or metastable states","When quantum descent misses, it still lands on metastable states","New method lands on ground or metastable states","Energy descent yields physical states, not just ground"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The no-barren-plateau guarantee assumes that the two-qubit gates in the SSGD circuit are effectively sampled from the Haar measure on SU(2), but the actual gates are rotations generated by a fixed local Pauli operator, so the cited variance bound may not apply.","fun_headline_variants_meta":{"raw":{"variants":["Quantum descent finds ground or metastable states","No barren plateaus; yields ground or metastable states","When quantum descent misses, it still lands on metastable states","New method lands on ground or metastable states","Energy descent yields physical states, not just ground"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001222,"raw_usage":{"total_tokens":5007,"prompt_tokens":908,"completion_tokens":4099,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":4024}},"tokens_in":524,"tokens_out":4099,"duration_ms":29569,"temperature":1.0,"reasoning_tokens":4024,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:26:25.162149+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the energy variance of the SSGD circuit for increasing system size N using the actual gates $e^{{-iθP}}$ with θ drawn from the same distribution used in training; if the variance decreases exponentially with N, the claimed barren-plateau avoidance fails. Alternatively, run SSGD on a 1D TFIM with a larger system and check whether the final-state energies continue to cluster at ground/metastable energies rather than at generic local minima.","supporting_citations":[{"cited_title":"Li, Quantum computers quickly find local minima, Nature Physics (2025)","cited_arxiv_id":null,"evidence_quote":"Supplies Theorem 1 giving the lower bound on energy variance that SSGD's brickwall circuit invokes to prove absence of barren plateaus."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The algorithm is inspired by ADAPT-VQE, which is the state-space approach that SSGD extends with ancilla and full gradient/Hessian updates."}],"review_version":1}