{"id":"00041960-6de3-4cc7-8f4a-041f13db20a6","arxiv_id":"2509.02658","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A single shared trunk plus one linear head per state can represent a degenerate eigenspace exactly if the trunk width is at least the combined linear rank of target log-moduli and phases minus one on the common support.","lead":"A neural-network quantum state method is proposed in which many degenerate quantum states share one feature-extracting network, with a small linear head for each state. The paper proves an exact-representability condition in terms of trunk width and demonstrates the approach on small frustrated spin chains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem A.1 proves representability for arbitrary feature maps; the transfer to width-h MLP trunks rests on the unproven assertion in Remark A.3, so the minimal-width practical claim is not established.","rationale":"The paper's core linear-algebra theorem is sound as a statement about unrestricted feature maps: the affine-rank bounds in Lemma A.1 and Corollary A.1 correctly limit what a fixed trunk can realize, and the converse direction is valid. The numerical proof-of-principle for small chains is consistent with the claimed mechanism, and the explicit Majumdar-Ghosh ground states provide a clean test bed. The load-bearing gap is the transfer from this abstract theorem to an actual neural-network trunk. The sufficiency proof chooses trunk features as basis vectors of Rboth; showing that a standard MLP of the same width can realize those basis vectors is exactly the unproven assertion in Remark A.3. For fixed width h, MLPs are not universal interpolators on a finite set when h is small relative to |S|; the MG N = 8 common support has 28 points, and a width-2 MLP with two hidden layers lacks the parameters and linear regions to realize the indicator functions forming Rboth. Thus the minimal-width statement h* = rboth - 1 is established only for abstract feature maps, not for the NQS used in practice. This does not invalidate the conditional theorem ('if the feature map's column space contains Rboth') nor the numerical demonstrations with wider trunks; it means the paper should either prove the MLP realizability for the relevant architectures and sizes or soften the minimal-width claim. The common-support restriction is a lesser issue: for VMC expectation values, configurations where a target state vanishes are never sampled, so the inability of the exponential ansatz to be exactly zero is not fatal, though it weakens the literal reading of 'exact'. Because the reader's conditional verdict already captures the main gap, I recommend no change to the verdict.","tokens_in":25461,"tokens_out":14465,"duration_ms":147627,"concrete_test":"Run a supervised exact-fit check for the N = 8 MG chain: with the same two-hidden-layer ReLU MLP trunk used in the paper and width h = 2 (the theorem's h* = rboth - 1), solve directly for trunk parameters to reproduce the three basis vectors 1_S, 1_{suppPhiA\\suppPhiB}, and 1_{suppPhiB\\suppPhiA} on the 28-point common support S. If the least-squares residual cannot be driven to numerical zero, then Remark A.3's existence assertion is false for this standard architecture and the minimal-width practical claim is unsupported; if the residual is zero, the transfer holds for this case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central representability theorem (Theorem A.1, Appendix A) is proven for an unrestricted feature map f_theta: S -> R^h. The sufficiency direction constructs the trunk by setting its components to a basis of Rboth, which is a valid existence argument for abstract maps but not for a neural network. The paper's practical claim that a real NQS trunk of width h = rboth - 1 suffices depends entirely on Remark A.3, which asserts without proof that this construction is implementable exactly on S by standard MLP trunks at the same width. That assertion is not a consequence of the theorem and is in general false for fixed-width MLPs: a two-hidden-layer MLP with h hidden units per layer is a restricted function class with O(hN + h^2) parameters and a bounded number of linear regions, while exact interpolation on a large finite support requires matching many independent values. For the MG model with N = 8, the common support has 28 configurations and rboth = 3, so the theorem predicts h = 2; a width-2 two-hidden-layer ReLU MLP cannot in general realize the indicator functions that form the basis of Rboth. Thus the minimal-width statement h* = rboth - 1 is established only for abstract feature maps, not for the NQS ansatz used in practice. Additionally, exactness is only asserted on the common support S, where all targets are nonzero; the exponential ansatz cannot vanish off S, so the 'exactly' in the conclusion is weaker than it appears, although this is likely not fatal for VMC expectation values since zero-amplitude configurations are not sampled under the target state.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a single-trunk multi-head (ST-MH) neural network quantum state (NQS) ensemble for simultaneously approximating K degenerate eigenstates. A shared feature-extracting trunk feeds lightweight complex linear heads, one per target state. The authors derive analytic gradients for the weighted energy plus orthogonality-penalty cost in variational Monte Carlo (VMC), prove a representability theorem (Theorem A.1, Appendix A) that gives a rank condition h+1 >= rboth on the trunk width for an abstract feature map, provide a qualitative cost analysis predicting a roughly K-fold saving in the trunk-dominated regime, and validate the method on the spin-1/2 J1-J2 Heisenberg model at the Majumdar-Ghosh point for N=4,6,8. The numerical experiments report high fidelities, full rank of the ground-space projection matrix, and reduced runtime relative to a multi-trunk multi-head (MT-MH) ensemble.","tokens_in":25805,"tokens_out":12211,"duration_ms":114312,"significance":"If the central claims hold, the ST-MH ansatz is a useful architectural compression for degenerate eigenspace learning, with potential savings in parameters and runtime, and the paper provides a concrete proof-of-principle on a frustrated model with exact ground states. The work has genuine strengths: Theorem A.1 is a clean linear-algebra statement whose sufficiency direction is constructive; the analytic gradients in Appendix B follow standard VMC identities; the numerical validation is internally consistent, including exact post-training overlap checks and comparisons to exact diagonalization and exact Majumdar-Ghosh ground states; and the cost model is explicit rather than fitted. However, the transfer of the representability theorem from abstract feature maps to fixed-width MLP trunks is asserted rather than proven, and the theorem's necessity direction has a phase-branch modulo-2pi gap. These issues affect the paper's strongest claims and need to be addressed.","major_comments":[{"comment":"The proof of the converse direction assumes that the chosen target phase branches Omega_(j) are exactly realized as affine functions of the trunk features. But the condition psi_k = Psi^(k) on S only gives pointwise equality of phases modulo 2pi; the realized phase phi_k f + gamma_k can differ from the chosen branch by a configuration-dependent integer multiple of 2pi. Consequently, the chosen branch evaluation vectors need not lie in col(X), and the inequality rboth > h+1 does not, by itself, rule out the existence of a representation. A concrete counterexample: take S = {x1, x2}, D = 1, Psi(x) = 1 identically, and choose the phase branch Omega = (0, 2pi). Then rboth = 2, yet a width-0 trunk with beta = 0 represents the state exactly, contradicting the claimed equivalence. The theorem should define rboth as the minimum attainable rank over admissible branch choices, or explicitly restrict the statement to branches that are exactly realized by the linear heads.","section":"Appendix A, Theorem A.1 (necessity direction)"},{"comment":"The assertion that the constructive proof of Theorem A.1 is implementable exactly on S by standard MLP trunks at width h is unproven and, at the minimal width h = rboth - 1, generally false for fixed-width MLPs. For the N = 8 MG example, rboth = 3 so the theorem predicts h = 2, while the common support S has 28 configurations; a two-hidden-layer ReLU MLP with 2 units per layer is a restricted function class and cannot in general realize the basis vectors of Rboth (which include indicator-like functions) exactly on 28 points. The paper should either provide a rigorous expressivity statement for the specific architecture, or present the MLP implementability as a separate hypothesis/empirical observation and soften the 'minimal width' claims in the abstract, Section 2.2.3, and conclusion.","section":"Appendix A, Remark A.3"},{"comment":"The claim that ST-MH 'can represent every degenerate eigenstate exactly' is stronger than Theorem A.1, which establishes equality only on the common support S (Appendix A). Off S, the exponential ansatz cannot vanish, so states with nodal configurations are not represented exactly. The paper should qualify all exactness statements as 'exact on the common support' and note the implication for nodal configurations (such as the Neel states in the MG model). The fidelities below 1 in Table 1 are consistent with this caveat and could be discussed in that light.","section":"Abstract and Conclusion"}],"minor_comments":[{"comment":"The sentence 'Lastly we provide a qualitative computational cost analysis which incentivise the applicability...' contains a grammatical error; 'incentivise' should be 'incentivizes' or 'motivates', and the intended meaning could be made clearer.","section":"Abstract"},{"comment":"Both equations contain a garbled LaTeX artifact ('/Leftr⫯g⊸tl⫯ne⇒') that makes the display unreadable and should be corrected.","section":"Equation (8) and Equation (C.11)"},{"comment":"The phrase 'the compute time required for sampling amortises that of the gradient computations' is unclear; likely the intended statement is that sampling time dominates or swamps the gradient computation time for small networks.","section":"Section 2.2.2"},{"comment":"The notation h*_both for the minimal representation width and h*_(s) for the cost threshold is easily confused; consider renaming one of the two quantities.","section":"Section 2.3.2"},{"comment":"The sentence 'One either support, the amplitudes have flat modulus...' contains a typo; it should read 'On either support'.","section":"Section 2.2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solo-authored proof-of-principle with a helpful numerical validation. The main theoretical contribution, Theorem A.1, is a clean linear-algebra statement, but its necessity direction is not correct as stated due to the modulo-2pi phase issue, and the transfer to MLP trunks is an unproven expressivity assumption. Both issues are fixable within the manuscript's scope by adjusting the statements and explicitly labeling the MLP implementability as a separate assumption. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a real look. The shared-trunk multi-head construction for degenerate NQS eigenspaces is a sensible transfer from multi-task learning, and the representability theorem (Appendix A) gives a clean characterization: on the common support, exact representation is possible iff h+1 >= rboth. That is a simple linear-algebra statement, proven correctly in both directions. The analytic gradients in Appendix B follow standard VMC identities, and the numerical validation on the Majumdar-Ghosh chain is convincing for the sizes shown: post-training exact overlaps, fidelities above 0.997, and full-rank projection matrices for N=4,6,8. The cost analysis is honest, with the K-fold saving explicitly confined to the trunk-dominated regime.\n\nThe main soft spot is the gap between theorem and practical ansatz. Theorem A.1 proves existence for arbitrary feature maps by setting trunk features to a basis of Rboth; it does not show that a fixed-width MLP trunk can realize that basis on a finite support. Remark A.3 asserts exactly this without proof. That matters because the paper's practical claim is that a real neural network of minimal width h*=rboth-1 suffices. For N=8, the support has 28 configurations and h=2; a two-hidden-layer ReLU MLP with two units per layer is a very restricted class, and it is not obvious it can represent the required basis. The h=2 ablation at N=4 gives empirical support, but only as a single data point. This is fixable: either prove a finite-support expressivity statement for standard MLPs, or present the minimal-width result as an existence claim for abstract feature maps and treat the MLP realization as a reasonable assumption.\n\nMinor issues: the abstract and conclusion say 'exactly' without always carrying the common-support qualifier, though Appendix A acknowledges that zero-amplitude configurations are not representable by the exponential ansatz. And no code or data is provided, making direct replication harder; Appendix D gives full hyperparameters but not the resulting states.\n\nOverall the central argument holds up; the MLP transfer gap is real but addressable and does not sink the approach. It deserves a serious referee, and I would expect minor-to-moderate revision.","headline":"Clean representability theorem and sensible architecture, but the practical minimal-width MLP claim rests on an unproven expressivity assumption.","tokens_in":26314,"tokens_out":4008,"would_cite":true,"duration_ms":36529,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single neural network with multiple linear heads can represent every state of a degenerate ground-state manifold exactly whenever its width meets a rank bound, and it does so at a fraction of the cost of K independent networks.","keywords":["neural network quantum states","variational Monte Carlo","degenerate ground states","single-trunk multi-head ansatz","representability theorem","shared trunk","orthogonality penalty","J1-J2 chain"],"falsifier":"Take a small configuration space where exact enumeration is possible, choose two target states whose log-moduli and phases have rboth = 4, and train the ST-MH ensemble with a trunk of width h = 2 plus the orthogonality penalty; the theorem predicts exact representation on the common support is impossible, so exact reproduction would falsify the necessity direction. Conversely, for the J1-J2 ring at J2=J1/2 with N=8 the theorem predicts h* = 2, so a trunk of width 1 that still reaches unit ground-subspace fidelity would falsify the minimal-width claim.","tokens_in":25234,"feed_emoji":"⚛️","tokens_out":10001,"duration_ms":87441,"temperature":0.7,"pith_summary":"The paper proposes a single-trunk multi-head (ST-MH) neural network quantum state that approximates all states of a degenerate ground-state manifold in one training run, instead of optimizing K separate networks. Its central claim is an exact representability condition: on the common support where every target state is non-vanishing, the ST-MH ansatz represents all D degenerate eigenstates exactly if and only if the shared trunk width h satisfies h+1 ≥ rboth, where rboth is the linear rank of the combined span of the states' log-moduli, their chosen phase branches, and the constant function. When that condition holds, the parameter count drops by roughly a factor equal to the degeneracy K relative to a multi-trunk ensemble with the same per-trunk width, and the leading variational Monte Carlo cost is reduced by a similar factor whenever the trunk dominates the computation. A proof-of-principle on the frustrated spin-1/2 J1-J2 chain at the solvable coupling J2=J1/2 resolves the two momentum eigenstates with high fidelity and near-zero overlap error, including with a trunk width of only 2. The practical importance is that degenerate eigenspaces, which are normally expensive and prone to redundant convergence, become accessible with one compact shared representation.","feed_headline":"One shared network can capture every degenerate ground state","feed_subtitle":"A single trunk plus K linear heads represents the whole manifold exactly when a rank condition holds, at a fraction of the cost.","key_machinery":"The load-bearing object is the single-trunk multi-head ansatz ψ_k(x)=exp[χ_k·f_θ(x)+c_k], where a single nonlinear feature map f_θ from configurations to R^h is shared by K lightweight complex-linear heads parametrized by χ_k and c_k. The paper's argument rests on an affine rank bound (Lemma A.1 and its phase analogue): for any fixed trunk, the realized log-moduli of all heads, and likewise their phase lifts, lie in an affine subspace of dimension at most h+1, spanned by the h feature columns plus the constant vector. Theorem A.1 converts this bound into an equivalence by requiring the combined linear span of the targets' log-moduli, phase branches, and constant to fit inside that same column space; the combined rank rboth then sets the minimal width. A separate cost model counts forward and backward FLOPs through a two-layer trunk, showing that trunk cost is shared across heads and becomes independent of K up to an O($K^{2}$) overlap-penalty term when 3F_T dominates 6Kh.","core_discovery":"The paper's central claim is Theorem A.1, stated for a finite configuration space and D target eigenstates. After fixing single-valued phase branches and letting S be the common support where all target states have non-zero amplitude, define rboth as the dimension of the linear span of the constant vector together with the target log-moduli and phase branches on S. Then a single-trunk multi-head ansatz of the form ψ_k(x) = exp(χ_k · f_θ(x) + c_k), with one shared feature map f_θ and K complex linear heads, can represent every target eigenstate exactly on S if and only if h+1 ≥ rboth. The minimal width is therefore h*_both = rboth − 1, and if rboth > h+1 no amount of extra heads can compensate, because all heads are limited to linear combinations of the same h+1 feature coordinates. The same construction, in trunk-dominated regimes and with equal per-trunk widths, cuts the parameter count and the leading gradient cost by a factor close to K relative to K independent trunks. Exactness is deliberately scoped to S: where a target state has a zero, the exponential ansatz cannot vanish exactly, so the paper's 'exactly' means pointwise on the common support. Numerically, the two degenerate momentum eigenstates of the J1-J2 ring at J2=J1/2 are resolved for N=4,6,8 with ground-subspace fidelities above 0.998 and overlap-matrix deviations below 0.01, and an ablation with trunk width 2 confirms the predicted minimal width.","pith_inferences":["The rank condition can be used as an expressivity diagnostic: for a given ansatz family, the smallest trunk width at which exact representation becomes achievable empirically estimates how well that family realizes arbitrary feature functions, independent of the Hamiltonian's details.","Because exactness holds only on the common support, configurations where a target state vanishes will carry small nonzero amplitudes in the exponential ansatz; observables concentrated on such 'nodal' configurations could accumulate errors that the K-fold speedup would need to be weighed against.","The shared-trunk construction is not tied to degeneracy: the same mechanism should extend to excited-state manifolds and to transfer-learning settings where one trunk serves multiple Hamiltonians, with only the rank rboth changing and the threshold moving accordingly.","For very large degeneracies K, the benefit shrinks twice: rboth can grow with K, forcing a wider trunk, and the O(K^2) overlap term eventually competes with the shared trunk cost, so the regime of clear ST-MH advantage is small-to-moderate K."],"forward_implications":["If rboth ≤ h+1, a single shared trunk represents the entire degenerate ground-state manifold exactly on the common support, so K independent networks are not needed for representability.","In trunk-dominated regimes with equal per-trunk widths, the ST-MH parameter count and leading variational Monte Carlo cost scale as roughly 1/K of the multi-trunk ensemble's, with only the shared O(K^2) pairwise-overlap term growing in K.","The minimal width h*_both = rboth−1 is a concrete, checkable number for small systems, and the N=4 ablation with h=2 supports the theorem's prediction for the J1-J2 ring at J2=J1/2.","If rboth > h+1, adding more heads cannot restore exact representability; the ensemble must widen the trunk or fall back to separate trunks.","On the examined J1-J2 rings (N=4,6,8), ST-MH resolves the two translation eigenstates with ground-subspace fidelities above 0.998, showing the resource saving does not sacrifice accuracy."],"supporting_citations":[{"why":"Defines the neural network quantum state in exponential/log-feature form that the ST-MH ansatz adapts by sharing a trunk.","marker":"[16]"},{"why":"Introduces the ensemble variational Monte Carlo cost with an overlap and orthogonality penalty that ST-MH adopts and that defines the MT-MH comparison baseline.","marker":"[38]"},{"why":"Supplies the log-derivative identities used to derive the closed-form energy and overlap gradients for the shared-trunk ensemble.","marker":"[47,48]"},{"why":"Provides the exact degenerate ground states of the J1-J2 chain at the solvable coupling used in the numerical proof-of-principle.","marker":"[44–46]"}],"fun_headline_variants":["Single trunk, K heads: one network for all degenerate states","Exact degenerate eigenstates from one shared feature trunk","One net, many states: ST-MH cuts cost by factor K","Shared trunk exactly captures degenerate manifolds","ST-MH: one trunk represents K degenerate states exactly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a real neural network trunk of width h can exactly produce, on the finite common support, any chosen feature functions that span the combined modulus-phase space; if standard MLP trunks cannot realize those features at that width, the minimal-width claim fails, and in any case 'exact' applies only where every target state is non-vanishing, since the exponential ansatz cannot vanish exactly.","fun_headline_variants_meta":{"raw":{"variants":["Single trunk, K heads: one network for all degenerate states","Exact degenerate eigenstates from one shared feature trunk","One net, many states: ST-MH cuts cost by factor K","Shared trunk exactly captures degenerate manifolds","ST-MH: one trunk represents K degenerate states exactly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2888,"prompt_tokens":1186,"completion_tokens":1702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":802,"completion_tokens_details":{"reasoning_tokens":1622}},"tokens_in":802,"tokens_out":1702,"duration_ms":12387,"temperature":1.0,"reasoning_tokens":1622,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:36:49.681529+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small configuration space where exact enumeration is possible, choose two target states whose log-moduli and phases have rboth = 4, and train the ST-MH ensemble with a trunk of width h = 2 plus the orthogonality penalty; the theorem predicts exact representation on the common support is impossible, so exact reproduction would falsify the necessity direction. Conversely, for the J1-J2 ring at J2=J1/2 with N=8 the theorem predicts h* = 2, so a trunk of width 1 that still reaches unit ground-subspace fidelity would falsify the minimal-width claim.","supporting_citations":[],"review_version":2}