{"id":"d4028ee6-c983-4d5d-ae8c-791164148b38","arxiv_id":"2608.05682","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A damped alternating direction method of multipliers with subspace splitting provides the first convergence guarantee for constrained DFT calculations.","lead":"This paper introduces DASSP, an optimization algorithm for constrained density functional theory (CDFT) that separates the Stiefel manifold constraint from nonconvex quadratic constraints using a subspace variable. DASSP comes with a convergence proof, and numerical tests show it is more robust and often faster than existing CDFT solvers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central guarantee rests on Assumption 3.1(A2), a uniform eigengap for A(k); the paper neither proves it from problem data nor verifies it on its charge-transfer benchmarks, so the claim of a rigorous CDFT convergence guarantee is conditional.","rationale":"The reader's weakest_assumption is exactly Assumption 3.1(A2), and my independent reading of the proof confirms it is load-bearing: Lemma 3.1's decrease from the X-update relies on it, and without it the descent argument for the surrogate sequence collapses. The paper's Section 4.4 support is a single numerical case and not the paper's headline charge-transfer benchmarks, so the assumption remains unverified for the main experiments. I also note the secondary mismatch that the recommended delta = 0 is outside the proven regime, but the decisive technical point is the eigengap assumption. I did not find an internal inconsistency in Lemmas 3.1-3.7; the algebra of the surrogate sequence checks out. The paper is honest about the assumption and provides partial numerical evidence, which is credit. The central claim should be read as convergence for damped DASSP under an unverified uniform eigengap condition, which is a substantial but conditional contribution. Since the reader's CONDITIONAL verdict already captures this, no adjustment is needed.","tokens_in":31483,"tokens_out":9206,"duration_ms":89065,"concrete_test":"Rerun the HCN, C2H4, and CH3COCH3 charge-transfer tests (Tables 1-2) with Algorithm 1 instrumented to record gamma_k := lambda_p(A(k)) - lambda_{p+1}(A(k)) at every iteration, using a value of delta that satisfies Assumption 3.1 (e.g., delta = 10^-8, with tau chosen to satisfy (A5)) and also the default delta = 0. If any run with a proven delta has min_{k <= K} gamma_k = 0 or no positive lower bound before termination, Assumption 3.1(A2) fails on the paper's own benchmarks and Lemma 3.1's descent mechanism cannot explain the observed convergence. An even sharper analytic test: for p = 1, n = 2, m = 1 with W chosen so the dual update forces a crossing of the two eigenvalues of A(k), compute gamma_k explicitly and check whether the Lemma 3.1 inequality holds at the crossing; a violation would show the assumption is not merely technical.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The convergence theorem is no stronger than its descent mechanism. Lemma 3.1 is the only step that produces a decrease from the X-update, and it invokes [45, Lemma 3.1] with the uniform eigengap gamma from Assumption 3.1(A2). If lambda_p(A(k)) - lambda_{p+1}(A(k)) touches zero at some k, the inequality degenerates and the telescoping descent of the surrogate Psi_k in Lemma 3.5 has no control over the X-block. Assumption 3.1(A2) is not a consequence of the blanket Assumption 1.1 or of the problem structure; Remark 3.1 only argues local plausibility near a fixed point with large beta, and Section 4.4 documents a positive eigengap plateau for one molecule (water dimer cation, with delta = 10^-10). The charge-transfer systems in Tables 1 and 2, where Lambda(k) varies substantially before convergence, are not monitored for gamma_k. The paper itself shows how severe near-degeneracies can be in CDFT (effective Hamiltonian eigengap about 1.1e-12 in Section 4.4), so the analogous degeneracy for A(k) is a live risk. In addition, the practical default delta = 0 is outside the delta in (0,1) regime of Assumption 3.1, so the theorem covers a damped variant rather than the algorithm used in the headline experiments. Together these gaps mean 'rigorous convergence guarantee for DASSP on CDFT' is not fully established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reformulates the discretized constrained density functional theory problem (1.2) by introducing the projector variable Y = XX^T, which turns the nonconvex quadratic constraints into linear constraints in Y. A single-loop damped ADMM (DASSP) is proposed: an X-step that computes a dominant p-dimensional eigenspace of A(k) = Y(k) + (1 - δ)/β Λ(k), a Y-step that performs one projected gradient step on the affine set Y, and a damped dual update for Λ. The main convergence theorem (Theorem 3.1) states that under Assumption 3.1, including a uniform positive eigengap of A(k) and δ ∈ (0,1), every accumulation point is a δ-approximate KKT point with explicit stationarity and feasibility error bounds. Numerical experiments compare DASSP with quadratic-penalty and double-loop methods on charge-transfer systems and on a water-dimer cation charge-localization test, reporting high feasibility accuracy and speedups.","tokens_in":31745,"tokens_out":17216,"duration_ms":162284,"significance":"If the theorem is accepted, the paper delivers a genuinely single-loop algorithm with a nontrivial convergence guarantee for a nonconvex problem that has lacked such guarantees; the surrogate-sequence proof is explicit, the descent constants are derived from the stated parameter assumptions, and the numerical study is broader than typical for an optimization methods paper. The paper is also honest about the scope of the theory, explicitly flagging the δ = 0 limitation. The main gap is that the central guarantee is conditional on an eigengap assumption that is not verified for the main benchmarks, and the recommended experimental setting (δ = 0, adaptive stepsize, τ = 0.2) is outside the theorem's parameter regime; therefore the advertised 'first rigorous convergence guarantee' is not yet demonstrated for the algorithm as actually used.","major_comments":[{"comment":"The convergence theorem is proved only for δ ∈ (0,1): Assumption 3.1(A5) contains δ in the numerator and requires c > 2(1−δ)(2−δ)/δ, so the case δ = 0 is excluded. Section 4.1 nevertheless adopts δ = 0 as the default for all subsequent experiments, with the sentence 'even though its convergence is not covered in this work.' The abstract and contribution statements should therefore be scoped to the damped variant; as written, the claim that DASSP is the first algorithm with rigorous convergence guarantees for CDFT overstates what is proved. Please either extend the analysis to δ = 0 or restrict the claim and describe the undamped/default version as a heuristic whose convergence is only demonstrated numerically.","section":"§4.1 and Theorem 3.1"},{"comment":"The uniform eigengap γ_k ≥ γ > 0 for A(k) is load-bearing: Lemma 3.1 uses it to turn the X-update into a descent step, and without it the telescoping argument for the surrogate sequence Ψ_k has no control over the X-block. The paper does not derive this condition from problem data; Remark 3.1 only argues local plausibility near a fixed point with large β. The only direct numerical evidence is Section 4.4 for the water-dimer cation with δ = 10^-10, and the charge-transfer benchmarks in Tables 1 and 2 do not report γ_k. Given that the paper itself finds effective-Hamiltonian eigengaps as small as ~10^-12 in a CDFT problem, a vanishing gap for A(k) is a live risk in the tested applications. Please either prove a sufficient condition for A2 from the problem structure, or monitor γ_k for every reported benchmark and state the convergence claim as conditional on this verified condition.","section":"Assumption 3.1(A2), Lemma 3.1, §4.4"},{"comment":"The certified parameter regime is much narrower than the experimental one. Assumption 3.1(A4) requires a constant stepsize η ∈ (0, 2/(β+L)), while the default η_k = ηABB is a clipped adaptive Barzilai-Borwein stepsize. Moreover, for the value δ = 10^-10 used in the δ-sweep and in Section 4.4, Assumption 3.1(A5) forces τ to be of order δ/(2c) ≈ 10^-21 (with c ≈ 2(1−δ)(2−δ)/δ), whereas the experiments use τ = 0.2. Thus even the damped experiments do not satisfy the theorem's parameter conditions. The remark in Assumption 3.1(A4) that variable stepsizes can be handled by a 'straightforward but tedious' extension is not a proof. Please either prove convergence for the adaptive, under-relaxed parameter choices actually used, or report experiments in the certified regime (constant η, sufficiently small τ) and confirm that the qualitative conclusions are unchanged.","section":"Assumption 3.1(A4)-(A5) vs §4.1 parameters"}],"minor_comments":[{"comment":"The invertibility of the Gram matrix M is attributed to Assumption 1.1(A2); since A2 is stated for the matrices P⊥_X W_j X at local minimizers, not for the W_j themselves, a one-sentence justification outside the proof would help the reader.","section":"Eq. (2.7)-(2.8)"},{"comment":"The text says the initial eigengap of A(0) may vanish and the figure reports a minimal value of 1.110e-16; please clarify whether this value occurs at k = 0 and whether the plateau value is reached before the iteration count reported, so the reader can judge how representative the eigengap history is.","section":"§4.4, Figure 8"},{"comment":"Several axis labels and legends in Figures 6 and 7 contain garbled text ('le el', 'Oute SCF ite ation', 'Subspace di tance') that should be corrected in a revised version.","section":"Figures 6-7"},{"comment":"The set Y defined in (2.2) is an affine subspace, not a manifold; calling TY its 'tangent space' is nonstandard and could be replaced by 'parallel subspace' or a similar term.","section":"§2, Y-update"}],"recommendation":"major_revision","confidential_remarks":"The theoretical core appears internally consistent; I found no algebraic error in Lemmas 3.3-3.6. The main risk is external validity: Theorem 3.1 covers a damped, constant-stepsize, tiny-τ regime, while the paper's recommended and tested configuration is undamped with adaptive stepsize. This is fixable by either adding the missing analysis or repositioning the claims, and by verifying Assumption 3.1(A2) on the benchmarks that support the 'first convergence guarantee' claim. The use of [45] as an external lemma is appropriate and does not raise a self-citation concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a strong, honest optimization paper, but the headline guarantee is more conditional than the abstract suggests. The subspace-splitting reformulation is genuinely new for CDFT, and the damped ADMM analysis with a carefully constructed surrogate sequence is a real step for nonconvex, nonlinearly constrained problems. The numerical work is broad and supports the practical claims: DASSP hits feasibility around 1e-9 and beats the double-loop WV method in wall-clock time, especially when the effective Hamiltonian has a tiny eigengap. The water dimer case is convincing.\n\nSoft spots, in order of importance. First, Assumption 3.1(A2) — uniform positive eigengap for A(k) — is load-bearing: Lemma 3.1, the only descent control on the X-block, uses it directly. The paper gives one numerical check on water dimer and a local-plausibility remark, but does not verify the eigengap on the charge-transfer benchmarks where Λ(k) moves a lot. If A(k) degenerates at some iteration, the proof stops. The stress-test is right about this, and the paper's own Section 4.4 shows how close CDFT systems can get to degeneracy. Second, the recommended default δ=0 is outside the theorem's δ>0 regime. The authors are explicit about this, and the δ=10^-10 curves are practically identical to δ=0, so it is a minor-to-moderate gap rather than a fatal one, but it does mean the benchmarked algorithm is not the one with the rigorous guarantee. Third, the guarantee is δ-approximate KKT; that is a reasonable price for damping, not a flaw.\n\nOn the citation pattern: Lemma 3.1 borrows from [45], which one of the authors co-authored, but that result is published and checkable; no circularity burden. No code is released, which slows reproducibility but is not a scientific defect.\n\nOverall: a solid paper, honestly written, with a clear new contribution. The main claim should be softened in the abstract — \"first damped ADMM variant for CDFT with rigorous guarantees under an eigengap assumption\" rather than the unqualified version. I would send it to a serious referee on optimization and electronic structure; the referee should push for a sharper treatment of Assumption 3.1(A2) or a verification on the actual test problems, and ideally an analysis covering δ=0 or a change to the default parameter.","headline":"A serious, honest optimization paper that gives the first convergence guarantee for a damped subspace-splitting ADMM for CDFT, under an unverified eigengap assumption and for a parameter setting the authors do not use in their own benchmarks.","tokens_in":32357,"tokens_out":3514,"would_cite":true,"duration_ms":34468,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C30","65K05","65F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Constrained DFT can be solved by a single-loop damped ADMM whose every accumulation point is a $\\delta$-approximate KKT point, giving the field its first rigorous convergence guarantee.","keywords":["constrained density functional theory","CDFT","alternating direction method of multipliers","subspace splitting","Stiefel manifold","nonconvex quadratic constraints","damped dual ascent","approximate KKT points"],"falsifier":"Engineer a CDFT instance where the $p$-th and $(p+1)$-th eigenvalues of $A^{(k)}$ become equal at some iteration (for example, two identical, widely separated fragments with a degeneracy in the density-matrix subspace) and check whether the iterates still converge to a $\\delta$-approximate KKT point; if they do not, Theorem 3.1 is false as stated. A complementary check is to measure the residuals at the accumulation point for $\\delta=10^{-2}$, $10^{-4}$, and $10^{-6}$ and verify they decrease proportionally to $\\delta$, as the theorem's bounds predict.","tokens_in":31176,"feed_emoji":"⚛️","tokens_out":9475,"duration_ms":89240,"temperature":0.7,"pith_summary":"Constrained density functional theory (CDFT) is the standard way to describe charge-localized and excited electronic states, but its discretized optimization problem couples an orthogonality constraint on orbitals with nonconvex quadratic density constraints, and existing solvers either violate the constraints or rely on fragile nested iterations. The paper shows that the two groups of constraints can be decoupled exactly: by exploiting rotation invariance, the density matrix $XX^\\top$ is treated as a separate variable $Y$ constrained only by linear trace conditions, linked back to $X$ by $Y=XX^\\top$. On this reformulation the paper builds DASSP, a single-loop damped alternating direction method of multipliers, and proves that under natural smoothness and eigengap assumptions every accumulation point is a $\\delta$-approximate KKT point whose stationarity and feasibility errors are bounded by constants times $\\delta$. A sympathetic reader would care because this is the first CDFT solver with a rigorous convergence guarantee, and the accuracy versus speed trade-off is controlled by one damping parameter.","feed_headline":"One-loop solver gives constrained DFT its first convergence proof","feed_subtitle":"DASSP splits the density matrix from the orbital constraints, turning limit points into δ-accurate KKT points.","key_machinery":"The load-bearing object is the subspace-splitting reformulation $Y=XX^\\top$ combined with the damped augmented Lagrangian $\\mathcal{L}_{\\beta,\\delta}(X,Y,\\Lambda)=f(Y)+\\langle(1-\\delta)\\Lambda,\\,Y-XX^\\top\\rangle+(\\beta/2)\\|Y-XX^\\top\\|_F^2$. The affine set $\\mathcal{Y}=\\{Y:\\,\\mathrm{Tr}(W_jY)=b_j\\}$ makes the nonconvex quadratic constraints linear, and the $X$-subproblem becomes a spectral step: compute the dominant $p$-dimensional invariant subspace of $A^{(k)}=Y^{(k)}+(1-\\delta)/\\beta\\,\\Lambda^{(k)}$. The $Y$-subproblem is a single projected gradient step onto $\\mathcal{Y}$, with a closed-form projection requiring only the solution of the $m\\times m$ normal system $M\\hat\\mu=-c$, and the dual variable is updated by the damped ascent $\\Lambda^{(k+1)}=(1-\\delta)\\Lambda^{(k)}+\\tau\\beta(Y^{(k+1)}-X^{(k+1)}X^{(k+1)\\top})$. The convergence proof is carried by the surrogate sequence $\\Psi_k$, which adds a scaled dual-difference penalty to the damped augmented Lagrangian; Lemmas 3.5 and 3.6 show $\\Psi_k$ descends each iteration and is bounded below, yielding the vanishing increments and the $\\delta$-proportional stationarity and feasibility bounds.","core_discovery":"The central claim is that DASSP, a damped ADMM for the subspace-splitting reformulation of CDFT, converges to approximately KKT points with an error that the user can preset. Theorem 3.1 states that if Assumption 3.1 holds, every accumulation point $(X^\\star,Y^\\star,\\Lambda^\\star)$ admits multipliers $\\Sigma^\\star$, $\\mu^\\star$ such that the stationarity residual $\\|F(X^\\star,\\mu^\\star)X^\\star-X^\\star\\Sigma^\\star\\|_F$ is at most $L\\sqrt{p(n-p)}/(\\tau\\beta)\\|\\Lambda^\\star\\|_F\\,\\delta$ and each quadratic constraint violation $|\\mathrm{Tr}(X^{\\star\\top}W_jX^\\star)-b_j|$ is at most $(1/(\\tau\\beta))\\|\\Lambda^\\star\\|_F\\|W_j\\|_F\\,\\delta$. The convergence mechanism is a surrogate-sequence argument: a constructed sequence $\\Psi_k$ decreases at every iteration and is bounded below, which forces the primal and dual increments to vanish and controls the dual iterates, the usual obstacle for nonconvex ADMM. The feasibility error comes from the limiting alignment residual $R^\\star=Y^\\star-X^\\star X^{\\star\\top}$; because $Y^\\star$ is exactly feasible, the residual is absorbed by the dual update as $\\delta/(\\tau\\beta)\\|\\Lambda^\\star\\|_F$. The paper also reports numerical evidence that DASSP reaches feasibility violations near $10^{-9}$ while the double-loop baseline struggles on systems with tiny eigengaps.","pith_inferences":["A continuation schedule on $\\delta$ (large damping for stability, then shrinking $\\delta$ to refine accuracy) is a natural practical extension that the paper does not test; the $\\delta$-proportional error bounds make the accuracy at each stage predictable.","Because DASSP does not rely on the effective Hamiltonian's eigengap but on the eigengap of $A^{(k)}$, it may be stable in charge-localization regimes where standard SCF-based CDFT oscillates; the water dimer cation experiment supports this, and a broader comparison on magnetic-moment constraints would test it.","The same subspace-splitting reformulation should extend to noncollinear CDFT with local magnetic-moment constraints, since those constraints are also quadratic in the orbitals; the paper lists this as future work, and the optimization proof is constraint-agnostic apart from the linear independence assumption.","The $\\delta$-approximate KKT certificate could be turned into a practical stopping rule: stop when increments are small and report $\\|\\Lambda\\|_F/(\\tau\\beta)\\delta$ as a certified bound, though the paper does not propose such a protocol."],"forward_implications":["CDFT calculations can be run as a single loop with no inner Newton iterations, so per-iteration cost is one eigensolve, one projected gradient step, and one closed-form dual update, all of standard numerical linear algebra.","The final accuracy is set in advance: choosing a smaller damping $\\delta$ produces limit points that are $\\delta$-proportionally closer to being stationary and feasible, at the cost of slower dual movement.","Penalty-based methods no longer need to be pushed to very large penalty parameters to force feasibility, so the conditioning problems documented for quadratic penalty SCF can be avoided.","The convergence theorem applies to any smooth objective on the Stiefel manifold with finitely many quadratic constraints satisfying the assumptions, not just to the Kohn-Sham energy; the paper's claims are formulated at that general level.","The undamped limit $\\delta=0$, which the numerical experiments find fast and stable, is not covered by the theorem and is explicitly left open by the authors."],"supporting_citations":[{"why":"Supplies the eigenvalue-subspace perturbation inequality used in Lemma 3.1 to show the X-update strictly decreases the damped augmented Lagrangian.","marker":"[45]"},{"why":"Defines the double-loop constrained-DFT method that DASSP competes against and whose eigengap instability motivates the single-loop design.","marker":"[61]"},{"why":"Provides the water-dimer-cation charge-localization setup used to test DASSP's robustness against perturbed initializations.","marker":"[2]"},{"why":"The electronic-structure program in which DASSP and all comparison algorithms are implemented, so it carries the numerical experiments.","marker":"[55]"},{"why":"Cited for the eigengap condition and level-shift strategy used to frame Assumption 3.1(A2) and to interpret the robustness tests.","marker":"[4]"},{"why":"Supplies the clipped alternating Barzilai-Borwein stepsize selected as the default Y-update stepsize in the parameter tuning.","marker":"[13]"}],"fun_headline_variants":["First proven-convergent solver for constrained DFT","DASSP: single-loop CDFT with rigorous guarantees","Damped ADMM gives CDFT its first convergence proof","Constrained DFT now has a convergent algorithm","One-loop CDFT solver with preset error bounds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"At every step, the matrix whose top $p$ eigenvectors define the new orbitals must keep its $p$-th and $(p+1)$-th eigenvalues separated by a fixed positive gap; if the gap closes at any iteration, the proof's central descent estimate fails.","fun_headline_variants_meta":{"raw":{"variants":["First proven-convergent solver for constrained DFT","DASSP: single-loop CDFT with rigorous guarantees","Damped ADMM gives CDFT its first convergence proof","Constrained DFT now has a convergent algorithm","One-loop CDFT solver with preset error bounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1351,"prompt_tokens":1112,"completion_tokens":239,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":728,"completion_tokens_details":{"reasoning_tokens":161}},"tokens_in":728,"tokens_out":239,"duration_ms":3136,"temperature":1.0,"reasoning_tokens":161,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:58:12.959856+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Engineer a CDFT instance where the $p$-th and $(p+1)$-th eigenvalues of $A^{(k)}$ become equal at some iteration (for example, two identical, widely separated fragments with a degeneracy in the density-matrix subspace) and check whether the iterates still converge to a $\\delta$-approximate KKT point; if they do not, Theorem 3.1 is false as stated. A complementary check is to measure the residuals at the accumulation point for $\\delta=10^{-2}$, $10^{-4}$, and $10^{-6}$ and verify they decrease proportionally to $\\delta$, as the theorem's bounds predict.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the eigenvalue-subspace perturbation inequality used in Lemma 3.1 to show the X-update strictly decreases the damped augmented Lagrangian."},{"cited_title":"Wu and T","cited_arxiv_id":null,"evidence_quote":"Defines the double-loop constrained-DFT method that DASSP competes against and whose eigengap instability motivates the single-loop design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the water-dimer-cation charge-localization setup used to test DASSP's robustness against perturbed initializations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The electronic-structure program in which DASSP and all comparison algorithms are implemented, so it carries the numerical experiments."},{"cited_title":"Bai, R.-C","cited_arxiv_id":null,"evidence_quote":"Cited for the eigengap condition and level-shift strategy used to frame Assumption 3.1(A2) and to interpret the robustness tests."},{"cited_title":"Dai and R","cited_arxiv_id":null,"evidence_quote":"Supplies the clipped alternating Barzilai-Borwein stepsize selected as the default Y-update stepsize in the parameter tuning."}],"review_version":1}