{"id":"2bba1725-bdb8-45c7-a9c7-584e0fd0cd2b","arxiv_id":"2508.20280","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Nonlinear splitting of gradients and constraints yields semi-implicit gradient methods and cheaper adjoint gradients, demonstrated on nonconvex benchmarks, quantum tomography, and kinetic transport.","lead":"This paper introduces a way to split the gradient of an optimization problem into a function of two arguments, so each step can be partly implicit and more stable. It applies the same idea to PDE-constrained problems by splitting the constraint equation, which reduces the number of expensive simulation solves needed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 is not the modified adjoint implied by the paper's own splitting construction: the derivatives in (3.26) are evaluated at x_{k+1} while the splitting eF(x,y)=G(x)-y requires evaluation at x_k. This breaks the claimed systematic derivation behind the constrained-case results.","rationale":"The reader's weakest assumption, that the split gradient in Algorithm 1 is not proven to be a descent direction, is real and is explicitly conceded in Section 5. My stress-test surfaced a more specific and more fundamental problem: for the constrained case, Algorithm 1 as written is not a consequence of the splitting construction the paper derives. The governing equations use derivative evaluations at x_{k+1}, whereas the splitting eF(x,y)=G(x)-y requires them at x_k. This is not a matter of missing convergence analysis; it is an internal inconsistency in the central derivation. Since the headline factor-of-three efficiency gain comes from the constrained neutron transport example, the correctness of this derivation is load-bearing. I do not question the numerical results themselves: the unconstrained examples, the Levenberg-Marquardt equivalence, and the stability observations may well be reproducible, and the paper is otherwise clearly written and self-contained. But a framework paper whose central constrained-case construction is inconsistent in its equations needs correction before the efficiency claims can be accepted. The proper fix is to either change Algorithm 1 so that it follows from Eqs. (3.25)-(3.27), or explicitly present Algorithm 1 as a heuristic and separate the claimed systematic construction from the numerical demonstrations. The reader's CONDITIONAL verdict remains appropriate: acceptance should require fixing the derivation, clarifying the status of Algorithm 1, and rerunning the Section 4.2.1 comparison with the correctly derived split gradient if it differs. My concern is complementary to the reader's: even after the derivation is fixed, a descent or convergence guarantee for the corrected split gradient is still absent, and Section 5's deferred analysis is an unresolved limitation.","tokens_in":23297,"tokens_out":16940,"duration_ms":155707,"concrete_test":"Re-derive Algorithm 1 from (3.25)-(3.27) on a 1D analogue: F(x)=(1+alpha x^2)x-q, G(x)=q/(1+alpha x^2), J(x)=1/2(x-x*)^2. Compute the split gradient two ways: (a) Algorithm 1 with D1F(x+) and partial F(x+)/partial theta; (b) the general construction with D1G(x)-I and partial G(x)/partial theta evaluated at (x,G(x)). For alpha>0, q neq 0 and x+ neq x the two expressions differ, which can be checked by exact algebra or a three-line script. If they differ, the paper must either modify Algorithm 1 to match (3.27) or reclassify it as a heuristic, and the constrained-case efficiency comparisons in Section 4.2.1 need to be rerun against the correctly derived split gradient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.2 the paper reduces a nonlinear solver iteration to a splitting eF(x,y;theta)=G(x;theta)-y (Eq. 3.25) and gives a general modified-adjoint construction (Eqs. 3.27a-3.27c). Applied to that eF at the point (x_k,x_{k+1}) with x_{k+1}=G(x_k;theta_k), the adjoint equation must read (D1G(x_k;theta_k)-I)^* lambda = -partial J(x_{k+1})/partial x, and the split gradient must use partial G(x_k;theta_k)/partial theta. Algorithm 1 instead states D1F(x_{k+1};theta_k)^* lambda = ... with F=G-x, so D1G(x_{k+1})-I, and partial F(x_{k+1})/partial theta = partial G(x_{k+1})/partial theta. These agree only if D1G and partial G/partial theta are independent of their state argument, which is false for the linearized transport map in Section 4.2.1: the scattering coefficient is evaluated at phi_{k+1} in one case and phi_k in the other. Consequently, the constrained-case gradient used in the numerical experiments is not the split gradient whose construction is the paper's central methodological contribution; it is a heuristic standard-adjoint evaluation at an inexact state. The factor-of-three claim therefore rests on an algorithm whose derivation is not established, and Section 5 explicitly defers convergence and stability analysis, so nothing in the text bridges this gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a nonlinear splitting framework for gradient-based optimization. For unconstrained problems, it defines a splitting of the gradient as a function g(x,y) with g(x,x)=∇C(x), leading to semi-implicit gradient methods and accelerated variants (NS-GD, NS-NAG, NS-GD-Newton(1), NS-NAG-Newton(1), NS-AA-Newton(1)). For equality-constrained problems, it splits the constraint F(x,θ) and uses a modified adjoint to compute a 'split gradient' that does not require solving the constraint exactly, yielding algorithms NS-Adj-GD, NS-Adj-NAG, and NS-Adj-AA. Numerical experiments on nonconvex benchmark functions, quantum tomography, neutron transport source optimization, and an ODE-constrained problem demonstrate improved stability, large-step capability, and, in the transport example, a claimed reduction of high-dimensional kinetic solves by up to a factor of three compared with standard adjoint optimization.","tokens_in":23673,"tokens_out":14106,"duration_ms":115174,"significance":"If correct, the framework provides a general and potentially useful construction for semi-implicit gradient methods and inexact-adjoint optimization. The unconstrained part is elegant: it unifies existing methods such as proximal gradient and Levenberg–Marquardt as special cases, connects to discrete gradient methods, and the empirical large-step stability on nonconvex benchmarks is striking. The constrained part addresses a practically important bottleneck in PDE-constrained optimization. However, the significance is tempered by the absence of descent or convergence analysis, which the paper itself defers, and by an internal inconsistency between the derived modified adjoint and the algorithm used in the constrained-case experiments. These issues undermine the claimed factor-of-three result as a validation of the proposed systematic construction.","major_comments":[{"comment":"The modified adjoint construction applied to the splitting eF(x,y;θ)=G(x;θ)−y yields the adjoint equation (D1G(x_k;θ_k)−I)^*λ = −∂J(x_{k+1})/∂x and the split gradient ∂J/∂θ + ⟨λ, ∂G(x_k;θ_k)/∂θ⟩. Algorithm 1 instead states D1F(x_{k+1};θ_k)^*λ = −∂J(x_{k+1})/∂x and ∂F(x_{k+1};θ_k)/∂θ, which, with F(x;θ)=G(x;θ)−x, means evaluating D1G and ∂G/∂θ at x_{k+1} rather than x_k. These expressions agree only if G is affine in its state argument. For the linearized neutron transport map of §4.2.1, G depends on ψ_k through φ_k in the scattering cross-section, so that condition fails. Consequently, the algorithm whose numerical results are reported in §4.2.1 is not the split gradient derived in §3.2; it is a heuristic standard-adjoint evaluation at an inexact state. This disconnect directly affects the factor-of-three claim in the abstract.","section":"§3.2, Eqs. (3.25)–(3.27) and Algorithm 1"},{"comment":"No descent or convergence analysis is provided for any of the proposed methods. For NS-GD, the discrete-gradient property (Example 3.2) guarantees monotonic decrease only when the splitting satisfies the mean-value condition (3.6), which the linearly-implicit splittings of §4.1.1 do not. For NS-Adj-GD, e∇θJ is not the true gradient of the reduced objective when x_{k+1}≠x_k, and there is no guarantee that the update θ_{k+1}=θ_k−γe∇θJ decreases J. The paper explicitly defers this analysis to future work in Section 5. Given that the central claim is improved accuracy and efficiency, the absence of even a local descent result leaves the empirical gains unexplained and limits the generality of the proposed framework.","section":"§5 and §3.2"},{"comment":"The statement that 'the term ∂F/∂θ depending on both x_k and x_{k+1}' is inaccurate: for the splitting eF(x,y;θ)=G(x;θ)−y, the partial derivative with respect to θ depends only on the first argument x_k. This confusion appears to stem from conflating the original constraint with the fixed-point residual F(x;θ)=G(x;θ)−x. Please clarify which operator is used in (3.26b)–(3.26c) and whether the modified adjoint is derived from the splitting eF or from the redefined constraint F.","section":"§3.2, discussion after Eq. (3.26)"}],"minor_comments":[{"comment":"The abstract states that the total number of kinetic solves is reduced by a factor of three, while Figure 6(B) and the text report '2−3× less'; please make the quantitative claim consistent.","section":"Abstract and §4.2.1"},{"comment":"The Rastrigin splitting contains factors y_i/x_i, which are undefined at x_i=0. Please specify the treatment of this singularity or restrict the initial conditions to avoid it.","section":"§4.1.1, Rastrigin splitting"},{"comment":"The text '46 = 4098' should read 4^6 = 4096; correspondingly, the number of Pauli operators is 4096, not 4098.","section":"§4.1.2"},{"comment":"The runtime comparison with MATLAB built-in optimizers is difficult to interpret because the NS methods are implemented in unoptimized MATLAB loops while the built-ins are compiled; the authors acknowledge this in the text, but a caveat in the table caption would be helpful.","section":"Table 2"},{"comment":"The statement that a closed one-form on R^N has an associated cost function is true but could be stated more precisely, e.g., 'on a simply connected domain, a closed one-form is exact'.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between the derivation and Algorithm 1 is the main concern. If the authors fix the algorithm to match the derivation, the numerical experiments in §4.2.1 may need to be rerun, and the factor-of-three claim could change. I would also encourage the editor to ask for at least a preliminary convergence or descent result for a restricted class of splittings, as the paper currently makes no theoretical guarantees. The unconstrained part of the paper is well executed and could be resubmitted more quickly if separated from the constrained part."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the unconstrained nonlinear splitting part is neat and mostly solid: it unifies discrete gradients, Levenberg–Marquardt, and inertial LM, and the stability gains on the benchmarks and quantum tomography look plausible. Second, the constrained-case Algorithm 1 does not actually implement the derivation in Section 3.2, and the stress-test is right about that.\n\nThe general modified adjoint construction (3.27) is coherent: for the splitting eF(x,y)=G(x)-y, the correct adjoint uses (D1G(x_k)-I)^* and ∂G(x_k)/∂θ, evaluated at the pair (x_k, x_{k+1}). Algorithm 1 instead evaluates D1F and ∂F/∂θ at x_{k+1}=G(x_k). Those agree only if G's derivatives are state-independent, which they are not in the neutron transport example: the scattering coefficient is evaluated at φ_{k+1} instead of φ_k. So the numerical gains in Section 4.2.1 come from a heuristic 'adjoint at an inexact state,' not from the paper's systematic construction. That is a load-bearing inconsistency—the abstract's factor-of-three claim rests on it.\n\nThe paper does several things well. The unconstrained framework is clearly explained, the LM connection is a nice unification, and the numerical demonstrations are diverse and honestly reported (they even warn about slow MATLAB loops). The general construction in (3.27) is checkable and correct as far as it goes. The authors are also explicit that convergence and stability analysis are future work, which is honest but leaves the efficiency claims entirely empirical.\n\nOther soft spots: the tomography experiment omits the LM baseline even though LM is a special case; no code or data is provided; and the factor-of-three comparison counts forward solves only, excluding adjoint solves. These are fixable.\n\nI would send this to peer review—the framework is interesting and the unconstrained part deserves scrutiny—but not in current form. The authors need to either correct Algorithm 1 so that it matches the derivation, or reframe the constrained experiments as a heuristic with appropriate caveats and analysis. Add the LM baseline and release code/data, and this could become a useful paper.","headline":"A promising nonlinear splitting framework whose constrained-case Algorithm 1 does not follow from the paper's own derivation; the unconstrained part is cleaner and the paper deserves refereeing but needs real fixing.","tokens_in":24196,"tokens_out":6017,"would_cite":false,"duration_ms":51137,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C30","90C26","65K10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Nonlinear splitting—replacing a gradient or constraint by a two-argument function that agrees on the diagonal—turns explicit gradient methods into semi-implicit ones stable at large step sizes and, in a neutron transport test, needs one…","keywords":["nonlinear splitting","semi-implicit gradient methods","adjoint optimization","fixed-point acceleration","momentum acceleration","nonconvex optimization","PDE-constrained optimization","neutron transport"],"falsifier":"For the scalar problem $\\min_\\theta (x-\\theta)^2$ subject to $x=\\theta$, take the solver iteration $G(x;\\theta)=x+2(\\theta-x)$, which has the correct fixed point $\\theta$ but overshoots on every step. Applying Algorithm 1 at a state $x\\neq\\theta$ and checking whether the split-gradient direction $\\widetilde{\\nabla_\\theta J}(\\theta)[x,G(x;\\theta)]$ has negative inner product with $\\theta-\\theta^*$ settles the descent question directly: if that inner product is positive for any $\\theta$, the central premise of the constrained construction fails. This one-dimensional calculation can be done by hand.","tokens_in":23090,"feed_emoji":"📉","tokens_out":15057,"duration_ms":127863,"temperature":0.7,"pith_summary":"The paper introduces nonlinear splitting as a general construction for making gradient-based optimization semi-implicit. For an unconstrained cost $C$, the gradient $\\nabla C(x)$ is replaced by a smooth two-argument map $\\widetilde{\\nabla C}(x,y)$ with $\\widetilde{\\nabla C}(x,x)=\\nabla C(x)$, and the update $x_{k+1}=x_k-\\gamma_k\\widetilde{\\nabla C}(x_k,x_{k+1})$ requires only a linear or cheaply implicit solve for common splittings. For equality-constrained problems, the constraint $F(x;\\theta)=0$ is split in the same way, and a single iteration of any fixed-point solver induces a modified adjoint gradient that can be evaluated without solving the constraint exactly. On standard nonconvex benchmarks, quantum tomography, and a nonlinear neutron transport source problem, the resulting methods remain stable at large step sizes and reach the same objective with a fraction of the expensive solves, including a reported factor of three fewer kinetic solves. The paper does not yet supply a descent or convergence theorem; it states in the conclusion that stability and convergence analysis is deferred.","feed_headline":"Nonlinear splitting stabilizes gradient methods and triples efficiency","feed_subtitle":"Two-argument splits of gradients or constraints let optimizers take large steps and cut kinetic solves threefold.","key_machinery":"The load-bearing object is the nonlinear splitting: a smooth two-argument function $g(x,y)$ that coincides with the original function on the diagonal, $g(x,x)=f(x)$. Applied to gradients, it turns explicit descent into the semi-implicit NS-GD step; applied to the equality constraint, it yields $\\widetilde{F}(x,y;\\theta)$ with $\\widetilde{F}(x,x;\\theta)=F(x;\\theta)$, and the induced split gradient is defined through a modified adjoint system. The paper's main constructive device is the single iteration of a fixed-point solver: writing $x_{k+1}=G(x_k;\\theta_k)$ and setting $\\widetilde{F}(x,y;\\theta)=G(x;\\theta)-y$ converts any existing nonlinear solver into a constraint splitting, so the adjoint can be evaluated at a state that does not yet satisfy the constraint. A secondary device is the linearly-implicit gradient splitting $\\widetilde{\\nabla C}(x,y)=L(x)y+g(x)$, which makes each NS-GD step one linear solve, and the one-step linearized version of the implicit update, which produces a preconditioned gradient method with preconditioner determined by the splitting. These two devices together carry the reported gains in stability and in reduced solver work.","core_discovery":"The central discovery, stated as a construction, is that evaluating a gradient or an equality constraint on two arguments instead of one splits the original object and yields a family of semi-implicit optimization methods. In the unconstrained case, a nonlinearly split gradient $\\widetilde{\\nabla C}(x,y)$ with diagonal value $\\nabla C(x)$ gives the NS-GD update (3.2); when the splitting has the linearly-implicit form $L(x)y+g(x)$, each step is a single linear solve. A single linearized solve of the implicit step produces a preconditioned gradient descent whose preconditioner is determined by the chosen splitting, and for nonlinear least squares the split gradient $J(x)^\\top r(y)$ recovers the classical damped least-squares update as a special case. In the equality-constrained case, any nonlinear solver written as $x_{k+1}=G(x_k;\\theta_k)$ defines a constraint splitting $\\widetilde{F}(x,y;\\theta)=G(x;\\theta)-y$; using one solver iteration as the state update and solving a modified adjoint equation gives a split gradient that is consistent with the true adjoint gradient when $x_{k+1}$ actually solves the constraint. The paper demonstrates numerically that these methods take larger stable steps on three two-dimensional nonconvex benchmark functions, stabilize and converge in a 128-real-dimensional quantum tomography least-squares problem, reduce forward transport solves by a factor of two to three in the neutron transport example, and add a momentum-like term to the initial condition of an ODE-constrained problem that prevents divergence at large step size.","pith_inferences":["Editorial inference: If a broad descent condition can be proven for split gradients—for example, a monotonicity assumption that generalizes the mean-value property already known to make discrete gradients descent directions—then the construction would unify proximal gradient, damped least-squares, preconditioned gradient descent, and adjoint methods as instances of one recipe.","Editorial inference: The constraint-splitting recipe is transferable: any optimization loop that already contains a fixed-point solver for the state equation can switch to one-solver-iteration state updates and a modified adjoint, converting solver iterations into optimization progress. The neutron transport result suggests this can be tested in other expensive implicit-constraint settings, where ","Editorial inference: The diagonal condition $\\widetilde{\\nabla C}(x,x)=\\nabla C(x)$ is only consistency; all of the reported stability must come from off-diagonal behavior of the chosen splitting. A natural testable extension is to design splittings whose off-diagonal term equals a desired preconditioner, or a quadratic surrogate whose minimizer is the global minimizer, which the benchmark results","Editorial inference: Since the paper's theory is deferred, the safest reading is that the framework is a method generator whose effectiveness is demonstrated empirically on a handful of problems; the factor-three solve reduction is a single experiment, not yet a theorem, so a broader benchmark across initial conditions and problem scales would determine how general the gain is."],"forward_implications":["NS-GD and its momentum-accelerated variant remain stable at step sizes where explicit gradient and momentum methods diverge, allowing the iterates on nonconvex benchmark problems to escape local minima and reach the global minimum.","For nonlinear least squares, the split gradient $\\widetilde{\\nabla C}(x,y)=J(x)^\\top r(y)$ turns the one-linearized-solve variant of NS-GD into the classical damped least-squares update, and the fixed-point-accelerated version converges where standard gradient methods diverge at step size $\\gamma=10$.","In adjoint-based optimization, one nonlinear-solver iteration per optimization step is enough to produce a usable split gradient; the neutron transport experiments reach the same objective and constraint residual with two to three times fewer forward transport solves than adjoint methods that solve the constraint to tolerance $10^{-6}$ each step.","Splitting the equality constraint of an initial-value problem adds a momentum-like term to the initial condition, and the example shows that the resulting method ($\\alpha=0.5$ or $1$) reaches the minimum at a step size that destabilizes the standard method ($\\alpha=0$).","Both momentum and fixed-point acceleration compose with the nonlinear splitting construction, and the paper reports that the accelerated variants match or beat their standard counterparts in objective and runtime on the tested problems."],"supporting_citations":[{"why":"Supplies the general nonlinear solver framework in which one fixed-point iteration G(x;θ) induces a constraint splitting and modified adjoint gradient.","marker":"[14]"},{"why":"Supplies the nonlinearly split time-integration methods from which the nonlinear splitting construction is imported.","marker":"[15; 51]"},{"why":"Defines discrete gradients as nonlinear splittings with a mean-value property that guarantees monotone descent.","marker":"[20]"},{"why":"Provides the semi-implicit integrator ideas that motivate distributing implicitness through the second argument.","marker":"[11]"},{"why":"Provides the Lipschitz-gradient inequality that derives gradient descent and frames the explicit/implicit contrast.","marker":"[26, Lemma 54]"},{"why":"The classical damped least-squares algorithm recovered as a special case of the one-linearized-solve variant.","marker":"[30; 35]"},{"why":"The momentum-accelerated damped least-squares variant reproduced by the accelerated nonlinear splitting method.","marker":"[29]"},{"why":"Provides the momentum-accelerated gradient template used in the accelerated nonlinear splitting methods.","marker":"[37; 38]"},{"why":"Provides the fixed-point acceleration technique used in the fixed-point-accelerated variants.","marker":"[40; 44; 54]"}],"fun_headline_variants":["Split gradients, reduce kinetic solves by threefold","Nonlinear splitting stabilizes gradient methods","Two-argument gradients for stable, fast optimization","Nonlinear splitting accelerates adjoint and gradient methods","Semi-implicit splits enable larger stable steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single iteration of a nonlinear solver leaves the state slightly off the constraint yet still produces a split gradient that points downhill, so that updating the parameters along that direction decreases the objective; the paper gives no descent or convergence theorem for this, and Section 5 explicitly defers stability and convergence analysis.","fun_headline_variants_meta":{"raw":{"variants":["Split gradients, reduce kinetic solves by threefold","Nonlinear splitting stabilizes gradient methods","Two-argument gradients for stable, fast optimization","Nonlinear splitting accelerates adjoint and gradient methods","Semi-implicit splits enable larger stable steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000903,"raw_usage":{"total_tokens":3950,"prompt_tokens":1075,"completion_tokens":2875,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":691,"completion_tokens_details":{"reasoning_tokens":2806}},"tokens_in":691,"tokens_out":2875,"duration_ms":21435,"temperature":1.0,"reasoning_tokens":2806,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:47:26.384398+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the scalar problem $\\min_\\theta (x-\\theta)^2$ subject to $x=\\theta$, take the solver iteration $G(x;\\theta)=x+2(\\theta-x)$, which has the correct fixed point $\\theta$ but overshoots on every step. Applying Algorithm 1 at a state $x\\neq\\theta$ and checking whether the split-gradient direction $\\widetilde{\\nabla_\\theta J}(\\theta)[x,G(x;\\theta)]$ has negative inner product with $\\theta-\\theta^*$ settles the descent question directly: if that inner product is positive for any $\\theta$, the central premise of the constrained construction fails. This one-dimensional calculation can be done by hand.","supporting_citations":[{"cited_title":"Brune, Matthew G","cited_arxiv_id":null,"evidence_quote":"Supplies the general nonlinear solver framework in which one fixed-point iteration G(x;θ) induces a constraint splitting and modified adjoint gradient."},{"cited_title":"A geometric integration approach to smooth optimization: foundations of the discrete gradient method","cited_arxiv_id":null,"evidence_quote":"Defines discrete gradients as nonlinear splittings with a mean-value property that guarantees monotone descent."},{"cited_title":"Boscarino, F","cited_arxiv_id":null,"evidence_quote":"Provides the semi-implicit integrator ideas that motivate distributing implicitness through the second argument."},{"cited_title":"On inertial Levenberg-Marquardt type methods for solving nonlinear ill-posed operator equations","cited_arxiv_id":"2406.07044","evidence_quote":"The momentum-accelerated damped least-squares variant reproduced by the accelerated nonlinear splitting method."}],"review_version":1}