{"id":"bf2310d9-51ba-4c2a-94b8-5ff97abb2524","arxiv_id":"2504.15753","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The Green's function for an LTV Itô diffusion with quadratic state killing is a Gaussian in the endpoints whose quadratic form comes from a Riccati ODE.","lead":"Closed-form Markov kernels are derived for linear time-varying diffusions with quadratic killing of probability mass, generalizing earlier noise-only results. This makes linear-quadratic Schrödinger bridge problems with arbitrary non-Gaussian endpoints solvable via explicit Sinkhorn recursions rather than Monte Carlo.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The main theorem depends on the unproved identity dM/dt=S; Appendix C calls it a by-product but never proves it, so Theorem 1 is not yet established.","rationale":"The reader's conditional verdict correctly identifies the unproven identity and the limit claim as the weakest points. Among these, dot M=S is the more fundamental: it is a necessary condition for the Gaussian ansatz to satisfy the PDE (33a) at all, and it is asserted in Appendix C as a by-product without proof. The paper does not verify it from the definitions, and it cannot be derived merely by equating terms unless one already assumes the ansatz solves the PDE. A direct differentiation check is routine but nontrivial; until it is performed, Theorem 1 remains conditional. The limit in (35) is secondary but also asserted in Remark 3 rather than proved. I do not see a deeper flaw: the distance-from-OCP construction, the Riccati substitution, and the recovery of known special cases are coherent. The appropriate disposition is to keep the paper conditional pending the missing algebra, which is exactly the reader's verdict.","tokens_in":22114,"tokens_out":7484,"duration_ms":72180,"concrete_test":"Verify dot M_{tt0}=S_{tt0} directly. Differentiate the three blocks of (30) with respect to t, using d Phi_{tt0}/dt = hat A_t Phi_{tt0}, d Gamma_{tt0}/dt = hat B_t hat B_t^T + hat A_t Gamma_{tt0} + Gamma_{tt0} hat A_t^T, and the t-derivative of the Riccati solution Pi(t0,0,t) with Pi(t,t)=0; substitute into (72a)-(72c). This can be checked symbolically for n=1 and numerically by finite differences for random LTV triples (A_t,B_t,Q_t). If the identity holds, the proof gap is closed; if not, Eq. (36) fails to solve (33).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing unproved step is in Step 1 of Appendix C. Substituting the Gaussian ansatz (32) into the PDE (33a) leads to Eq. (71), whose left side contains dot M_{tt0} and whose right side contains S_{tt0}. The argument then equates spatially independent and quadratic terms; the quadratic equality is exactly the identity dot M_{tt0}=S_{tt0}. The paper labels this as a \"by-product\" and says it will not be used hereafter, but no proof is supplied. Without this identity, the ansatz is not shown to satisfy the PDE, so Theorem 1 is not established. The existence claim for the limit in (35), asserted in Remark 3, addresses only the initial condition; even if granted, the identity dot M=S is needed before the prefactor can be justified. Thus the central claim rests on an unverified matrix ODE identity that is not derived from the definitions of hat Phi, hat Gamma, and Pi.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives a closed-form Markov kernel for a linear time-varying Itô diffusion with an additional quadratic killing rate, i.e. the Green's function for the linear reaction-advection-diffusion PDE (33). The derivation follows the template: Markov kernel ← distance function ← deterministic optimal control problem. The distance is obtained as the value of a finite-horizon LQ optimal control problem, and the prefactor is fixed by the Dirac-delta initial condition. The main result, Theorem 1, gives the kernel in Eq. (36) in terms of the solution of a Riccati matrix ODE. The authors then explain how this kernel enables dynamic Sinkhorn recursions for the linear quadratic non-Gaussian Schrödinger bridge. They also show that known kernels in the literature are recovered as special cases in Proposition 3 and Corollary 2.","tokens_in":22342,"tokens_out":6611,"duration_ms":60639,"significance":"If the result is fully established, this is a valuable contribution to stochastic control and Schrödinger bridge theory. It generalizes previously known kernels for the zero-drift, identity-diffusion case to general controllable LTV systems with time-varying quadratic killing, and it provides a systematic method ('Markov kernel from optimal control distance') that goes beyond Hermite-polynomial and Weyl-calculus approaches. The paper also demonstrates the usefulness of the kernel for solving non-Gaussian distribution steering problems with an explicit mixture-of-Gaussians example. However, the proof of the main theorem currently has load-bearing gaps: an unproved matrix identity and an unproved limit assertion. These gaps must be addressed before the central claim can be regarded as established.","major_comments":[{"comment":"The proof asserts 'As a by-product of the above calculation, we get ˙Mtt0 = Stt0, but this will not be used hereafter.' This identity is not proved, yet it is needed to verify that the Gaussian ansatz (32) satisfies the PDE (33a) for all x and y: equating the quadratic terms in (71) is exactly the identity ˙Mtt0 = Stt0. Without a derivation of this identity from the definitions of Φ̂tt0, Γ̂tt0, and Π(t0,0,t), Theorem 1 is not established. Please supply a proof, for example by differentiating (30) and using the Riccati equation (25) together with the transition-matrix and Gramian identities.","section":"Appendix C, Step 1, Eq. (71)"},{"comment":"The existence, finiteness, and positivity of the limit defining the constant a is only asserted in Remark 3, with the statement that it is 'not too difficult to show'. This limit is load-bearing because it determines the prefactor of the kernel and the positivity property κ>0. Please provide a complete proof of the existence and positivity of this limit under Assumptions A1–A2. The special-case computation in Appendix D does not cover the general time-varying case.","section":"Theorem 1, Eq. (35), Remark 3"},{"comment":"The interchange of limit and integral is justified by saying the integrand is bounded in x, but boundedness alone does not provide an integrable dominating function independent of t as t↓t0, so the dominated convergence theorem is not applicable as stated. Since the Gaussian integral is available in closed form via Lemma 1 for each fixed t, the evaluation in (77) can be obtained without exchanging the limit and the integral. Please either remove the DCT step and evaluate the integral first, or supply a valid uniform integrability argument.","section":"Appendix C, Step 2, Eqs. (76)–(77)"}],"minor_comments":[{"comment":"The displayed computation of θ(s) appears to have a typo: the argument of coth should be 2√Dii(s−t0) rather than √Dii(s−t0), to be consistent with the following integral and with ωi = 2√Dii in the final result.","section":"Appendix D, Eq. (83)"},{"comment":"References [9] and [10] appear to be duplicate citations of the same ACC 2015 paper by Chen, Georgiou, and Pavon; please merge them.","section":"References"},{"comment":"The abstract says the linear quadratic non-Gaussian Schrödinger bridge is 'exactly solvable', but the proposed dynamic Sinkhorn recursion is an iterative algorithm; this wording could be misread as a fully closed-form solution. Consider rephrasing to 'the kernel is explicit, making the dynamic Sinkhorn recursion implementable'.","section":"Abstract and Introduction"},{"comment":"The operator notation changes from 'BtB⊤t ∆xκ' in the introduction to '⟨BtB⊤t, ∇²xκ⟩' in the PDE (33a); these are consistent for scalar functions, but the equivalence could be stated explicitly to avoid confusion.","section":"Introduction, Eq. (4) and Eq. (33a)"},{"comment":"Lemma 1 is referenced in the text before it is stated in Appendix A; consider adding a forward reference or moving the lemma to an earlier position.","section":"Section IV-C, after Eq. (43)"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly a continuation of the authors' own prior work on Schrödinger bridges with quadratic cost, and the new generalization to LTV dynamics is the main contribution. The proof gaps in Appendix C are the central obstacle; they are serious but appear fixable within the manuscript's scope. The paper would also benefit from removing the duplicate reference and tightening the 'exactly solvable' wording."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this paper derives a genuinely new closed-form Markov kernel for an LTV diffusion with quadratic killing, and uses it to make non-Gaussian LQ Schrödinger bridges solvable by Sinkhorn iteration. The claimed generalization is real—the kernel in Theorem 1 covers generic controllable (A_t,B_t) and time-varying Q_t, and the earlier kernels from [5],[8] fall out as special cases. The new technique, finding the distance in the exponential from a deterministic OCP, is a clean idea and it works. The paper is well-written and the derivation is self-contained up to two gaps.\n\nWhat I like: the kernel is not just a guess. They derive the distance from a min-energy control problem, then verify the Gaussian ansatz against the PDE and fix the prefactor via the Dirac-delta initial condition. The special-case recovery in Proposition 3 and Corollary 2 is strong evidence. The Sinkhorn application is practical, and the mixture-of-Gaussians example shows the kernel is actually usable. The citation pattern is fine: earlier work by the same group is the natural baseline and is cited as such.\n\nThe soft spots are real. In Appendix C, Step 1, equation (71) holds only if the quadratic forms match, i.e., dot M = S. The paper calls this a \"by-product\" and never proves it. That identity is load-bearing: without it, you cannot separate the constant terms and get the ODE for c. This is not a cosmetic omission. The limit defining a in (35) is also only asserted in Remark 3; no proof or reference is given. Recovering the known kernels suggests the formula is right, but a closed-form Green's function needs those two steps nailed down. They may be routine—matrix algebra plus Riccati theory—but right now Theorem 1 rests on an unverified identity.\n\nWho should read this: anyone working on Schrödinger bridges, stochastic control with distribution steering, or exact kernels for reaction-diffusion PDEs. It deserves a serious referee. My recommendation: send it to peer review, and ask the referee to verify dot M = S from the definitions and to supply a proof of the limit in (35). If those check out, this is a publishable result. If they don't, the kernel formula needs revision.","headline":"A genuinely new Markov kernel for LQ non-Gaussian steering, but the main theorem's proof rests on an unproved matrix identity that a referee must check.","tokens_in":22811,"tokens_out":3061,"would_cite":true,"duration_ms":29126,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N10","93E20","60J60","35K57","60H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A closed-form Markov kernel is derived for a linear time-varying diffusion with quadratic killing of probability mass, making the linear-quadratic non-Gaussian Schrödinger bridge exactly solvable.","keywords":["Markov kernel","Schrödinger bridge","linear quadratic control","non-Gaussian distribution steering","reaction-advection-diffusion PDE","Riccati matrix ODE","optimal control distance","Sinkhorn recursion"],"falsifier":"Choose a nontrivial scalar example, say $A_t=-t$, $B_t=1$, $Q_t=1+t^2$; solve the Riccati ODE (25) numerically, form $M_{tt_0}$ from (30), assemble the kernel (36), and test directly whether it satisfies the reaction-advection-diffusion PDE (33) for all $x,y$ and whether the $t\\downarrow t_0$ limit is a Dirac delta. The claim fails if the residual is not identically zero or if the limit in (35) is not finite and positive.","tokens_in":21931,"feed_emoji":"🎛️","tokens_out":15057,"duration_ms":119767,"temperature":0.7,"pith_summary":"The paper derives a closed-form Markov kernel for a controlled linear time-varying diffusion whose probability mass is killed at a time-varying quadratic rate. The kernel is the Green's function of a linear reaction-advection-diffusion PDE, written as a Gaussian in the endpoint pair whose matrix is built from a solution of a Riccati matrix ODE. If correct, this makes the linear-quadratic non-Gaussian Schrödinger bridge exactly solvable: any endpoint distributions with finite second moments can be steered to each other by dynamic Sinkhorn recursions that use kernel-based integral transforms instead of path-integral Monte Carlo. The derivation rests on a template: postulate the kernel as a distance-based exponential, obtain the distance from a deterministic optimal control problem, then fix the prefactor from the Dirac-delta initial condition.","feed_headline":"Exact kernel solves quadratic Schrödinger bridges","feed_subtitle":"A closed-form Green's function makes distribution steering with arbitrary endpoints a Sinkhorn recursion.","key_machinery":"The carrying machinery is the template 'Markov kernel $\\leftarrow$ distance $\\leftarrow$ deterministic optimal control problem.' The paper postulates $\\kappa=c(t,t_0)\\exp(-\\tfrac12\\mathrm{dist}_{tt_0}^2(x,y))$, defines $\\mathrm{dist}$ as the value of the LQ optimal control problem (24), and uses the Riccati ODE (25) together with Proposition 1 to convert the soft quadratic state cost into the modified LTV system $\\hat A_\\tau=A_\\tau-\\hat B_\\tau\\hat B_\\tau^\\top\\Pi(\\tau,0,t)$. The distance then becomes the quadratic form governed by the matrix $M_{tt_0}$ in (30). The prefactor is not fixed by normalization; it is obtained by substituting the ansatz into the PDE and matching the Dirac-delta initial condition, which yields $\\dot c=-\\theta(t)c$ and the constant $a$ in (35). The proof also relies on the by-product identity $\\dot M_{tt_0}=S_{tt_0}$ to separate the spatially independent and quadratic terms.","core_discovery":"Under assumptions A1 and A2, the paper claims that the Markov kernel solving the reaction-advection-diffusion initial value problem (33) is\n$$\\kappa(t_0,x,t,y)=a\\,$e^{{-\\int_{t_0}}$^{t}\\$\\theta$(s)\\,ds}\\exp\\big(-\\tfrac12\\binom{x}{y}^{\\top} M_{tt_0}\\binom{x}{y}\\big),$$\nwhere $M_{tt_0}$ is the positive-definite matrix in (30) formed from the modified transition matrix $\\hat\\Phi_{tt_0}$, the controllability Gramian $\\hat\\Gamma_{tt_0}$, and the Riccati solution $\\Pi(t_0,0,t)$; $\\theta$ is defined by (34) and the prefactor $a$ by the limit (35). This is the Green's function for the forward Kolmogorov-type operator with advection, diffusion, and quadratic killing. As direct consequences, the kernel specializes to the heat kernel, the linear LTV kernel without killing, and the zero-drift constant-$Q$ kernel of previous work, and it makes the dynamic Sinkhorn iteration (41) implementable through the closed-form integral transforms (37).","pith_inferences":["The same distance-from-optimal-control template may extend to non-quadratic killing rates whenever the associated deterministic optimal control problem has an explicit value function; the paper does not attempt that extension.","Since the kernel is a Gaussian in the endpoint pair, approximating endpoint densities by conic mixtures of Gaussians yields closed-form Sinkhorn updates, and Example 1 supplies the needed formula; this suggests a practical fixed-point algorithm only sketched in the paper.","The by-product identity $\\dot M_{tt_0}=S_{tt_0}$ could itself serve as an evolution law for the kernel's covariance, potentially bypassing recomputation of the Riccati solution at every time step.","Pairing the distance form with Varadhan-type short-time asymptotics suggests that $\\mathrm{dist}_{tt_0}$ is the geodesic distance of a metric associated with the modified controlled dynamics, a geometric reading the authors flag but do not develop."],"forward_implications":["The LQ non-Gaussian Schrödinger bridge is exactly solvable: the optimal controlled density can be produced by the dynamic Sinkhorn recursion (41), with each pass applying the kernel-based transforms (37).","Endpoint distributions with finite second moments are allowed; no Gaussianity or other structural restriction on the marginals is needed.","The kernel unifies previously disjoint cases: the Euclidean heat kernel, the LTV advection-diffusion kernel, and the constant-Q zero-drift kernel of prior work all appear as specializations of (36).","For mixture-of-Gaussian endpoint functions the Sinkhorn updates can be evaluated in closed form, giving a concrete numerical recipe.","Because the kernel is explicit, the Feynman-Kac/Monte Carlo approximation for each Sinkhorn pass is no longer needed."],"supporting_citations":[{"why":"Derives the special-case kernel for zero drift and constant quadratic killing via Hermite polynomials; the present result must generalize it.","marker":"[5]"},{"why":"Derives the same special case via Weyl calculus; the present kernel generalizes and reduces to that formula.","marker":"[8]"},{"why":"Gives the Gaussian-endpoint steering solution that this paper extends to non-Gaussian endpoints with finite second moments.","marker":"[11]"},{"why":"Supplies the transition-probability kernel for the LTV advection-diffusion case without killing, a special case recovered by the new distance template.","marker":"[14]"},{"why":"Provides the standard minimum-effort state steering solution used to identify the distance in the linear-kernel exemplar.","marker":"[24]"},{"why":"Gives Proposition 1 and the Riccati linear-fractional machinery that changes the soft quadratic cost into a modified LTV system.","marker":"[29]"},{"why":"Cited for existence, uniqueness, and positive semidefiniteness of the Riccati solution, assumptions behind the distance formula.","marker":"[30]"},{"why":"Cited for invariance of controllability under state feedback, ensuring the modified Gramian used in (30) is nonsingular.","marker":"[32]"},{"why":"Gives the Hilbert-metric contraction result that guarantees convergence of the dynamic Sinkhorn recursion applied with the kernel.","marker":"[36]"}],"fun_headline_variants":["Exact Green's function for LTV Schrödinger bridges","Closed-form kernel enables exact Sinkhorn recursions","Riccati ODE yields closed-form kernel for LTV diffusion","Sinkhorn recursions exact for non-Gaussian control","Generalized kernel solves Schrödinger bridge for LTV"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the limit defining the constant $a$ in (35) exists and is positive, and that the derivative identity $\\dot M_{tt_0}=S_{tt_0}$ holds; the paper asserts existence in Remark 3 and states the identity as a by-product without a standalone proof, so the closed-form kernel stands or falls on these two points.","fun_headline_variants_meta":{"raw":{"variants":["Exact Green's function for LTV Schrödinger bridges","Closed-form kernel enables exact Sinkhorn recursions","Riccati ODE yields closed-form kernel for LTV diffusion","Sinkhorn recursions exact for non-Gaussian control","Generalized kernel solves Schrödinger bridge for LTV"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000743,"raw_usage":{"total_tokens":3422,"prompt_tokens":1158,"completion_tokens":2264,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":774,"completion_tokens_details":{"reasoning_tokens":2183}},"tokens_in":774,"tokens_out":2264,"duration_ms":16618,"temperature":1.0,"reasoning_tokens":2183,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:18:22.621902+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a nontrivial scalar example, say $A_t=-t$, $B_t=1$, $Q_t=1+t^2$; solve the Riccati ODE (25) numerically, form $M_{tt_0}$ from (30), assemble the kernel (36), and test directly whether it satisfies the reaction-advection-diffusion PDE (33) for all $x,y$ and whether the $t\\downarrow t_0$ limit is a Dirac delta. The claim fails if the residual is not identically zero or if the limit in (35) is not finite and positive.","supporting_citations":[{"cited_title":"Schr\\\"{o}dinger Bridge with Quadratic State Cost is Exactly Solvable","cited_arxiv_id":"2406.00503","evidence_quote":"Derives the special-case kernel for zero drift and constant quadratic killing via Hermite polynomials; the present result must generalize it."},{"cited_title":"Weyl calculus and exa ctly solvable Schr¨ odinger bridges with quadratic state cost,","cited_arxiv_id":null,"evidence_quote":"Derives the same special case via Weyl calculus; the present kernel generalizes and reduces to that formula."},{"cited_title":"Optimal steering of a linear stochastic system to a final probability distribution—part iii,","cited_arxiv_id":null,"evidence_quote":"Gives the Gaussian-endpoint steering solution that this paper extends to non-Gaussian endpoints with finite second moments."},{"cited_title":"Optimal steering of a linear stochastic system to a final probability distribution, part i,","cited_arxiv_id":null,"evidence_quote":"Supplies the transition-probability kernel for the LTV advection-diffusion case without killing, a special case recovered by the new distance template."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the standard minimum-effort state steering solution used to identify the distance in the linear-kernel exemplar."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives Proposition 1 and the Riccati linear-fractional machinery that changes the soft quadratic cost into a modified LTV system."},{"cited_title":"A review of the matrix Riccati equation,","cited_arxiv_id":null,"evidence_quote":"Cited for existence, uniqueness, and positive semidefiniteness of the Riccati solution, assumptions behind the distance formula."},{"cited_title":"On invariance of degre e of control- lability under state feedback,","cited_arxiv_id":null,"evidence_quote":"Cited for invariance of controllability under state feedback, ensuring the modified Gramian used in (30) is nonsingular."},{"cited_title":"Entropic and displa cement interpolation: a computational approach using the Hilbert metric,","cited_arxiv_id":null,"evidence_quote":"Gives the Hilbert-metric contraction result that guarantees convergence of the dynamic Sinkhorn recursion applied with the kernel."}],"review_version":1}