{"id":"688d3696-720a-41ae-bf38-692a9b65398a","arxiv_id":"2411.13765","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The Schrödinger bridge for a jump diffusion is represented and approximated by harmonic h-transforms, with a PIDE formulation for smooth jump-diffusion flows.","lead":"This paper extends the Schrödinger bridge problem from continuous diffusions to random motions with jumps, expressing the bridge as a limit of harmonic h-transforms. It provides a rigorous foundation for modeling the most likely jumpy dynamics between two observed distributions, with applications in finance and generative modeling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Strong-convergence theorem is conditional on Assumption (A3); for stable-like jump diffusions only finite-dimensional convergence with mollified references is proved, so the abstract's 'mild assumptions' overstate the scope.","rationale":"The reader's weakest assumption, (A3), is indeed the condition that makes the approximating h_k harmonic, and the paper concedes it is scarce for jump diffusions. I partially agree with the reader's framing: the strong convergence of the measures in Theorem 4.4 does not actually require harmonicity. The functions h_k defined by (4.8) always enjoy the mean-value property by the Markov property, and the Scheffé/L^1 argument in Theorem 4.4 goes through without any C^{1,2} regularity. Thus (A3) is load-bearing for the claim that the approximants are harmonic h-transforms whose generators and SDEs are given by Section 3, not for the raw total-variation convergence of the sequence. The more serious scope issue is that the stable-like operators in Section 5.3 are explicitly outside (A3), and for them the paper proves only finite-dimensional convergence of h-transforms of mollified reference measures, not strong convergence of h-transforms of the original reference. That is a real gap between the abstract's 'mild assumptions' / 'relatively general class' and what is actually proven. I did not find an internal inconsistency in the main computations: the h-transform drift and jump-intensity formulas in Lemma 3.7 are consistent, the Poisson example in Section 4.1 checks out, and the Girsanov signs in Theorem 3.9/Corollary 3.10 are coherent once the minus signs in (3.14) and (3.15) are tracked. The conditional verdict remains appropriate: the paper is a useful but assumption-heavy extension, and the claimed breadth should be stated more cautiously.","tokens_in":39120,"tokens_out":34645,"duration_ms":313220,"concrete_test":"Fix a one-dimensional stable-like reference measure from Section 5.3, for instance a≡1, b≡0, κ≡1, α∈(0,2), and a fixed g∈C_c^∞(R). Check analytically or numerically whether h(t,x)=∫g(y)p(t,x,T,y)dy is C^{1,2} and satisfies (∂_t+L)h=0 on (0,T)×R. If A3 fails or cannot be verified, Theorem 4.4 does not apply to the stable-like class, confirming the scope gap. If A3 holds in this constant-coefficient case, repeat with a Hölder-continuous variable coefficient κ(t,x) to see whether the smoothing property persists; the outcome determines whether the strong-convergence claim can be extended beyond the Kunita-flow setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result (Theorem 4.4) is that the SBP solution \\hat P is the strong limit of harmonic h-transforms. The only condition making the approximants h_k harmonic is Assumption (A3): for every g in C_c^\\infty, h(t,x)=∫g(y)P_{t,T}(x,dy) must be C^{1,2} and solve the backward equation. (A3) is not implied by (A1)/(A2) or by existence of a transition density; it is a genuine smoothing assumption. Section 5.3 explicitly says the stable-like operators considered there 'may not support (A3)', and for those operators Theorem 5.6 proves only convergence in finite-dimensional distributions. Moreover, the approximating Q_k there are h-transforms of the mollified reference measures P_k (associated with L_{m_k}), not h-transforms of the original reference R. Thus the strong-convergence statement is not established for the stable-like class the paper highlights; for that class the abstract's 'strong-convergence limit' is replaced by f.d.d. convergence of h-transforms of a different process. This is a scope gap relative to the claimed 'mild assumptions'. It is worth noting that the measure-convergence step in Theorem 4.4 itself uses only the automatic mean-value property of h_k, not their harmonicity; the load-bearing part of (A3) is specifically the harmonicity that lets the approximants be interpreted through the h-transform generator and SDE theory of Section 3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the Schrödinger bridge problem (SBP) when the reference measure is the law of a jump diffusion. It develops an h-transform theory for jump diffusions in two complementary formulations: a martingale-problem/generator approach under Assumption (A1) and an SDE/Girsanov approach under Assumption (A2). The transformed drift and jump measure are given explicitly in Lemma 3.7, and the corresponding martingale problem and SDE statements appear in Theorem 3.8 and Corollary 3.10. The paper then connects the SBP solution to an h-transform: under a regularity assumption (A3), Theorem 4.4 shows that the SBP solution is the strong (total-variation) limit of h-transforms by harmonic functions, with the approximating functions constructed explicitly from the Schrödinger system data. Under additional smoothness assumptions and existence of a transition density, the paper derives a dynamic Schrödinger system of PIDEs and a product-form marginal density (Theorems 5.3 and 5.4). Finally, for α-stable-like jump operators that may fail (A3), a mollification procedure is proposed and convergence in finite-dimensional distributions is claimed in Theorem 5.6.","tokens_in":39445,"tokens_out":29307,"duration_ms":278983,"significance":"If the results hold as stated, this is a useful contribution to the Schrödinger bridge literature, which has largely been confined to diffusion references. The paper correctly identifies the role of transition-semigroup regularity, gives explicit formulas for the transformed generator and SDE, and provides a constructive approximation scheme. The proof of Theorem 4.4 is coherent, and the construction is genuinely data-driven: the approximating functions g_k are built from R, ρ0 and ρT, not assumed to solve the target problem. The stable-like extension is potentially valuable but is the weakest part of the paper, and its current proof and framing need work. The paper does not provide code or machine-checked proofs; its value is theoretical.","major_comments":[{"comment":"The abstract's claim of obtaining the SBP solution 'under mild assumptions' as 'the strong-convergence limit of a sequence of harmonic h-transforms' is not supported for the advertised general jump-diffusion class. Assumption (A3), stated in Section 1.2 around Eq. (1.4), is a strong C^{1,2} regularity condition on all smoothed conditional expectations; it is not implied by (A1)/(A2), and the paper itself notes in Section 5.3 that the stable-like operators of Eq. (5.8) may not support (A3). For that class, Theorem 5.6 proves only convergence in finite-dimensional distributions, and the approximating Q_k are h-transforms of the mollified measures P_k associated with L_{m_k}, not h-transforms of the original reference R. The abstract and Section 1.2 should be revised to state explicitly that strong convergence of harmonic h-transforms requires (A3).","section":"Abstract and Section 1.2"},{"comment":"The proof of finite-dimensional convergence in Theorem 5.6 is incomplete. After establishing pointwise convergence of h_k^{(m_k)} to h (Lemma 5.5) and uniform-on-compact convergence of p_{m_k} to p, the displayed formula for Q_k(∩_{i=0}^N {X_{t_i}∈B_i}) is passed to the limit without a dominated-convergence or uniform-integrability argument. The ratio h_k^{(m_k)}(T,x_N)/h_k^{(m_k)}(0,x_0) may be large where h(0,x_0) is small, and no bound is supplied; the same issue affects the assertion that r_k→1. A rigorous argument is needed, for example by using the explicit construction of g_k and the mean-value property to obtain L1 control, or by restricting to compact sets and appealing to tightness. This is load-bearing because Theorem 5.6 is the paper's only positive statement for the stable-like class.","section":"Section 5.3"}],"minor_comments":[{"comment":"The proof of strong convergence uses only the mean-value property of h_k, which is automatic from the definition h_k(t,x)=∫g_k(y)P_{t,T}(x,dy); the harmonicity supplied by (A3) is not used in the L1-convergence argument. The paper should state explicitly that the measure-convergence part does not require (A3) and that (A3) is needed only to interpret the approximants as harmonic h-transforms with the generator/SDE theory of Section 3.","section":"Theorem 4.4"},{"comment":"The definition of Phk in Eq. (4.9) divides by r_k, which may vanish for finitely many k; since the proof shows r_k→1, the statement should be restricted to k sufficiently large or allow discarding finitely many terms.","section":"Theorem 4.4"},{"comment":"The Girsanov representation is cleaner if the base measure is taken to be \\tilde P = 1_{X0∈A0}/r0 P; with the current normalization 1_{τ0>T}/r0 P, the condition P(τ0>T)=r0 in Theorem 3.9 is presented as necessary for the Girsanov interpretation, while Corollary 3.10 does not restate it. Please clarify the normalization and state explicitly that Corollary 3.10 is understood as a Girsanov transformation with respect to \\tilde P.","section":"Theorem 3.9"},{"comment":"There is a typo in the sentence after Lemma 3.4: 'andn which' should read 'and which'.","section":"Section 3.1"},{"comment":"The conditional expectation E_R0T[g_k(y)|x] is defined only for R0-a.e. x; writing 'for every x∈Rn' is an abuse of notation and should be qualified.","section":"Lemma 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's conditional results appear mathematically sound, and the main issue is the gap between the abstract's claims and what is actually proved for the stable-like class. I recommend major revision to fix the proof of Theorem 5.6 and to recalibrate the statements of the abstract and Section 1.2. The proof of Theorem 5.6 currently lacks a justification for the passage to the limit, which is essential for the paper's advertised extension to stable-like jump diffusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does what it claims in the large: it carries the Schrödinger bridge machinery from diffusions to jump diffusions, and the main theorems are mostly proved. The genuinely new pieces are the explicit h-transform generator transformation (Lemma 3.7), the strong-convergence approximation of the bridge by harmonic h-transforms under Assumption (A3) (Theorem 4.4), the PIDE formulation for smooth stochastic flows (Theorem 5.3), and the honest finite-dimensional convergence result for stable-like operators (Theorem 5.6). The construction of the approximating functions from the Schrödinger data is explicit and usable, and the paper carefully handles non-negative h and zero sets. The Poisson example is a nice sanity check. Citations to the diffusion SBP and jump-process literature look fair and appropriate.\n\nThe soft spot is the gap between the abstract and the theorems. Calling (A3) a 'mild assumption' is too generous: it requires the semigroup applied to every smooth compactly supported g to produce a C^{1,2} solution of the backward equation. The paper itself concedes this is scarce for jump diffusions. The strong convergence of Theorem 4.4 rests on it. For the stable-like class that the paper highlights, Theorem 5.6 gives only convergence in finite-dimensional distributions, and the approximating Q_k are h-transforms of the mollified reference measures P_k, not of the original reference R. So the abstract's 'strong-convergence limit' is not established for that class. This is a framing issue rather than a hidden mathematical error; the proofs themselves look coherent, and the paper is open about the limitation in Section 5.3. The fix is straightforward: temper the abstract and state the scope precisely.\n\nMinor nit: the reader's report flags that Corollary 3.10 does not restate a condition from Theorem 3.9. I checked and the 'under the same setting' wording does cover it, so this is a readability issue, not a real gap.\n\nWho is this for? People working on Schrödinger bridges, stochastic control with jumps, or generative modeling with jump processes will get a rigorous but conventional extension rather than a conceptual break. The paper deserves a serious referee. I would send it to review and ask the authors to align the abstract with what is actually proved, but the core is solid and publishable after that revision.","headline":"Solid jump-diffusion Schrödinger bridge theory with a real approximation theorem, but the abstract oversells Assumption (A3): for the stable-like class only finite-dimensional convergence of h-transforms of a mollified reference is proved.","tokens_in":39927,"tokens_out":5749,"would_cite":true,"duration_ms":933718,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35Q93","45K05","60H10","60H20","60H30","94A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that for jump-diffusion reference measures, Schrödinger bridges are strong limits of h-transforms by harmonic functions, with explicit drift and jump-measure formulas.","keywords":["Schrödinger bridges","h-transform","jump diffusions","non-local Lévy-type operators","KL divergence","Schrödinger system","martingale problem","stable-like processes"],"falsifier":"Take the α-stable-like jump operator of Section 5.3 with a discontinuous Schrödinger-system function g and check whether sup_A |P_{h_k}(A)-\\hat{P}(A)| tends to 0; the paper proves only finite-dimensional convergence in that case, so exhibiting a counterexample to the vanishing total-variation distance would show the strong-convergence claim depends essentially on (A3).","tokens_in":38904,"feed_emoji":"🌉","tokens_out":8673,"duration_ms":85873,"temperature":0.7,"pith_summary":"The paper tackles the Schrödinger bridge problem when the reference process is a jump diffusion: given initial and terminal distributions, find the path measure closest to the reference in relative entropy. It establishes that the optimal bridge is an $h$-transform of a modified reference, and when the bridge function is not harmonic, it can still be reached as the strong limit of $h$-transforms by genuinely harmonic functions. This matters because the classical diffusion theory of Schrödinger bridges does not automatically carry over: jump diffusions have non-local generators and limited regularity theory. The result gives explicit formulas for the transformed generator, including the new drift and the rescaled jump measure, and extends the dynamic Schrödinger system to the jump setting.","feed_headline":"Jump-diffusion Schrödinger bridges are limits of harmonic h-transforms","feed_subtitle":"The bridge measure is built from harmonic functions, with explicit changes to drift and jump intensity.","key_machinery":"The central object is the $h$-transform of a jump-diffusion path measure, defined by $\\mathbf{P}^h=\\mathbf{1}_{\\{X_0\\in A_0\\}}r_0^{-1}(h(T,X_T)/h(0,X_0))\\mathbf{P}$ for a non-negative $h$ satisfying the mean-value property. If $h$ is harmonic, i.e. $(\\partial_t+L)h=0$, then $\\mathbf{P}^h$ solves the martingale problem for a transformed jump-diffusion operator $L^h$ with the explicit drift and jump measure above; equivalently it is a Girsanov transform with those coefficients. The approximation machinery approximates the Schrödinger-system function $g$ by $C_c^\\infty$ functions $g_k$ (via simple functions, Lusin's theorem, and mollification), and Assumption (A3) makes each $h_k=\\int g_k(y)P_{t,T}(x,dy)$ harmonic, so the corresponding $h$-transforms converge strongly to $\\hat{\\mathbf{P}}$.","core_discovery":"Under Assumptions (A1)/(A2) and (A3), the solution $\\hat{\\mathbf{P}}$ to the SBP for a jump-diffusion reference $\\mathbf{R}$ is the $h$-transform of $\\mathbf{P}$ by $h(t,x)=\\int g(y)P_{t,T}(x,dy)$, where $(f,g)$ solves the Schrödinger system. When $h$ is only measurable, the bridge is still recovered as $\\lim_k \\sup_{A\\in\\mathcal{F}}|\\mathbf{P}_{h_k}(A)-\\hat{\\mathbf{P}}(A)|=0$, with $h_k(t,x)=\\int g_k(y)P_{t,T}(x,dy)$ for smooth compactly supported approximations $g_k$ of $g$; each $h_k$ is harmonic, so each $\\mathbf{P}_{h_k}$ is a classical harmonic $h$-transform. In the smooth stochastic-flow case, $\\hat{\\mathbf{P}}$ is itself an $h$-transform by a harmonic $\\phi$, the marginal density of $\\hat{\\mathbf{P}}$ at time $t$ is $\\phi(t,x)\\hat{\\phi}(t,x)$, and $(\\phi,\\hat{\\phi})$ solves the backward-forward PIDE system. The generator of the bridge has drift $b^h=b+\\sigma\\sigma^T\\nabla\\log h+\\int_{|z|\\le 1}(h(x+\\gamma)-h(x))/h(x)\\,\\gamma\\,\\nu(dz)$ and jump measure $\\nu^h(dz)=(h(x+\\gamma)/h(x))\\nu(dz)$.","pith_inferences":["The explicit g_k construction suggests a numerical route to the bridge: approximate g from endpoint data and evaluate h_k by Monte Carlo transition kernels, without needing a smooth transition density, as long as (A3) is verified.","The tilt factor h(x+γ)/h(x) is the jump analogue of the diffusion drift pull toward high-h regions; one could test in simulation whether bridges of jump diffusions select paths whose jump destinations track the level sets of h.","For stable-like generators, the gap between finite-dimensional and strong convergence is likely not just technical: if (A3) fails, strong convergence would probably require heat-kernel or gradient estimates that the paper does not supply."],"forward_implications":["For any jump-diffusion reference satisfying (A3), the bridge can be approximated by harmonic h-transforms built from explicit smooth approximations of the Schrödinger-system function, so the diffusion-case approximation picture carries over.","The bridge dynamics are explicit: the drift gains the term σσᵀ∇log h plus a jump-induced correction, and the jump measure is tilted by h(x+γ)/h(x), so jumps into regions of higher h become more likely.","In the smooth stochastic-flow setting, the marginal density of the bridge factorizes as φφ̂ and the dynamic Schrödinger system becomes a pair of backward and forward PIDEs, extending the classical diffusion results.","For α-stable-like jump operators where (A3) is not known, h-transforms of mollified operators still converge to the bridge in finite-dimensional distributions, giving a weaker but usable approximation."],"supporting_citations":[{"why":"Originates the problem: find the most probable explanation for observed endpoint distributions under a Markov prior.","marker":"[57, 58]"},{"why":"Supplies the large-deviations formulation that identifies the Schrödinger bridge with KL-divergence minimization.","marker":"[26]"},{"why":"Supplies the product-form representation \\hat P = f(X0)g(XT)R and the Schrödinger system used throughout the paper.","marker":"[43]"},{"why":"Supplies the diffusion-case dynamic Schrödinger system that the jump-diffusion results extend.","marker":"[33]"},{"why":"Supplies the smooth stochastic-flow framework: adjoint operator, diffeomorphic jump maps, and transition-density regularity used in Section 5.","marker":"[41]"},{"why":"Supplies heat-kernel estimates and transition densities for non-symmetric jump operators used in the stable-like approximation.","marker":"[12]"}],"fun_headline_variants":["Jump-diffusion Schrödinger bridges via harmonic h-transforms","Solving Schrödinger bridges for jump diffusions","Jump-diffusion bridge measures from harmonic functions","Harmonic h-transforms solve jump-diffusion Schrödinger problem"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on Assumption (A3): the conditional expectation h(t,x)=E[g(X_T)|X_t=x] must be smooth enough in time and space and satisfy the backward equation for every smooth compactly supported terminal function g, and if that regularity fails the approximating functions are no longer harmonic and the strong-convergence proof stops working.","fun_headline_variants_meta":{"raw":{"variants":["Jump-diffusion Schrödinger bridges via harmonic h-transforms","Solving Schrödinger bridges for jump diffusions","Jump-diffusion bridge measures from harmonic functions","Harmonic h-transforms solve jump-diffusion Schrödinger problem"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2723,"prompt_tokens":1087,"completion_tokens":1636,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":703,"completion_tokens_details":{"reasoning_tokens":1571}},"tokens_in":703,"tokens_out":1636,"duration_ms":14085,"temperature":1.0,"reasoning_tokens":1571,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:55:51.308822+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the α-stable-like jump operator of Section 5.3 with a discontinuous Schrödinger-system function g and check whether sup_A |P_{h_k}(A)-\\hat{P}(A)| tends to 0; the paper proves only finite-dimensional convergence in that case, so exhibiting a counterexample to the vanishing total-variation distance would show the strong-convergence claim depends essentially on (A3).","supporting_citations":[{"cited_title":"Random fields and diffusion processes","cited_arxiv_id":null,"evidence_quote":"Supplies the large-deviations formulation that identifies the Schrödinger bridge with KL-divergence minimization."},{"cited_title":"A survey of the Schrödinger problem and some of its connections with optimal transport, 2013","cited_arxiv_id":null,"evidence_quote":"Supplies the product-form representation \\hat P = f(X0)g(XT)R and the Schrödinger system used throughout the paper."},{"cited_title":"The Markov processes of Schrödinger","cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion-case dynamic Schrödinger system that the jump-diffusion results extend."},{"cited_title":"Stochastic Flows and Jump-Diffusions","cited_arxiv_id":null,"evidence_quote":"Supplies the smooth stochastic-flow framework: adjoint operator, diffeomorphic jump maps, and transition-density regularity used in Section 5."},{"cited_title":"Heat kernels for non- symmetric diffusion operators with jumps","cited_arxiv_id":null,"evidence_quote":"Supplies heat-kernel estimates and transition densities for non-symmetric jump operators used in the stable-like approximation."}],"review_version":1}