{"id":"cef1c505-1de3-4e62-846a-03618ed53a66","arxiv_id":"2512.06839","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Jump diffusions admit a generator split into a symmetric dissipative part and an antisymmetric circulation part; the free energy dissipation rate equals minus a nonlocal Fisher information, with housekeeping heat as the gap to entropy production.","lead":"This paper derives a thermodynamic bookkeeping for jump diffusions — processes that mix Brownian motion and sudden random jumps — by splitting energy loss into entropy production and \"housekeeping heat\" needed to keep a non-equilibrium steady state. It also decomposes the process generator into a reversible, dissipative part and a circulation-only part, extending a known structural result from smooth diffusions to discontinuous trajectories.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prop 4.2's conclusion F_a'(ρ)=0 is false: the proof drops a nonzero symmetrization term, so L_a is not dissipation-free and Prop 4.3's dF/dt=-I[ρ] does not follow.","rationale":"The reader's weakest_assumption focuses on Assumption (E), a regularity/integrability condition. That is a legitimate concern but not the main defect. The paper's central decomposition result, as used thermodynamically, depends on Prop 4.2, and Prop 4.2's proof contains a concrete algebraic error: after the standard antisymmetrization of the nonlocal integral, a nonzero term proportional to (f_x-f_y)(log f_x-log f_y) is dropped. The error is visible even in the pure-jump, uniform-stationary three-state cycle, where all the paper's auxiliary identities (4.3)-(4.4) hold and yet F_a'≠0 for a nonuniform density. Therefore the claim that L_a generates a canonical conservative flow with zero free-energy dissipation is not merely unproven under heavy-tailed kernels; it is false in the well-behaved jump sector. Consequently Prop 4.3's headline relation dF/dt=-I[ρ] fails, and the physical interpretation of housekeeping heat as the difference between the symmetric and antisymmetric sectors is not supported. Theorem 4.1's formal decomposition L=L_s+L_a is itself fine, and parts of Section 3 are coherent, so the paper contains useful pieces; but the central advertised result is incorrect as stated, so the appropriate verdict is REJECT rather than CONDITIONAL.","tokens_in":20488,"tokens_out":64152,"duration_ms":495916,"concrete_test":"Evaluate the symmetrization step in Prop 4.2 on the jump-only 3-state cycle with uniform ρ_ss, antisymmetric k_a(1,2)=k_a(2,3)=k_a(3,1)=1/2 and reversed entries -1/2, and with f=(0.5,0.3,0.2). Compute βF_a' = Σ_{x,y} p_x k_a(x,y)(log f_y-log f_x), with p_x=f_x/3. The result is 0.0050, not 0. Separately compute the paper's symmetrized expression; it gives 0. The discrepancy isolates the missing (f_x-f_y)(log f_x-log f_y) term. Equivalently, in Example 5.1 with α=1.5, β=1, and initial Gaussian ρ_0, numerically evaluate βF_a' = ∫∫ f_0(x)(ρ_ss(x)-ρ_ss(y))k(x,y)[1+½(log f_0(y)-log f_0(x))]dy dx using ρ_ss from (5.2); it will be nonzero.","verdict_should_be":"REJECT","load_bearing_attack":"The central thermodynamic claim rests on Prop 4.2 (F_a'(ρ)=0) and Prop 4.3 (dF/dt=-I[ρ]). Prop 4.2's proof is algebraically invalid. In the notation of §4.1, βF_a' = ∫ ρ(t,x)v_a·∇log f dx + ∫∫ ρ(t,x)k_a(x,y)(log f(y)-log f(x)) dy dx, with f=ρ/ρ_ss. Using v_aρ_ss = j_loc^ss and stationarity ∇·j_loc^ss = -∫j_nl^ss, this reduces to βF_a' = ∫∫ f(x) j_nl^ss(x,y)[1 + ½(log f(y)-log f(x))] dy dx, which is generically nonzero. The symmetrization in Prop 4.2 replaces part of the integrand with a 1/4 ρ_ss k_a (f_x+f_y)(log f_y-log f_x) term and discards the 1/2 ρ_ss k_a (f_x-f_y)(log f_y-log f_x) term; (4.4) kills the first but not the second. A minimal jump-only counterexample (3-state cycle with uniform ρ_ss and antisymmetric k_a with zero row sums, f=(0.5,0.3,0.2)) gives βF_a'=0.005≠0, even though both (4.3) and (4.4) hold. Thus L_a is not dissipation-free for generic non-equilibrium densities, and Prop 4.3's identification of total free-energy decay with the symmetric Dirichlet form is unsupported. This is independent of Assumption (E) and of integrability tails.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the stochastic thermodynamics of jump diffusions on R^n combining Brownian and Poisson noise. It derives a free-energy dissipation formula dF/dt = Q_hk - β^{-1}e_p, proves (or claims) nonnegativity of Q_hk and e_p, and then decomposes the generator L = L_s + L_a in L^2(R^n, ρ_ss dx), with L_s self-adjoint and L_a antisymmetric. The central claim is that the antisymmetric part generates a 'canonical conservative dynamics' with zero free-energy dissipation, F_a'(ρ)=0, so that the total dissipation is carried by the symmetric part and equals minus a nonlocal Fisher information I[ρ]: dF/dt = -I[ρ]. Two numerical examples illustrate the intended decomposition.","tokens_in":20911,"tokens_out":30380,"duration_ms":234524,"significance":"The paper contains a useful algebraic decomposition of the generator into self-adjoint and antisymmetric parts (Thm. 4.1) and proposes a nonlocal Fisher information via the Dirichlet form of the symmetric part. If the dissipation-free property of L_a were true, it would provide a clean geometric picture extending Helmholtz-Hodge decompositions to jump diffusions. The numerical examples illustrate the intended phenomenology. However, the central result is not correct as stated, and the supporting calculations contain systematic factor errors; the claimed physical interpretation therefore does not follow.","major_comments":[{"comment":"The claim F_a'(ρ)=0 is false. In the proof, the identity I2 = (1/(2β))∫∫ρ k_a (log f(y)-log f(x)) is missing a factor 1/β; the correct identity from the preceding line is I2 = (1/β)∫∫ρ k_a(φ(y)-φ(x)). Consequently the two symmetrized expressions are each (1/2)I2 with opposite signs, not I2 itself, so they cannot imply I2=0. Concrete counterexample within the hypotheses: on a three-point space with uniform ρ_ss and k(1,2)=k(2,3)=k(3,1)=2, k(1,3)=k(2,1)=k(3,2)=1, take f=(0.5,0.3,0.2). Then (4.3) and (4.4) hold but βF_a'=0.015≠0. Hence L_a is not dissipation-free and Prop. 4.3 is unsupported.","section":"§4.1, Prop. 4.2"},{"comment":"The step replacing (1/(2β))∫∫ j_nl log C, with C=kρ_ss/(k'ρ_ss'), by (1/(4β))∫∫ j_nl [log C + log(1/C)] = 0 is algebraically invalid: the two terms in the bracket cancel pointwise for the same (x,y), yielding zero, while the original integral is generically nonzero. A correct proof must use, e.g., the known relative-entropy contraction or a proper entropy-production inequality. As written, the proof of dF/dt≤0 does not follow.","section":"§3, proof of Thm. 3.1, Eqs. (3.11)-(3.12)"},{"comment":"Detailed balance is defined by j_loc(t,x)=0 and j_nl(t,x,y)=0 for all t>0 and all x,y. This is inconsistent with relaxation: a reversible jump diffusion starting from a non-stationary density has nonzero currents at finite times. The definition should be stated for the stationary currents (j_loc^ss=0, j_nl^ss=0), or equivalently in terms of the invariant measure/reversibility, matching Prop. 2.2.","section":"§2, Def. 2.1"}],"minor_comments":[{"comment":"The stationary Fokker-Planck equation as written is missing the factor β^{-1} multiplying A(x)∇ρ_ss. The correct equation is ∇·(bρ_ss - β^{-1}A∇ρ_ss) + ∫(ρ_ss(y)k(y,x)-ρ_ss(x)k(x,y))dy = 0. This appears to be a typo, since the prior line uses the β^{-1} factor.","section":"Eq. (3.10)"},{"comment":"Since F = β^{-1}H, the free-energy dissipation should read dF/dt = β^{-1}⟨f, L_s log f⟩_{ss} = -β^{-1}I, not -I. The factor β^{-1} is dropped in the statement.","section":"Prop. 4.3"},{"comment":"The local contribution I1 is written with the sign ∇logρ_ss - ∇logρ, whereas the adjoint calculation gives the opposite sign. Both signs yield zero by (4.3), so this is harmless but confusing.","section":"Prop. 4.2 proof, I1"},{"comment":"Assumption (E) is quite strong: it requires bounded ∇logρ and k̄∈L∞, excluding many heavy-tailed asymmetric kernels for which the entropy production integrand may be infinite. The paper should state this limitation explicitly in the introduction or abstract.","section":"Assumption (E)"},{"comment":"There are several typos: 'detial', 'serveed', 'extendion', 'bablance', and inconsistent capitalization of 'Lévy'. Also, Figure 1 and Figure 2 have no descriptive captions in the text.","section":"General"}],"recommendation":"reject","confidential_remarks":"The false Prop. 4.2 is decisive. I would not encourage resubmission without a genuine re-derivation of the dissipation decomposition; the current claims are not salvageable by minor edits. The L^2 decomposition and the nonlocal Fisher information may be reusable in a different statement, but the present manuscript's main theorem is invalid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The punchline: the main structural result is false. Proposition 4.2 claims the antisymmetric part of the generator is dissipation-free (F_a'=0). The proof's symmetrization drops a term. The correct calculation gives βF_a' = ∫∫ f(x) j_nl^ss(x,y)[1 + ½(log f(y)-log f(x))] dy dx (plus a local term), which is generically nonzero. I checked a minimal 3-state cycle with uniform stationary density and a cyclic antisymmetric kernel satisfying (4.3)-(4.4); with f=(0.5,0.3,0.2) the expression is 0.01 (up to normalization), not zero. So L_a does dissipate free energy, and Proposition 4.3's identification of dF/dt with -I[ρ] does not follow.\n\nWhat's actually new and worthwhile: Section 3 derives explicit formulas for housekeeping heat and entropy production for general jump diffusions—that's a genuine extension of the diffusion/Markov-jump results. The nonlocal Fisher information with logarithmic mean is a nice object, even if it doesn't play the role the paper claims. The numerical examples are illustrative, not decisive.\n\nOther soft spots: Definition 2.1 defines detailed balance by vanishing currents at all times, which is stronger than the usual reversibility condition and inconsistent with transient relaxation. Equation (3.10) drops a β^{-1} from the stationary Fokker-Planck equation, though the conclusion there survives. There are several typos and no code. The regularity assumption (E) is strong; heavy-tailed asymmetric kernels are outside the stated scope.\n\nThe paper deserves a serious referee because the Section 3 results are substantial and the Section 4 error is subtle—it would take a careful reader to catch it. But as it stands, the advertised 'canonical conservative flow' interpretation and the decomposition of free-energy dissipation are unsupported. The authors need to either fix the proof (unlikely, since the claim is genuinely false) or substantially reframe the contribution around Section 3.\n\nI'd bring it to a reading group: the error is instructive, and the nonlocal Fisher information idea might be salvageable. But I wouldn't cite it in its current form.","headline":"Section 3 is a real extension, but the generator decomposition at the center of Section 4 doesn't hold—Prop 4.2's proof is wrong.","tokens_in":21380,"tokens_out":14467,"would_cite":false,"duration_ms":97079,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J75","60J60","82C31","35Q84"],"pacs":["05.40.-a","05.70.Ln"],"model":"deepseek-v4-flash","headline":"This paper proves that for jump diffusions without detailed balance, the free energy dissipation is carried entirely by the reversible part of the generator and equals a nonlocal Fisher information, while the antisymmetric part generates di","keywords":["free energy dissipation","jump diffusions","Lévy processes","Fisher information","entropy production","housekeeping heat","detailed balance","generator decomposition"],"falsifier":"Construct a well-posed jump diffusion with an asymmetric heavy-tailed kernel, such as k(x,y)=|x−y|^{-(1+α)}e^{y−x}, and solve for its stationary density. If the nonlocal part of e_p(t) diverges because k log(k(x,y)/k(y,x)) is unbounded, while the process itself has smooth finite-density marginals, then the dissipation identity dF/dt = Q_hk − β^{-1} e_p fails exactly when Assumption (E) is violated, marking the boundary of the theorem.","tokens_in":20366,"feed_emoji":"🔄","tokens_out":6894,"duration_ms":68428,"temperature":0.7,"pith_summary":"Jump diffusions—stochastic processes that combine smooth Brownian motion with discontinuous Poisson jumps—are used to model everything from intracellular cargo transport and animal foraging to financial returns and generative models. This paper proves that, whenever such a process has a smooth stationary density, its generator splits into two pieces in the Hilbert space weighted by that density: a symmetric piece that is a reversible jump diffusion, and an antisymmetric piece that circulates probability along closed loops without changing the stationary state. The free energy, defined as the relative entropy of the running distribution with respect to the stationary measure, always decreases; its entire dissipation rate equals the nonlocal Fisher information of the symmetric piece, while the antisymmetric circulation contributes zero dissipation. The same calculation yields a decomposition of the dissipation rate into housekeeping heat and entropy production, each nonnegative, and a Clausius inequality. This gives the nonlocal, jump-driven analogue of the familiar Helmholtz–Hodge decomposition of drift in Langevin dynamics: the irreversible part relaxes, the rotational part spins forever at no free-energy cost.","feed_headline":"Jump diffusions split into a dissipative half and a circulation half","feed_subtitle":"The reversible half sets the decay rate via a nonlocal Fisher information; the antisymmetric half spins without cost.","key_machinery":"The central object is the ρ_ss-weighted generator decomposition. In the Hilbert space L^2(R^n, ρ_ss dx), take the adjoint L^† of the generator with respect to the stationary density. Then L_s = (L+L^†)/2 is the symmetric, reversible part; its Dirichlet form E(f, log f) defines a nonlocal Fisher information I[ρ] that carries the full free-energy decay. L_a = (L−L^†)/2 is the antisymmetric part, with drift v_a = b − β^{-1}A∇logρ_ss and jump kernel k_a(x,y) = (1/2)(k(x,y)/ρ_ss(y) − k(y,x)/ρ_ss(x))ρ_ss(y). This kernel integrates to zero and the drift is divergence-free against ρ_ss, so L_a preserves the stationary density and generates pure circulation along its level sets.","core_discovery":"The central discovery is an operator-theoretic split of the generator of a general jump diffusion into L = L_s + L_a, taken with respect to the inner product weighted by the stationary density ρ_ss. L_s, defined as (L+L^†)/2, is self-adjoint and generates a time-reversible jump diffusion, so it satisfies detailed balance by construction; its Dirichlet form defines a nonlocal Fisher information I[ρ], and the free energy dissipation is exactly dF/dt = −I[ρ(t)]. L_a, defined as (L−L^†)/2, generates the canonical conservative dynamics: its drift v_a = b − β^{-1}A∇logρ_ss and jump kernel k_a are constructed so that (L_a)^*ρ_ss = 0 and F_a'(ρ) vanishes identically. The antisymmetric part produces","pith_inferences":["A practical design consequence, not drawn by the paper: adding an antisymmetric (curl-like) driving force to a jump diffusion can be expected to change the stationary circulation without changing the relaxation rate, since the decay rate is fixed solely by L_s; this could be tested by comparing equilibration times of two jump diffusions with the same L_s but different L_a.","The logarithmic mean appearing in the nonlocal Fisher information suggests a bridge between discrete-state Markov chains and continuum diffusions; in the small-jump limit I_nl should converge to the classical Fisher information, so the same functional family may interpolate across levels of coarse-graining.","The canonical conservative flow generated by L_a could be made explicit through a continuous-space cycle decomposition, analogous to the loop decomposition of master equations; expressing the antisymmetric nonlocal current as a superposition of closed loops x → y → z → x would give a graphic way to measure how far a jump diffusion is from detailed balance."],"forward_implications":["The free energy of any such jump diffusion is a Lyapunov function: F(ρ(t)) is nonincreasing, and the Clausius inequality θ dS/dt − dQ/dt = θ e_p(t) ≥ 0 holds.","Housekeeping heat Q_hk(t) equals the mechanical power of external driving, both local and nonlocal, and vanishes exactly when detailed balance holds.","Free energy dissipates only through L_s: dF/dt = −I[ρ(t)], where I is the nonlocal Fisher information built from the symmetric Dirichlet form and decomposes into a diffusion part and a jump part involving the logarithmic mean.","If the symmetric generator satisfies a modified logarithmic Sobolev inequality, F(t) ≤ F(0) e^{−t/(β C_I)}, so the relaxation rate is controlled by the spectral gap of L_s, not by the circulation L_a.","In a nonequilibrium steady state, dF/dt → 0 forces the stationary entropy production rate to equal β times the stationary housekeeping heat, so the steady state sustains positive entropy production without further free energy loss."],"fun_headline_variants":["Jump diffusions split into dissipative and circulation parts","Free-energy decay set by nonlocal Fisher information","Antisymmetric generator spins without dissipation","Reversible half dissipates, antisymmetric half circulates","Generator decomposition reveals cost-free circulation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole construction leans on the assumption that the process has a smooth positive stationary density and that the jump kernel's log-ratio k(x,y)log(k(x,y)/k(y,x)) is bounded along the time-marginal distributions; if either fails, the integral defining entropy production can be infinite and the L^2 decomposition is not established.","fun_headline_variants_meta":{"raw":{"variants":["Jump diffusions split into dissipative and circulation parts","Free-energy decay set by nonlocal Fisher information","Antisymmetric generator spins without dissipation","Reversible half dissipates, antisymmetric half circulates","Generator decomposition reveals cost-free circulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1096,"prompt_tokens":693,"completion_tokens":403,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":344}},"tokens_in":437,"tokens_out":403,"duration_ms":4593,"temperature":1.0,"reasoning_tokens":344,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T18:03:21.303093+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a well-posed jump diffusion with an asymmetric heavy-tailed kernel, such as k(x,y)=|x−y|^{-(1+α)}e^{y−x}, and solve for its stationary density. If the nonlocal part of e_p(t) diverges because k log(k(x,y)/k(y,x)) is unbounded, while the process itself has smooth finite-density marginals, then the dissipation identity dF/dt = Q_hk − β^{-1} e_p fails exactly when Assumption (E) is violated, marking the boundary of the theorem.","supporting_citations":[],"review_version":1}