{"id":"4c1e47dc-2b38-44c5-93fb-90f493da1d8b","arxiv_id":"2608.11585","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A single adjoint framework identifies when the same physical hardware that computes forward can also compute exact gradients, with distinct conditions for linear and nonlinear systems.","lead":"Physical devices can train themselves by computing their own gradients, but only under specific conditions. The paper gives a unified theory: linear systems need reciprocity, nonlinear systems need time-reversal symmetry and nudging, and a broader class of non-reciprocal systems can still work via a symmetry trick.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Nonlinear trajectory 'formally exact' gradients (Thm 6.1, Eq. 31) require ε→0 while the (w_ε−w)/ε extraction amplifies re-injection/measurement error as η/ε; exactness thus holds only in a jointly noiseless, infinitesimal limit the paper does not quantify.","rationale":"Good-faith reading: the paper is a theory contribution with explicitly conditional theorems; its central claim is that stated sufficient conditions permit exact same-hardware adjoint computation, and it carefully recovers existing algorithms as instances. I re-derived the key algebraic steps and found them internally consistent: the forward-time adjoint transformation (Appendix A.10, Eqs. 124–127), the linear reciprocal identities (Eqs. 26–27; Secs. 7.1, 9.1), the damping-obstruction analysis (Appendix B), the nonlinear undamped matching of the nudged response (Eq. 29 with D=0 and F_u^T=F_u) to the adjoint equation (4), the intertwining construction with S K^T S^{-1}=K (Theorem 10.1, Eq. 82), and the Hatano–Nelson twisted-reciprocal example (Sec. 10.1). I could not find a mathematical contradiction, a circular step, or an incorrect recovery claim among the cases I checked (EP, FFM, scattering-matrix rules, free-space optics). The single most load-bearing concern is the operational claim hidden in the nonlinear trajectory construction: exactness requires the limits ε→0 (to kill the finite-nudge bias) and η→0 (to kill re-injection/measurement error) simultaneously, but the extraction (31) amplifies the latter as η/ε. The paper flags the O(ε) bias and offers central differencing, but never analyzes the error floor set by re-injection precision, even though Sec. 6.2 explicitly requires initializing the nudged run from measured (u(T), u̇(T)). This matches and sharpens the prior review's weakest-assumption identification, and I agree with that identification. It does not, however, overturn the reader's ACCEPT: the theorems are true under the stated idealizations, the paper already lists infinitesimal nudging and the TRM as requirements in Table 1 and Sec. 6, and the finite-amplitude linear and twisted-reciprocal constructions — which carry the other half of the central claim and the FFM extension — do not suffer this mechanism. The right remedy is a stated accuracy floor (or a clarifying sentence that 'formally exact' means the mathematical limit), not a rejection. Hence UNCHANGED.","tokens_in":48717,"tokens_out":39298,"duration_ms":373816,"concrete_test":"Extend the paper's own Taylor-test protocol (Appendix I) to the full nonlinear algorithm of Eqs. (30)–(31) on a simulated Hamiltonian lattice (8–16 FPU/Klein–Gordon oscillators, quadratic terminal loss, D=0, reciprocal F_u), with the reference gradient g from a digital adjoint solve. Inject controlled re-injection errors η ∈ {10^{-4}, 10^{-6}, 10^{-8}} into u(T) and u̇(T), sweep ε from 10^{-1} to 10^{-6}, and record ‖ĝ_ε − g‖/‖g‖. If the error follows η/ε + O(ε) with an accuracy floor ~√η at ε* ~ √η rather than decreasing monotonically to machine precision, then exactness in Eq. (31) holds only in the noiseless limit and the paper should state the achievable gradient-error floor as a function of measurement/re-injection precision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is conditional: exact same-hardware adjoints are obtained under stated sufficient conditions, and I verified substantial portions of the algebra (linear reciprocal case, Eqs. 26–27; nonlinear undamped perturbation match, Eqs. 29–31 vs. Eq. 4; intertwining Theorem 10.1, Eq. 82; Hatano–Nelson example, Eqs. 85–87). The load-bearing weakness is the operational status of the nonlinear trajectory construction, which the prior review also identified. In Sec. 6.2, a(t) = lim_{ε→0}(w_ε(t) − w(t))/ε, with w the recorded reversed trajectory and w_ε a fresh physical run initialized from the re-injected full state (Eq. 30). The paper acknowledges only the O(ε) forward-difference bias (Eq. 31, central-difference fix Eq. 32). It does not address that w_ε must reproduce the recorded w to within the re-injection/measurement error η in the measured (u(T), −u̇(T)); that error propagates along the reversed trajectory (without decay under conservative dynamics, and with growth under chaotic ones) and enters the extracted field as O(η/ε), diverging as ε→0. Bias O(ε) and noise O(η/ε) trade off at ε* ~ √η, imposing a gradient-accuracy floor ~√η, so 'formally exact' is unreachable on hardware with finite measurement and actuation precision. The same η/ε structure afflicts the PT-TRM Schrödinger variant (Eq. 58) and the EP stationary limit (Eq. 76), making it the generic price of contrastive physical gradients rather than a corner case. The theorems remain mathematically true under their stated idealizations; what is unsupported is the suggestion that the nonlinear route delivers formally exact gradients outside a jointly noiseless, infinitesimal limit. The finite-amplitude linear (Secs. 6.1, 7.1, 9.1) and twisted-reciprocal (Sec. 10) constructions are immune to this mechanism. Because the paper itself lists infinitesimal nudging and the TRM as requirements (Table 1, Sec. 6), this sharpens an acknowledged limitation rather than contradicting the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a unifying theory, based on the adjoint method, for when the adjoint field required for exact gradient computation can be generated on the same physical hardware that ran the forward computation. It derives sufficient conditions for linear and nonlinear second-order, first-order, Schrödinger-type, and stationary systems; recovers Equilibrium Propagation, Hamiltonian echo backpropagation, fully forward mode training, and photonic in situ backpropagation; and introduces an intertwining condition that extends exact same-hardware adjoints to a class of non-reciprocal systems illustrated by a Hatano–Nelson chain. The appendices provide self-contained derivations of the adjoint equations and gradient formulas for all four system classes, including the damping obstruction for nonlinear trajectories and the V-reciprocal generalization.","tokens_in":49076,"tokens_out":8188,"duration_ms":94599,"significance":"If the results hold as stated, this is a valuable unification: it places several existing exact physical-learning algorithms in a common adjoint framework, clarifies the differing roles of linearity versus nonlinearity, and provides constructive sufficient conditions for designing new on-device gradient schemes. The derivations are self-contained; Theorems 6.1, 7.1, 8.1, 9.1, and 10.1 are explicit; and the Hatano–Nelson example gives a concrete, falsifiable construction. The paper also correctly distinguishes exact from approximate prior algorithms. The main limitation is operational: the nonlinear trajectory construction is exact only in an idealized noiseless infinitesimal limit, and the paper does not quantify how finite measurement and actuation precision degrade the extracted gradient.","major_comments":[{"comment":"The claim that the nonlinear adjoint is obtained exactly from a(t) = lim_{ε→0}(w_ε(t)-w(t))/ε assumes that the full final state (u(T), -u̇(T)) can be measured and re-injected without error and that w_ε can be recorded without error. With measurement/re-injection error η, the extracted field carries error O(η/ε), while the truncation bias is O(ε); balancing the two gives ε* ~ √η and a gradient accuracy floor ~ √η. Thus 'formally exact' is accurate only in the jointly noiseless, infinitesimal limit. The same η/ε structure affects the PT-TRM Schrödinger construction (Eq. (58)) and the Equilibrium Propagation stationary limit (Eq. (76)). The authors should add a quantitative error-propagation statement, or explicitly qualify every 'formally exact same-hardware' claim as holding in the noiseless mathematical model, rather than leaving the idealization implicit.","section":"Sec. 6.2, Eq. (31); also Sec. 8.2, Eq. (58) and Sec. 9.2, Eq. (76)"}],"minor_comments":[{"comment":"The initial conditions w_ε(0) = u(T) - ε M^{-1} Pψ_u̇ and ẇ_ε(0) = -u̇(T) - ε M^{-1} Pψ_u are written with what appear to be unprojected loss derivatives; consider making the notation ψ_u = ∂ψ/∂u and ψ_u̇ = ∂ψ/∂u̇ explicit so the connection to the adjoint initial conditions (6)-(7) is immediate.","section":"Sec. 6.2, Eq. (30)"},{"comment":"The sentence 'The construction requires time reversal symmetry, so that complex conjugating c(T) re-generates the adjoint state evolution' is valid only when K is real; the preceding paragraph has only K^T = K. The authors do state later that time reversal symmetry makes K real, but the logic would be clearer if this assumption were introduced before the intensity-reconstruction protocol.","section":"Sec. 8.1, 'Holographic vs. intensity measurements'"},{"comment":"The table entries for damped nonlinear trajectory systems say 'None' and mark the case obstructed; this is correct for the same-device setting, but the mass-proportional damping construction of Appendix C (which requires flipping damping to gain and hence a different hardware setting) is summarized only in the text. A parenthetical pointer to Appendix C in the table would prevent the impression that all damped nonlinear trajectory gradients are unobtainable in every experimental configuration.","section":"Table 1, nonlinear damped trajectory rows"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically strong and the mathematical core is sound; the requested revision is driven by the gap between 'formally exact' and physically implementable under finite precision. The authors should calibrate the abstract and conclusions so that the infinitesimal, noiseless nature of the nonlinear constructions is explicit, and add a short error-propagation discussion. This is a focused revision rather than a request for new results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know upfront. This is the first credible unification of on-device physical backpropagation: Equilibrium Propagation, Hamiltonian echo, fully forward mode training, and the integrated-photonic in-situ methods all show up as instances of one adjoint construction, and the linear/nonlinear split the paper draws is real. Linear reciprocal systems need only a single finite-amplitude adjoint run, with damping and gain allowed; nonlinear trajectory systems need a time-reversal mirror, reciprocal linearized dynamics, and infinitesimal nudging. Second, the linear half of the paper is solid, while the nonlinear trajectory construction is mathematically right as stated but only operationally 'exact' in a noiseless infinitesimal limit the paper does not quantify.\n\nWhat is actually new: I would credit three results. Exact trajectory gradients in linear reciprocal overdamped systems (Sec 7.1). PT-symmetric nonlinear Schrödinger physical adjoints (Sec 8.2). And the intertwined-reciprocity construction (Sec 10), which extends exact same-device gradients to non-reciprocal systems, Hatano–Nelson chain included—that is a real piece of theory. The appendices are complete and self-contained; I checked the linear reciprocal algebra, the undamped nonlinear perturbation match, the intertwining theorem, and the Hatano–Nelson example, and they hold up. The theorems are conditional, carefully stated, and the derivations do not assume the target results. No circularity concerns.\n\nThe soft spot, in proportion: the nonlinear trajectory route (Thm 6.1, Eq. 31) extracts the adjoint as (w_ε−w)/ε. The paper addresses the O(ε) forward-difference bias with central differences, but the same division amplifies re-injection and measurement error η from the recorded final state as η/ε. Bias and noise trade off at ε*≈√η, giving a gradient-accuracy floor ∼√η. So 'formally exact' is not reachable on hardware with finite measurement and actuation precision. The same η/ε structure hits the PT-TRM Schrödinger variant and the EP stationary limit, so it is a generic property of contrastive physical gradients, not a corner case. This sharpens a limitation the paper already admits—Table 1 and Sec 6 list infinitesimal nudging and the TRM as requirements—so it is not a contradiction, but the abstract's 'formally exact' claim overpromises for the nonlinear case. The linear and twisted-reciprocal constructions are immune to this mechanism. The conclusions also honestly leave gradient degradation under model error open, which is fair.\n\nWho benefits: anyone designing physical neural networks or in-situ training schemes. It deserves serious referee time. My recommendation: send to review, accept with revision, and ask the authors to state the η/ε trade-off and soften the exactness claim for the nonlinear trajectory route.","headline":"A genuine unification of on-device physical backprop with a clean linear/nonlinear split; the nonlinear trajectory 'exactness' needs a noise-aware qualifier before publication.","tokens_in":49734,"tokens_out":4039,"would_cite":true,"duration_ms":38378,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49K15","68T05","81Q12","35Q41"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that physical hardware can compute the exact gradient of its own cost on-device: reciprocity for linear systems, a time-reversal mirror for nonlinear trajectories, an intertwining condition for non-reciprocal ones.","keywords":["physical backpropagation","adjoint method","on-device gradient computation","reciprocity","time-reversal mirror","equilibrium propagation","non-Hermitian systems","physical neural networks"],"falsifier":"On a damped resistor–capacitor network with unequal capacitances, run the forward trajectory, then run the single finite-amplitude adjoint experiment proposed here and compare its gradient with central-difference gradients of the measured cost; the theory predicts exact agreement, and deliberately adding a non-reciprocal element such as a gyrator should make the adjoint-run gradient deviate exactly as the intertwining condition predicts. Repeating the comparison on a one-way-gain Hatano–Nelson chain with mirror-symmetric on-site potentials would settle the twisted-reciprocal extension.","tokens_in":48497,"feed_emoji":"⚡","tokens_out":13571,"duration_ms":139466,"temperature":0.7,"pith_summary":"Physical computers—optical, mechanical, or electrical networks that compute through their own dynamics—are hard to train because their exact gradients are usually computed in a digital model that never quite matches the real device. This paper asks when the device itself can produce the gradient it needs, and answers with a unified theory built on the adjoint method, the standard technique of computing a cost gradient from two solves. The answer splits by regime: a linear system needs only reciprocity (its mass, damping, and stiffness matrices symmetric, in the Euclidean or Onsager sense), and then one extra finite-amplitude run of the same hardware—with sources and initial conditions swapped and time dependences reversed—yields the exact gradient, damping and gain included. A nonlinear trajectory system needs more: zero damping, a time-reversal mirror that regenerates the reversed trajectory from the final state, and a symmetric linearized internal force, with the gradient extracted from the infinitesimal response of a nudged reversed trajectory. The same framework recovers Equilibrium Propagation, Hamiltonian echo backpropagation, fully forward mode training, and photonic in-situ backpropagation as special cases, and a generalized intertwining condition extends exact on-device gradients to non-reciprocal systems.","feed_headline":"Physical devices can now compute their own exact gradients","feed_subtitle":"Reciprocity lets linear hardware train in one extra run; nonlinear trajectories need a time-reversal mirror.","key_machinery":"The central object is the forward-time adjoint field $a(t)=b(T-t)$: the Lagrange multiplier of the constrained optimization, re-indexed so it runs forward in time. It satisfies the adjoint equation, which involves transposed and linearized operators, and the gradient is built from overlap integrals between $a$ and parameter sensitivities of the forward operators. The load-bearing identity is that the adjoint propagator coincides with the forward propagator on the same hardware when the transposed operators are conjugate to the forward ones under a constant invertible matrix $S$—$S K^T S^{-1}=K$ (with $S=I$ for reciprocity, $S=V$ for Onsager reciprocity, and orthogonal $S$ for twisted reciprocity). For nonlinear trajectories the mechanism is the time-reversal mirror: initializing with the measured final state and flipped velocity regenerates the reversed trajectory $w(t)=u(T-t)$, and an infinitesimal nudge $\\epsilon P\\theta_u$ produces a response $\\delta w/\\epsilon$ that obeys the adjoint equation when $D=0$ and $F_u^T=F_u$ along the trajectory.","core_discovery":"The paper's central claim is that the adjoint field needed for formally exact gradients can be produced by the same hardware that ran the forward computation, under sufficient conditions that depend on the regime. For linear systems, reciprocity—$M^T=M$, $D^T=D$, $K^T=K$ at every instant, or the Onsager version $V A^T V^{-1}=A$—makes the adjoint equation identical to the forward equation, so the adjoint is obtained in a single finite-amplitude experiment with the loss derivatives as sources and initial conditions and any explicit time dependence replayed in reverse; damping and gain are admissible. For nonlinear trajectory systems the conditions are the existence of a time-reversal mirror (which regenerates $w(t)=u(T-t)$ from the measured final state and flipped velocity), an undamped dynamics $D=0$, and reciprocal linearization $F_u^T=F_u$ along the trajectory; the adjoint is then the infinitesimal limit $\\lim_{\\epsilon\\to 0}(w_\\epsilon-w)/\\epsilon$ of a nudged reversed trajectory, so multiple experiments and small nudges are required. In stationary problems the nonlinear construction reduces to the difference between free and nudged equilibria—Equilibrium Propagation—provided the fixed point is stable and its tangent operators self-adjoint. Finally, reciprocity is shown to be only the simplest case of an intertwining condition $S K^T S^{-1}=K$ for a constant invertible $S$, which yields exact physical gradients even in non-Hermitian, non-reciprocal systems, exemplified by a Hatano–Nelson chain with mirror-symmetric on-site potentials. The conditions are sufficient, not necessary, and the listed algorithms all follow as instances of the one adjoint construction.","pith_inferences":["The sufficiency, not necessity, of the conditions invites stronger statements: one can likely relax the intertwiner to time-dependent or parameter-dependent $S$ with closed-form updates, buying broader trainability at the price of dynamic range and recomputation.","The paper's separation of 'knowing the model' from 'knowing the parameters' predicts a concrete robustness effect: in devices whose parameters enter linearly, gradient estimates from measured forward and adjoint fields should be immune to drift and aging without recalibration—a testable laboratory prediction.","The appendix's mass-proportional damping construction, with scalar exponential reweighting, suggests a practical bridge for weakly damped nonlinear systems: even when the strict $D=0$ theorem fails, an approximate same-device gradient may be recovered and its error bounded by the size of the damping term.","The twisted-reciprocity condition is a trainability version of known non-Hermitian symmetry classifications; viewed that way, any hardware symmetry (rotation, inversion, exchange) with the correct parity signature under transposition confers exact on-device gradients, which could be used to screen candidate platforms before building them."],"forward_implications":["Linear reciprocal platforms with damping or gain can deliver exact trajectory gradients with one extra finite-amplitude run, so energy-dissipating hardware is trainable without a digital twin.","Nonlinear trajectory training on hardware is only exact in undamped, time-reversal-symmetric systems; damped nonlinear trajectories cannot produce exact gradients on the same device, but their steady states remain trainable by Equilibrium Propagation.","Existing exact algorithms—Equilibrium Propagation, Hamiltonian echo backpropagation, fully forward mode training, photonic in-situ backpropagation—are all special cases of one adjoint construction, giving them a common language for extension.","Symmetry-odd non-reciprocity (one-way gain on a mirror-symmetric structure) is trainable exactly, so non-Hermitian platforms like the Hatano–Nelson chain become usable for physical learning.","Encoding inputs in the parameters of a linear, damped, reciprocal device (structural nonlinearity) combines nonlinear computation with finite-amplitude exact gradients, a route the paper singles out as particularly attractive."],"supporting_citations":[{"why":"Supplies the adjoint method from waveform inversion that the whole construction is built on.","marker":"[28]"},{"why":"Early trainable hardware using error backpropagation through reciprocal physical media, the direct precedent of the linear reciprocal construction.","marker":"[20]"},{"why":"Equilibrium Propagation, recovered as the stationary nonlinear case where free and nudged equilibria are contrasted.","marker":"[30]"},{"why":"Hamiltonian echo backpropagation, recovered as the undamped nonlinear trajectory case with a time-reversal mirror.","marker":"[26]"},{"why":"Fully forward mode training, recovered as reciprocal hardware with a spatial symmetry relocating the adjoint injection to the input.","marker":"[32]"},{"why":"In-situ backpropagation gradients for photonic networks, the Helmholtz-type linear stationary instance of the construction.","marker":"[33]"},{"why":"Experimental realization of in-situ photonic backpropagation, confirming the stationary Helmholtz case on real hardware.","marker":"[23]"},{"why":"Scattering-matrix learning rules for wave systems, shown equivalent to the adjoint identity read at single ports.","marker":"[37]"},{"why":"The Hatano–Nelson chain, the example system for the twisted-reciprocal non-reciprocal extension.","marker":"[39]"}],"fun_headline_variants":["Self-gradients: hardware computes its own backprop","Reciprocity unlocks exact on-chip gradients","Time-reversal mirror enables nonlinear self-training","Unifying backprop for all physical learning","Adjoint method brings exact gradients to hardware"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction depends on the hardware being reversible in a specific sense: linear devices must be reciprocal, and nonlinear trajectory devices must be undamped and able to measure and re-inject their final state without error.","fun_headline_variants_meta":{"raw":{"variants":["Self-gradients: hardware computes its own backprop","Reciprocity unlocks exact on-chip gradients","Time-reversal mirror enables nonlinear self-training","Unifying backprop for all physical learning","Adjoint method brings exact gradients to hardware"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000452,"raw_usage":{"total_tokens":2392,"prompt_tokens":1182,"completion_tokens":1210,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":798,"completion_tokens_details":{"reasoning_tokens":1150}},"tokens_in":798,"tokens_out":1210,"duration_ms":11347,"temperature":1.0,"reasoning_tokens":1150,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:34:20.616813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a damped resistor–capacitor network with unequal capacitances, run the forward trajectory, then run the single finite-amplitude adjoint experiment proposed here and compare its gradient with central-difference gradients of the measured cost; the theory predicts exact agreement, and deliberately adding a non-reciprocal element such as a gyrator should make the adjoint-run gradient deviate exactly as the intertwining condition predicts. Repeating the comparison on a one-way-gain Hatano–Nelson chain with mirror-symmetric on-site potentials would settle the twisted-reciprocal extension.","supporting_citations":[{"cited_title":"Fully forward mode training for optical neural networks.Nature, 632(8024):280–286, 2024","cited_arxiv_id":null,"evidence_quote":"Fully forward mode training, recovered as reciprocal hardware with a spatial symmetry relocating the adjoint injection to the input."},{"cited_title":"Training of photonic neural networks through in situ backpropagation and gradient measurement.Optica, 5(7):864–871, 2018","cited_arxiv_id":null,"evidence_quote":"In-situ backpropagation gradients for photonic networks, the Helmholtz-type linear stationary instance of the construction."},{"cited_title":"Fully nonlinear neuromorphic computing with linear wave scattering.Nature Physics, 20(9):1434–1440, 2024","cited_arxiv_id":null,"evidence_quote":"Scattering-matrix learning rules for wave systems, shown equivalent to the adjoint identity read at single ports."},{"cited_title":"Localization transitions in non-Hermitian quantum mechan- ics.Physical Review Letters, 77(3):570–573, 1996","cited_arxiv_id":null,"evidence_quote":"The Hatano–Nelson chain, the example system for the twisted-reciprocal non-reciprocal extension."}],"review_version":1}