{"id":"d309c91f-e3ae-4630-a0ca-42c1e16e5d4d","arxiv_id":"2412.00617","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"A flow-matching control law, defined as the conditional expectation of stochastic-bridge control inputs, steers initial to target distributions through deterministic or stochastic linear control systems.","lead":"The paper derives feedback control laws that steer a probability distribution through a linear control system with limited control channels, using a flow-matching construction based on conditional expectations of stochastic-bridge controls. It gives explicit formulas for Gaussian and mixture-of-Gaussians targets and demonstrates the method numerically with neural-network approximations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 7 and the abstract overclaim the deterministic case: for ε=0, the conditional-expectation feedback cannot branch from an atom of P0, so arbitrary P0→P1 fails; a 1D two-point target is a counterexample.","rationale":"The reader's weakest assumption concerned controllability, the noise-channel restriction, and the lack of guarantees for the regression approximation. These are legitimate but do not attack the fully stated scope of the paper. The more load-bearing gap is the unqualified inclusion of the deterministic case. The paper explicitly says the deterministic construction is identical, and Theorem 7 is stated for ε ∈ R, which includes ε=0. For a deterministic controllable system, a feedback law depending only on the current state cannot make a single initial point branch into multiple distinct endpoints; the conditional-expectation construction averages over the target endpoints that share the initial point, so the closed-loop trajectory from an atom remains a single path. The 1D two-point example makes this concrete and satisfies all assumptions (controllability holds). This does not disprove the stochastic theorem for ε>0, where Brownian noise provides the needed randomness, but it shows the paper's central claim as written is false without an additional restriction such as ε>0 or, for the deterministic case, regularity/absolute continuity of P0 and well-posedness of the resulting flow. The verdict remains conditional because the stochastic core appears sound and the overclaim is fixable by amending the scope and adding existence hypotheses; however, the manuscript should not be accepted as-is without that correction.","tokens_in":10403,"tokens_out":22799,"duration_ms":240590,"concrete_test":"Run the closed-loop construction for the 1D system \\dot X=u with ε=0, P0=δ_0, P1=0.5δ_{-1}+0.5δ_1. Compute \\bar k from (5)/(10): X^z_t=ty, u^z_t=y, so \\bar k(t,ξ)=E[y|X^z_t=ξ]. Solve the ODE \\dot X_t=\\bar k(t,X_t) from X_0=0; the solution is unique (e.g., X_t≡0 with a consistent extension of \\bar k off the support {±t}), so the terminal law is δ_0, not P1. This settles that the deterministic case cannot solve Problem 3 for arbitrary distributions. If the intent is Theorem 7 only for ε>0, repeat with ε=0.1 and verify that the terminal law converges to the mixture target.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's stated scope includes the deterministic case (abstract; Section 3: 'the construction for the deterministic case is identical'), and there the central construction is not valid for arbitrary distributions. With ε=0, for each z the interpolant X^z_t is deterministic, and the feedback law \\bar k(t,ξ)=E[u^z_t|X^z_t=ξ] averages over all z mapping to ξ. At an atom of P0, several z share the same initial point, so \\bar k(0,x) is a mean control; a deterministic closed-loop ODE cannot split this point into multiple branches. Concretely, for the scalar system \\dot X_t = u_t (A=0, B=1, which is controllable), take P0=δ_0 and P1=(δ_{-1}+δ_1)/2. Formula (7) gives u^z_t=y and X^z_t=ty, hence \\bar k(0,0)=0 and the closed-loop trajectory from 0 is unique, so the terminal law is a single point, not P1. Thus Problem 3 as stated ('any pair of distributions') is false for ε=0. The proof of Theorem 7 silently assumes existence of a well-posed solution of (6) with the conditional-expectation drift; for ε=0 and atomic P0 no such solution can realize the interpolant's law. The stochastic case ε>0 is not refuted by this example, but the paper's deterministic claim and Theorem 7 as written (ε ∈ R) need a restriction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends flow matching to linear control systems of the form dX_t = AX_t dt + B(u_t dt + ε dW_t). The authors construct stochastic bridges between pairs (x,y) using the controllability Gramian, derive the explicit bridge control (7) and marginal (8), and then propose the feedback law (10), defined as the conditional expectation of the bridge control given the bridge state. Theorem 7 claims that this feedback law steers any initial distribution P0 to any target distribution P1, and that the controlled process has the same law as the bridge process. Analytical formulas are given for Gaussian and mixture-of-Gaussian targets, a numerical regression procedure is proposed, and experiments are shown for several 2D, 4D, and 8D systems.","tokens_in":10694,"tokens_out":9182,"duration_ms":136674,"significance":"The stochastic construction is conceptually attractive: it provides an explicit, parameter-free feedback law in the stochastic setting, avoiding the solution of an optimal control problem, and it extends flow matching to systems with constrained control channels. The tower-property argument in Theorem 7 is elegant and, under suitable well-posedness assumptions, gives a clean law-matching proof. The paper also ships reproducible code and compares the generated samples with the bridge samples using MMD and W2 distances. However, the claimed validity for arbitrary distributions in the deterministic case is false as stated, and the mixture-of-Gaussians formula in Corollary 8 contains a concrete error. These issues need correction before the paper can be accepted.","major_comments":[{"comment":"The claim that Problem 3 is solved for 'any pair of distributions' in the deterministic case is false. For the scalar system dX_t/dt = u_t (A=0, B=1, which is controllable), take P0 = δ_0 and P1 = (δ_{-1}+δ_1)/2. The bridge formulas (7)-(8) give X^z_t = ty and u^z_t = y for z=(0,y). The control law (10) at t=0 is kbar(0,0)=E[y|X^z_0=0]=(−1+1)/2=0, so the deterministic closed-loop ODE has the unique trajectory X_t≡0 and cannot realize the two-point target P1. Thus the statement in Section 3 that the deterministic construction is 'identical' to the stochastic one is not correct as written, and the theorem/abstract need a restriction (e.g., ε>0, or non-atomic P0 with additional regularity for ε=0).","section":"Section 3, Theorem 7 and Abstract"},{"comment":"The formula for the mixture weights w'_l is missing the component-dependent normalizing constants. Writing C_l = R_t Q0 R_t^T + S_t Q_l S_t^T + ε^2 Σ_t, the correct posterior weight is proportional to w_l |C_l|^{-1/2} exp(−(1/2)(ξ−R_t m0−S_t m_l)^T C_l^{-1}(ξ−R_t m0−S_t m_l)), with the determinant factor depending on l. As printed, the displayed formula is only correct when all Q_l are identical; otherwise the weights are incorrectly normalized and the conditional expectation (14) is wrong. This should be corrected or the determinant factors should be incorporated into the definition of w'_l.","section":"Corollary 8, Eq. (14)"},{"comment":"The proof shows that if a solution X_t of (6) with the feedback law (10) exists, then its law satisfies the same Fokker-Planck equation as the bridge process X^z_t. It does not establish existence or uniqueness of such a solution for arbitrary P0,P1. The drift kbar(t,ξ) can be undefined or irregular, as the deterministic atomic example in the first comment shows, and no well-posedness assumptions are stated. The theorem should be restated with explicit conditions under which the SDE (6) with drift kbar has a unique solution, or it should be formulated as a formal result pending those conditions.","section":"Theorem 7 proof"}],"minor_comments":[{"comment":"There is a duplicated article in 'through a a deterministic or stochastic linear system'; it should be 'through a deterministic or stochastic linear system'.","section":"Section 3, first paragraph"},{"comment":"The notation 'E(y|X^z_t=ξ]' has a mismatched parenthesis; it should be 'E[y|X^z_t=ξ]'.","section":"Equation (12)"},{"comment":"There are small typos: 'due the the equality' should be 'due to the equality', and 'It ˆo rule' contains a formatting artifact (the hat is misplaced).","section":"Theorem 7 proof"},{"comment":"'spacial case' should be 'special case'.","section":"Corollary 8 proof"},{"comment":"The caption lists three systems (a)-(c) but does not explain what the colors or curves represent, nor how the ε values are distinguished; please expand the caption.","section":"Figure 1 caption"},{"comment":"The text says 'The left panel shows...' and 'The second panel compares...', but the figure has three panels per row and the second and third rows are described only by row; the panel references should be made explicit.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The deterministic two-point counterexample is a genuine overclaim in the abstract and in Section 3, but it is localized to the ε=0 regime; the stochastic core with ε>0 appears sound under a well-posedness assumption. The mixture-weight error in Corollary 8 is concrete and fixable. I recommend asking the authors to add explicit assumptions, correct the deterministic claim, and fix the normalization in Eq. (14). This is a major revision rather than a rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: Theorem 7 is right for ε>0, but the deterministic-case claim is too broad. The stress-test example holds: for dx/dt=u, P0=δ_0, P1=½(δ_1+δ_{-1}), the bridge control u^z_t=y gives k̄(0,0)=0, so the closed-loop trajectory from 0 is unique and cannot reach both targets. This is a real counterexample to Problem 3 as stated, since the abstract and Section 3 explicitly include the deterministic case.\n\nWhat the paper does well: it adapts flow matching to linear control systems with control-channel constraints, derives the bridge-based feedback law (10), and gives closed-form Gaussian and mixture-of-Gaussians conditionals. The tower-property argument in the proof of Theorem 7 is short and correct for ε>0, and the bridge formulas (7)-(8) are consistent with Gaussian conditioning. The numerical experiments show the method works on several 2D and higher-dimensional systems, and the code appears to be available.\n\nThe soft spots, in proportion: the deterministic atom-splitting issue is load-bearing for the stated scope and needs a fix, either by restricting Theorem 7 to ε>0 or by assuming P0 and P1 are absolutely continuous in the deterministic case. The mixture formula (14) is stated without proof—'removed on account of space'—which is unsatisfying but probably not incorrect. The numerics would be stronger with error bars and a baseline comparison (e.g., unconstrained flow matching or an optimal transport solver); as they stand, they are qualitative proof-of-concept. Remark 6's restriction to noise entering through the same channel as control is a real limitation and should be surfaced in the abstract.\n\nOverall, the stochastic core is a legitimate contribution and the deterministic overclaim is fixable. I'd send this to a serious referee, with a request to address the counterexample and tighten the claims. The paper is worth discussing in a reading group because the deterministic flaw is instructive and the stochastic part is a useful tool.","headline":"The stochastic-case construction is sound and useful, but the paper overclaims the deterministic case: conditional-expectation feedback cannot split atoms, so 'any pair of distributions' is false for ε=0.","tokens_in":11204,"tokens_out":4106,"would_cite":true,"duration_ms":43269,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E20","93B05","60J60"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that a feedback law given by the conditional expectation of the control inputs of stochastic bridges of a linear system steers any initial distribution to any target distribution through that system.","keywords":["flow matching","stochastic linear control systems","probability density steering","stochastic bridge","feedback control","controllability Gramian","Gaussian mixture control","mean-field control"],"falsifier":"Fix a controllable linear system, take $P_0$ and $P_1$ to be two-point distributions, compute the exact conditional expectation (10) by replacing the bridge law with a fine histogram, and simulate the closed loop; the theorem predicts the time-marginals coincide with the bridge marginals at every time, so any systematic mismatch is a direct refutation of the claimed law-matching property.","tokens_in":10211,"feed_emoji":"🎛️","tokens_out":9658,"duration_ms":89750,"temperature":0.7,"pith_summary":"The paper shows that flow matching still works when the flow is constrained to a linear control system: instead of freely choosing a velocity field, one can only push the state through the given control channel. Its central claim is that a feedback law equal to the conditional expectation of the bridge control inputs steers the state distribution from any initial law to any target law, for both deterministic and stochastic linear systems. This matters for applications where agents or particles are manipulated only through control actions, such as robot swarms and stochastic thermodynamics, because it provides a constructive controller without solving an optimal control problem. For Gaussian and Gaussian-mixture targets the paper derives an explicit formula for the feedback law, and otherwise it proposes a regression-based approximation.","feed_headline":"One feedback law steers any start distribution to any target","feed_subtitle":"The conditional expectation of bridge controls reproduces the bridge's law; explicit formulas cover Gaussian and mixture targets.","key_machinery":"The central object is the stochastic bridge $X^z_t$ of the linear system, whose mean is $R_t x + S_t y$ and whose covariance is $\\epsilon^2\\Sigma_t$, together with its input $u^z_t = B^\\top e^{(1-t)A^\\top}\\Phi_{1-t}^{-1}(y - e^{(1-t)A}X^z_t)$. The mechanism that carries the argument is the identity $\\bar{k}(t,\\xi) = \\mathbb{E}[u^z_t \\mid X^z_t = \\xi]$: averaging the bridge controls over all endpoints that pass through $\\xi$ produces a feedback law whose drift matches the bridge's drift, so the closed-loop marginals coincide with the bridge marginals at every time. The controllability Gramian $\\Phi_t$, through the inverse $\\Phi_{1-t}^{-1}$, is what makes the bridge reach the prescribed endpoint and is therefore the reason the controllability assumption is needed.","core_discovery":"On the paper's own terms, the discovery is Theorem 7. For the stochastic linear system $dX_t = AX_t\\,dt + B(u_t\\,dt + \\epsilon\\,dW_t)$, if one builds, for each paired endpoint $z=(x,y)$ drawn from any coupling of $P_0$ and $P_1$, the stochastic bridge $X^z_t$ with bridge control $u^z_t$, then the closed-loop system driven by $\\bar{k}(t,\\xi) = \\mathbb{E}[u^z_t \\mid X^z_t = \\xi]$ has the same time-marginal law as $X^z_t$. In particular, $X_1$ follows $P_1$. The proof compares the evolution of expectations under the closed loop and under the bridge process, and the only identity needed is the tower property of conditional expectation. The same law-matching statement holds in the deterministic limit, where the bridge is the minimum-norm interpolating trajectory of the linear system.","pith_inferences":["Editorial inference: because the coupling $\\Pi$ between the endpoint distributions is arbitrary, one can choose it to shape trajectories or reduce control effort without changing the theorem, for example by taking an optimal-transport coupling.","Editorial inference: the proof's mechanism is essentially the tower property of conditional expectation, so in principle the same feedback construction would apply to any bridge process one can sample from; the paper's explicit linear-system formulas are what make the construction directly usable in that setting.","Editorial inference: a quantitative extension would be a finite-sample bound relating the regression error in (15) to the error in the terminal distribution, a step the paper does not attempt."],"forward_implications":["Any distribution pair can be connected through the constrained linear dynamics, not only through an unconstrained velocity field, and the terminal law is matched exactly in the ideal setting where the conditional expectation is known.","For Gaussian initial and Gaussian or mixture-of-Gaussians target laws, the feedback law is given explicitly by formula (14), so no iterative optimization is needed in that case.","The stochastic and deterministic problems are solved by exactly the same feedback law, and the stochastic bridge tends to the deterministic interpolant as the noise intensity goes to zero.","When the exact conditional expectation is unavailable, the proposed least-squares regression (15) approximates it, and the numerical experiments report small normalized MMD and $W_2$ distances in two-, four-, and eight-dimensional examples.","The argument extends to control-affine systems whenever a sampler for the corresponding stochastic bridge exists."],"supporting_citations":[{"why":"Introduces the flow matching methodology that the paper adapts to constrained linear control systems.","marker":"Lipman et al., 2022"},{"why":"Provides the conditional or rectified flow formulation whose regression-style control approximation the paper generalizes.","marker":"Liu et al., 2022"},{"why":"Supplies the stochastic interpolant view that underlies the bridge construction used here.","marker":"Albergo and Vanden-Eijnden, 2022"},{"why":"Derives stochastic bridges of linear systems, which the paper uses as the endpoint-to-endpoint interpolants.","marker":"Chen and Georgiou, 2015"},{"why":"Provides the reverse-time diffusion formalism used to derive the bridge control law in Proposition 4.","marker":"Anderson, 1982"},{"why":"Supplies the time-reversal of diffusions technique used in the proof of the stochastic bridge formula.","marker":"Haussmann and Pardoux, 1986"},{"why":"Gives the controllability Gramian and minimum-norm control results that underlie Proposition 1 and the invertibility of $\\Phi_t$.","marker":"Basar et al., 2020"},{"why":"Supplies the Gaussian-sum filtering identity used to write the explicit conditional expectation for mixture-of-Gaussians targets.","marker":"Alspach and Sorenson, 1972"}],"fun_headline_variants":["Single control law guides any distribution to any target","Steer arbitrary start and target laws via one feedback rule","Conditional expectation yields exact steering for linear systems","Universal distribution steering for stochastic linear controls","One bridge-based law moves any distribution to any other"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system must be controllable, and the noise must enter through exactly the same input channels as the control; otherwise the bridge formula the whole construction is built on is unavailable.","fun_headline_variants_meta":{"raw":{"variants":["Single control law guides any distribution to any target","Steer arbitrary start and target laws via one feedback rule","Conditional expectation yields exact steering for linear systems","Universal distribution steering for stochastic linear controls","One bridge-based law moves any distribution to any other"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1410,"prompt_tokens":846,"completion_tokens":564,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":492}},"tokens_in":462,"tokens_out":564,"duration_ms":6177,"temperature":1.0,"reasoning_tokens":492,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:11:04.837815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a controllable linear system, take $P_0$ and $P_1$ to be two-point distributions, compute the exact conditional expectation (10) by replacing the bridge law with a fine histogram, and simulate the closed loop; the theorem predicts the time-marginals coincide with the bridge marginals at every time, so any systematic mismatch is a direct refutation of the claimed law-matching property.","supporting_citations":[{"cited_title":"Nonlinear bayesian estimation using gaussian sum approximations","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian-sum filtering identity used to write the explicit conditional expectation for mixture-of-Gaussians targets."}],"review_version":1}