A Lagrangian View of Flow Matching

Abstract: Modern explicit-time generative models, such as Flow Matching [Lipman et al., 2023] and Rectified Flow [Liu et al., 2023], are typically derived top-down via Optimal Transport and the continuity equation. This standard Eulerian approach focuses on the macroscopic transport of probability mass. In this paper, we present an alternative, bottom-up mechanical derivation grounded in a Lagrangian (particle-centric) perspective. By analyzing the local Taylor expansion of a continuous denoiser, we motivate a strict invariance condition required for optimal, singlestep generation: the conservation of target identity. Enforcing this condition yields a governing quasi-linear advection Partial Differential Equation (PDE). We demonstrate that solving this PDE via the Method of Characteristics analytically yields the straight-line trajectories of Flow Matching. This geometric perspective isolates the Jacobian of the denoiser as the primary source of trajectory curvature, providing a direct mathematical explanation for why straight-line flows enable massive step sizes, and why empirical models require distillation to flatten intersecting characteristics.

• The paper derives a Lagrangian advection condition, ∂tf+Jfv=0, and shows that combining denoiser invariance with the residual field v=c(t)(x−f) recovers Flow Matching and Rectified Flow’s straight-line characteristics.
• The paper identifies the denoiser Jacobian as a source of target drift and numerical stiffness, while a two-mode example reports invariance-violation peaks of about 9.23 near characteristic crossings versus 2.65 for distant trajectories.
• The paper interprets Reflow and distillation as methods for reducing conflicting target assignments at intersecting characteristics, but emphasizes that uniqueness, zero posterior uncertainty, and universal Jacobian behavior require additional assumptions.

Problem formulation and central thesis

The paper develops a Lagrangian account of Flow Matching (FM) and Rectified Flow (RF), replacing the usual Eulerian derivation based on marginal probability transport with a particle-centric condition on the denoiser. Let $f(x,t)$ denote a continuous denoiser predicting the terminal clean state from an intermediate state $x$ at time $t$. The central proposal is that an optimal trajectory should preserve target identity:

$f(x(t),t)=x_0$

for every point along the trajectory associated with the clean sample $x_0$. Differentiating this condition yields

$$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$

where $v(x,t)=dx/dt$ and $J_f$ is the spatial Jacobian of the denoiser. The resulting first-order quasi-linear advection PDE expresses a local compatibility condition between denoiser evolution and particle motion.

The paper’s principal claim is not that this PDE uniquely derives the FM velocity field. It explicitly concedes that the PDE admits multiple velocity fields for a given denoiser. Instead, the argument proceeds by introducing the residual ansatz

$v(x,t)=c(t)\big(x-f(x,t)\big),$

motivated by Tweedie-style denoising flows, and then imposing target invariance along the resulting characteristics. Under this ansatz, the target remains constant along each self-consistent path, reducing the dynamics to

$\frac{dX(t)}{dt}=c(t)\big(X(t)-x_0\big).$

Integration gives

$x$0

with $x$1 and $x$2 an integration constant. These are the straight-line characteristics associated with FM and RF. The derivation therefore provides a complementary interpretation of straight trajectories: they arise from enforcing denoiser-output invariance on a residual-driven characteristic family, rather than from an initial Optimal Transport construction (2609.00198).

From denoiser drift to the advection PDE

The paper begins with the total differential of the denoiser along a finite trajectory increment:

$x$3

This decomposition identifies two sources of target variation. Spatial motion changes the prediction through $x$4, while the explicit time dependence of the denoiser contributes $x$5. A finite-step solver that evaluates the denoiser at $x$6 and then at $x$7 encounters both effects. Its second prediction differs approximately by

$x$8

The paper interprets this changing prediction as the local mechanism behind the small-step requirements of conventional diffusion samplers. If the denoiser target moves appreciably during a solver step, the initial residual no longer accurately represents the destination at the end of the step. The resulting discretization error is especially severe when the denoiser has large spatial sensitivity.

The Lagrangian formulation imposes the stronger condition that the prediction remain unchanged on a valid particle path. Applying the chain rule gives the advection PDE

$x$9

This equation is best understood as a consistency relation, not as a complete generative model. It constrains the velocity field only through its action under the Jacobian. In high-dimensional settings, $t$0 may have a nontrivial null space, and multiple velocity fields can satisfy the same constraint. The paper’s subsequent derivation consequently depends essentially on the residual velocity ansatz; the straight-line result does not follow from the PDE alone.

Recovery of straight-line Flow Matching characteristics

The Method of Characteristics is applied after selecting the residual velocity field. Along a curve satisfying

$t$1

the denoiser evolves according to

$t$2

The invariance requirement sets this derivative to zero. With the boundary condition $t$3, the denoiser output is therefore fixed to $t$4 along the characteristic. The residual field then becomes a linear displacement from the target, yielding the separable ODE and the solution $t$5.

This construction clarifies the logical structure of the result. The residual ansatz selects the class of admissible characteristics; invariance makes the denoiser target constant; and the boundary condition identifies that constant with the clean sample. The Method of Characteristics then integrates the resulting scalar schedule. Thus, the paper establishes that the residual ansatz is compatible with FM-style straight paths and that invariance is sufficient to determine those paths within this restricted construction. It does not establish uniqueness among all solutions of the advection PDE.

The distinction is important for interpreting the paper’s relationship to standard FM theory. Eulerian formulations construct a velocity field whose induced continuity equation transports one marginal distribution to another. The present derivation instead characterizes individual trajectories by preserving their terminal identity. It supplies a local, denoiser-level explanation for a global transport construction, but it does not replace the distributional conditions required to guarantee that the generated marginal matches the data distribution.

The Jacobian penalty and numerical stiffness

Rearranging the PDE gives

$t$6

The paper uses this expression to identify $t$7 as the mechanism by which motion induces target drift. If the trajectory follows a characteristic satisfying the invariance condition, this drift is canceled exactly and $t$8 vanishes along the path. If the learned vector field deviates from such a characteristic, the target changes during transport, and the solver must resolve that change numerically.

Under an idealized MMSE-denoising interpretation, the paper relates the denoiser Jacobian to posterior uncertainty, writing schematically $t$9. Assuming the Jacobian is symmetric positive semidefinite, it obtains

$f(x(t),t)=x_0$0

The claimed implication is that target drift is opposed to the direction of motion. This provides a local explanation for conservative solver behavior: large Jacobian curvature causes the denoiser’s target to retreat as the state advances, increasing the risk that a finite step overshoots the locally valid prediction.

The argument requires several assumptions that constrain its scope. The positive-semidefinite conclusion depends on a symmetric, MMSE-like denoiser Jacobian and does not automatically apply to an arbitrary neural denoiser. Moreover, the precise relationship between $f(x(t),t)=x_0$1 and posterior covariance depends on the observation model and noise parameterization; proportionality may include schedule-dependent factors and need not imply that the covariance itself diverges. In the paper’s toy model, the divergence is specifically attributable to the increasingly sharp posterior transition near the mode boundary and the associated schedule scaling.

Within these assumptions, the geometric interpretation is useful: straight characteristics suppress the component of target variation generated by motion, whereas curved trajectories expose the sampler to large local Jacobian effects. The paper therefore links the empirical advantage of large-step FM sampling to a differential invariance condition rather than treating straightness solely as a visual property of the transport paths.

Intersecting characteristics and the role of distillation

The paper argues that ideal straight characteristics are generally incompatible with a deterministic denoiser when paths carrying different terminal identities intersect. At an intersection, the same $f(x(t),t)=x_0$2 would require multiple values of $f(x(t),t)=x_0$3, one for each originating data point. A deterministic network trained with a regression objective instead represents an averaged prediction or velocity. This conditional averaging bends the learned field away from the independently specified straight segments.

This mechanism provides the paper’s Lagrangian interpretation of Reflow and related distillation procedures. Reflow constructs new training pairs from trajectories generated by an initial model, thereby replacing independently sampled, potentially conflicting couplings with empirically aligned source-target pairs. In the paper’s terminology, distillation attempts to uncross the characteristics. Once conflicting target assignments are removed, the denoiser can approximate a single-valued inverse flow map more accurately, reducing target drift and permitting fewer solver evaluations.

The manuscript makes a strong claim that Reflow drives posterior uncertainty toward zero, implying $f(x(t),t)=x_0$4 and thereby eliminating the source of target drift. This should be interpreted as an idealized limiting description rather than a general theorem. A globally deterministic flow map can still have nonzero spatial derivatives, and a zero denoiser Jacobian is not necessary for target invariance: the condition is $f(x(t),t)=x_0$5, not $f(x(t),t)=x_0$6. Likewise, finite-capacity networks, imperfect couplings, and residual characteristic overlap prevent exact uncertainty elimination in practical models. The conceptual contribution is therefore the identification of characteristic conflict as a local source of averaging and curvature, not a proof that distillation universally produces vanishing posterior covariance.

Two-mode crossing example

The one-dimensional two-mode model makes the proposed mechanism analytically explicit. The data variable takes values in $f(x(t),t)=x_0$7, while the interpolation is

$f(x(t),t)=x_0$8

The Bayes-optimal denoiser and residual velocity are given by

$f(x(t),t)=x_0$9

At high noise, $x_0$0 is comparatively smooth and averages the two modes. As $x_0$1 decreases, it becomes a sharp transition near $x_0$2. The corresponding Jacobian,

$x_0$3

concentrates and grows near the decision boundary. This is the precise setting in which a small region of unresolved target ambiguity produces strong local sensitivity.

Figure 1: The denoiser sharpens from a nearly linear average at high noise to a step-like mode selector as $x_0$4 decreases.

Trajectories initialized far from the crossing, such as $x_0$5, remain close to straight paths because the denoiser resolves the terminal mode early. Initializations near the boundary, such as $x_0$6, exhibit pronounced curvature and remain near the ambiguous region before eventually committing to a mode. The exactly symmetric initialization $x_0$7 remains at the unstable fixed point with $x_0$8 and never resolves the mode identity.

Figure 2: Backward-integrated residual trajectories are nearly straight for distant noise samples and strongly curved near the characteristic crossing.

The paper also evaluates the invariance violation $x_0$9 along these paths. Near-crossing trajectories sustain large violations over a longer interval and reach a reported peak of approximately $$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$0, compared with approximately $$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$1 for distant trajectories. The ratio, about $$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$2, quantitatively supports the proposed diagnostic: the trajectories with the strongest curvature also exhibit the largest failure of denoiser invariance.

Figure 3: The log-scale invariance diagnostic remains elevated near characteristic crossings and localizes unresolved target ambiguity.

The diagnostic is attractive because it requires only successive denoiser evaluations and no ground-truth target. It could therefore be used for adaptive step-size control or for identifying samples and regions that would benefit from additional distillation. However, the example does not establish calibration of the diagnostic on realistic data. Its interpretation is exact only in the analytically specified two-mode model; on complex manifolds, large denoiser variation may also reflect approximation error, schedule effects, or numerical differentiation error rather than characteristic intersection alone.

Limitations and open questions

The principal limitation is that the invariance principle is an idealization. A denoiser trained by conditional regression generally estimates a conditional expectation, not a globally valid inverse flow map. When the conditional distribution of terminal targets is multimodal, exact target identity cannot be recovered from $$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$3 alone. The paper’s framework identifies this non-injectivity but does not provide a criterion for when a learned representation is sufficiently close to injective for one-step sampling.

The derivation also depends on the residual ansatz $$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$4 and on a boundary condition that identifies $$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$5 with a fixed clean target. Other velocity fields may satisfy the advection PDE, and the paper does not characterize the complete solution set or establish an optimality principle selecting the residual field. In addition, the treatment of $$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$6 as a posterior-covariance-related positive-semidefinite operator is restricted to particular denoising models and cannot be assumed for arbitrary neural architectures.

Several empirical questions remain open within the paper’s framework. It is not shown whether the scalar statistic $$\frac{\partial f}{\partial t}(x,t)+J_f(x,t)v(x,t)=0,$$7 predicts local integration error better than conventional residual, curvature, or score-norm diagnostics. Nor is there a quantitative result connecting the distribution of this statistic to FID, likelihood, or perceptual degradation under reduced solver budgets. Finally, the one-dimensional crossing example demonstrates a mechanism but does not quantify how characteristic intersections should be defined or measured in high-dimensional latent spaces.

Conclusion

The paper presents a Lagrangian formulation of Flow Matching based on conservation of denoiser target identity along particle trajectories. Its advection PDE exposes the denoiser Jacobian as the local coupling between motion and target drift. When combined with a residual velocity ansatz, the invariance condition yields the straight-line characteristics of FM and RF. The two-mode example then illustrates how unresolved characteristic crossings produce denoiser sharpening, trajectory curvature, and measurable invariance violations. The framework is most valuable as a local, mechanistic interpretation of solver behavior and distillation; its stronger claims about uniqueness, covariance collapse, and universal Jacobian behavior remain conditional on the stated modeling assumptions.