Flow Matching under Noisy Latent Structure: Beyond Exact Low-Dimensional Support

1 Introduction

• The motivating statistical question is how finite-sample velocity error propagates to generated distributions, against prior FM analyses centered on ambient regularity or exact low-dimensional support.Earlier intrinsic-dimensional FM results cited here assume support on low-dimensional subspaces or smooth manifolds.Source: §1 · pp. 1–2
• The paper extends the low-dimensional FM question beyond exact-support settings to a Hӧlder latent generator perturbed by nondegenerate ambient Gaussian noise.The model has a full-dimensional target law, while systematic variation is parameterized by a d-dimensional latent variable.Source: §1 · p. 2

1.1 Main Contributions

• The velocity-estimation and W2 bounds have leading sample-size factors n^(-2β/(2β+d)) and n^(-β/(2β+d)), respectively, while ambient and noise dependence remains explicit.The analysis applies with fixed positive target noise, which keeps the interpolation nondegenerate over the full time interval.Source: §1 · p. 2, §1.1 · p. 2
• A spatially regular ReLU class approximates the population velocity with leading resolution complexity ε^(-d), despite the target's full ambient support.The corresponding excess-risk guarantee keeps dependence on D, H, σ, and confidence explicit.Source: §1.1 · p. 2

1.2 Other Related Work

• The theory distinguishes its setting from noisy-observation approaches that seek to recover a clean distribution: here, ambient noise is part of the target itself.The related-work discussion also distinguishes this statistical setting from dequantization and FM on non-Euclidean state spaces.Source: §1.2 · pp. 2–3

2.1 Notation and conventions

• The notation section defines Hβ(Ω, Rm, H) as the coordinatewise Hölder class used in the analysis and specifies the norm conventions.Its precise multi-index definition is deferred to Appendix A.1.Source: §2.1 · p. 3

2.2 Noisy latent Flow Matching

• The target is generated from a d-dimensional latent variable through g∗, then perturbed by ambient Gaussian noise; the interpolation starts from an independent standard Gaussian source.For fixed positive σ, the target is full-dimensional and the interpolation variance stays positive throughout [0, 1].Source: §2.2 · p. 3
• The population velocity decomposes into a known affine component and a posterior component based on the conditional mean of g∗(U).The latter component is the nonlinear part approximated by the neural network.Source: §2.2 · p. 4

2.3 Neural network class

• The paper uses sparse feed-forward ReLU networks, with depth, width, sparsity, and parameter magnitude controlling the network class.The velocity network maps from state-time inputs of dimension D + 1 to outputs of dimension D.Source: §2.3 · p. 4

3 Neural Approximation of the Population Velocity

• The approximation problem reduces to the latent posterior mean, while retaining spatial regularity needed for subsequent transport analysis.The posterior is defined on unbounded ambient state space even though its exploitable structure comes from the d-dimensional generator.Source: §3 · pp. 4–5
• The spatial envelope has integrated and uniform bounds scaling with 1 + σ^(-2) and 1 + σ^(-3), respectively.These bounds specify the noise dependence of the profile used in transport analysis.Source: §3 · p. 5
• The leading approximation complexity scales as ε^(-d), with ambient dimension D remaining in explicit polynomial factors and Γε.The comparator also satisfies the spatial profile γx used later in transport analysis.Source: §3 · p. 5
• Gaussian smoothing bounds the posterior mean's spatial Jacobian, and the neural realization controls first derivatives as well as function values.This regularity supports ODE stability and accompanies the integrated velocity approximation bound.Source: §3 · p. 5
• Generator-to-velocity stability transfers a latent-generator approximation to the corresponding population-velocity error under the same noisy interpolation.The proposition is stated for bounded measurable generators and σ > 0.Source: §3 · p. 6
• Gaussian concentration localizes approximation to a high-probability spatial region, avoiding a bounded ambient-support assumption.Within the region, posterior numerator and denominator reduce to low-dimensional integral terms that are realized while controlling the spatial Jacobian.Source: §3 · p. 6

4 Generalization

• Excess Flow Matching risk is equivalent to time-integrated L2(πt) velocity error, making velocity estimation the target of the generalization analysis.The auxiliary Gaussian source and time are integrated exactly into the loss, so finite-sample randomness comes only from target observations.Source: §4 · p. 6

4.1 Error Decomposition

• Although squared Flow Matching loss is unbounded, localized moment and tail bounds support an unbounded Bernstein inequality for the excess loss.The error is decomposed into approximation and estimation terms.Source: §4.1 · p. 7, §4.2 · p. 8

4.2 Covering and Generalization

• With ε = n^-1/(2β+d), the ERM has a high-probability finite-sample bound whose algebraic sample-size exponent is 2β/(2β+d), set by latent dimension d and smoothness β.The dependence on ambient dimension D, noise σ, Hӧlder radius H, confidence level, and logarithmic complexity remains explicit.Source: §4.2 · p. 8
• Finite network representatives, Gaussian localization, and local loss stability control population and empirical excess losses; balancing ε^2β approximation against ε^-d/n entropy yields the rate.The covering estimate is local to bounded state–time domains, and the ERM comparison and union bound extend control to the data-dependent estimator.Source: §4.2 · pp. 7–8

5.1 Continuous-time transport

• The population flow is globally well posed and transports π0 to πt for every t ∈[0, 1].The proof uses a finite spatial Lipschitz profile and uniform linear growth of the population velocity.Source: §5.1 · pp. 8–9
• The learned flow has a nonasymptotic W2 bound with leading sample-size factor n^-β/(2β+d), with probability at least 1 −δ.The bound follows by propagating velocity-estimation error through the learned ODE.Source: §5.1 · pp. 9–10

5.2 Euler discretization

• The Euler bound separates numerical discretization error from continuous-flow statistical error, with leading algebraic scales h and n^-β/(2β+d).For a uniform grid, h = M^-1, and M at least of order n^β/(2β+d) balances these scales.Source: §5.2 · p. 10

6 Discussion and Conclusion

• With full-dimensional support, the noisy latent-generator setting still yields velocity and W2 bounds whose leading sample-size exponents depend on latent dimension.The velocity bound has leading factor n^-2β/(2β+d), and the continuous-flow W2 bound has n^-β/(2β+d); fixed positive σ keeps interpolation nondegenerate on [0, 1].Source: §6 · p. 10
• The guarantees retain ambient-dimension dependence in prefactors and network complexity, and their constants deteriorate as σ decreases.The bounds are upper bounds, not a minimax characterization, and do not provide a uniform passage to the exact-support limit.Source: §6 · p. 11

A Population Identities and Structural Properties

• The posterior representation gives a version of the population velocity valid for every x and t, with endpoint formulas at t = 0 and t = 1.At t = 0, v*(x, 0) = EG − x; at t = 1, v*(x, 1) = x.Source: A Population Identities and Structural… · p. 15

A.1 Smoothness and network conventions

• Network depth counts affine layers, the final affine layer has no activation, and exact identity padding aligns networks of different depths.Width padding adds zero neurons without increasing sparsity or parameter-magnitude bounds.Source: §A.1 · p. 16

B.1 Consequences of the scale conditions

• For fixed d, β, D, H, and σ > 0, the range satisfying the scale assumptions is nonempty.The first two left-hand sides have the required limits as ε ↓0.Source: §B.1 · p. 16
• The last scale condition follows from ε log(1/ε) →0..Source: §B.1 · p. 16
• The scale conditions provide the small-cell bound used in the posterior construction, and conclusions remain pointwise in fixed σ > 0..Source: §B.1 · p. 16

B.2 Proof of the main approximation theorem

• The latent Taylor partition uses N^d cells, giving leading approximation complexity governed by the latent dimension.The partition contributes a factor of order ε^-d in network size.Source: §B.2 · pp. 16, 18
• Replacing the true generator with its latent surrogate changes the conditional velocity, and the proof transfers this generator error to integrated FM velocity risk.The argument uses the stability result under the true path law.Source: §B.2 · p. 16
• The proof localizes posterior approximation to a controlled, time-dependent state region and bounds the complement's velocity-risk contribution by order Dε^(2β)/σ.A uniform approximation on the compact region is paired with a Gaussian-tail bound outside it.Source: §B.2 · p. 17
• The network approximates a low-dimensional exponential-integral map whose nonlinear input dimension depends on (d, β), even though the state x lies in R^D.The cellwise integrals are assembled from Taylor coefficients, scalar coordinates, and time-dependent coefficients.Source: §B.2 · p. 17
• First-order realizations of the scalar cell modules preserve value approximation while supplying the spatial Lipschitz bound required by the velocity class.The resulting posterior network is inserted into the structured velocity formula to form the comparator.Source: §B.2 · pp. 17–18

B.3.1 Latent Taylor approximation

• The latent Taylor surrogate approximates the generator cellwise, with remainder controlled by the Hӧlder smoothness and cell diameter.Under the theorem's scale assumptions, the surrogate is bounded by 2.Source: §B.3.1 · pp. 18–19

B.3.2 Generator-to-velocity stability

• The generator-to-velocity argument uses relative entropy's O(t^2) behavior near the source endpoint to compensate for the score representation's time singularity.The endpoint-weight integral is π/(2σ).Source: §B.3.2 · pp. 19–20

B.3.3 Localization of the surrogate posterior

• The compact localization remains nondegenerate over the full time interval, allowing neural approximation without a positive stopping time.The state-space complement is handled through a Gaussian-tail contribution.Source: §B.3.3 · pp. 21–22

B.4 Neural realization of the surrogate posterior

• The surrogate posterior is realized from latent-cell value approximations, then refined to a first-order realization and extended to a global network.Its cell integrals depend on finitely many scalar coefficients associated with the latent Taylor polynomials.Source: §B.4 · p. 22, Its architecture satisfies · p. 22

Its architecture satisfies

• The stated architecture has L and log B bounded by order Γ_carch ε, while W and S are bounded by order D^2ε^-dΓ_carch ε.The bounds make the ambient-dimension and latent-cell contributions explicit.Source: Its architecture satisfies · p. 22

B.4.1 Value realization of the latent-cell posterior

• The local Taylor-coordinate decomposition is algebraically exact on each cell.Source: §B.4.1 · p. 26
• Under the small-cell condition, intermediate-coefficient bounds verify the hypotheses for ReLU approximation of each normalized local FM integral.The lemma’s hidden constants depend only on d and β.Source: §B.4.1 · pp. 28, 30, 31
• Assembling the N^d cell networks yields numerator and denominator estimates, with coordinatewise scalar division avoiding a Euclidean-norm conversion and its ambient-dimension factor.Source: §B.4.1 · p. 36

• Parallelizing the D numerator-coordinate networks makes ambient dimension enter network size, while each coordinate’s approximation tolerance remains of order ε′/N.Source: §B.4.1 · p. 36

Denominator control and posterior assembly

• For 1 ≤ l ≤ D, the ratios P◦_l/Q◦ are approximated by applying a robust division network after establishing that the exact denominator is uniformly bounded away from zero on C∗×[0,1].Source: §B.4.1 · pp. 36–37
• An exact affine layer forms D pairs (P_l/2, Q), after which D parallel copies of the scalar division network and exact scaling produce the D-output value network.Source: §B.4.1 · p. 37

• The value-only circuit’s hidden constants depend only on d and β; these are baseline value-module bounds, while the spatial refinement rebuilds the circuit for the final common envelope.Source: §B.4.1 · p. 38

B.4.2 First-order spatial refinement

• Hölder regularity of g* controls latent-variable approximation, while Gaussian smoothing supplies ambient spatial differentiability through the posterior covariance identity.On the tube, the refinement gives Jacobian error at most one and an absolute Jacobian profile 4t/q_t + 1.Source: §B.4.2 · pp. 38, 39, 44
• Fixed-box first-order ReLU modules realize cell integrals and their quotient, with stable division arranged so inactive dyadic blocks contribute zero value and derivative error.Holomorphic-map approximation provides value and first-derivative control; zero-preserving multiplication supports the division construction.Source: §B.4.2 · pp. 39, 40, 44
• The moving tube provides a denominator lower bound and an exact ReLU gate that equals one on the localization set, with detector width and sparsity O(DN^d).The tube is the domain where the network ratio and its spatial derivative are controlled.Source: §B.4.2 · p. 39
• The architecture has W, S ≲_{d,β} D^2ε^-dΓ_carch^ε and L, log B ≲_{d,β} Γ_carch^ε; refinement adds only polylogarithmic factors in 1/ε, not a new ε^-c factor.The spatial refinement preserves the latent-cell count.Source: §B.4.2 · p. 44
• Exact N^-d cell weights keep coordinatewise value error at order ε^β without growth across the N^d cells, which determine the algebraic approximation-scale dependence.The cellwise error therefore has no factor N^d.Source: §B.4.2 · pp. 43–44
• The assembled ReLU field has Jacobian bound ≲_{d,β} 1 + σ^-1 + t/q_t and is globally Lipschitz, while approximation of the spatial derivative is claimed only on the tube.Outside the tube, exact zero preservation supplies a global absolute Lipschitz bound; no global approximation of the target Jacobian is claimed.Source: §B.4.2 · p. 45

B.5 Auxiliary neural-network lemmas

• The auxiliary toolkit combines ReLU composition and identity padding with multiplication, reciprocal, division, and exponential approximations for posterior construction.Because 0 < σ < 1 keeps q_t bounded away from zero, the constructions can approximate time-dependent coefficients and the quadratic and linear interaction terms.Source: §B.5 · pp. 45–47

B.5.1 Constructions for the Flow Matching modules

• The constructions also give ReLU approximations of the quadratic and linear interaction terms by combining multiplication networks with spatial and time branches.The linear interaction uses an affine spatial map and a time-coefficient network before multiplication.Source: §B.5 · pp. 47–48, §B.5.1 · p. 52
• The construction first forms a neural approximation of q_t and establishes q_t ∈ [η, η^-1] for every t ∈ [0,1]; an approximation claim is proved for γ ∈ {0,1,2}.Source: §B.5.1 · pp. 48–49

C Proofs for Generalization

• The main theorem proof is presented first.The risk, truncation, entropy, and concentration lemmas follow in subsections.Source: C Proofs for Generalization · p. 54
• The appendix's proof material is organized around the theorem and its supporting lemmas.The theorem proof comes first, followed by those lemmas.Source: C Proofs for Generalization · p. 54
• The following subsections prove the lemmas used by the main theorem proof.They cover risk, truncation, entropy, and concentration.Source: C Proofs for Generalization · p. 54

C.1 Proof of the finite-sample theorem

• The approximation–estimation balance with ε = n^-1/(2β+d) yields the theorem's stated integrated velocity error.The proof chooses the class parameters from Theorem 3.1 and uses the risk identity to convert the excess-risk bound into integrated velocity error.Source: §C.1 · pp. 54–55

C.2 Risk identity and localized excess loss

• The proof controls localized excess loss using a second-moment bound and a sub-exponential envelope, then obtains fixed-field concentration.The excess-loss bounds use cancellation of the common affine velocity component and clipping of the nonlinear discrepancy; the Bernstein constant is independent of D, L, W, S, and B.Source: §C.1 · p. 54, §C.2 · pp. 57, 59, 60

C.3 Truncation, local loss stability, and finite reduction

• A bounded state–time network cover transfers local stability to a finite deterministic family, with Gaussian truncation controlling the unbounded domain.Constrained representatives are selected from cover elements intersecting the velocity class, and the resulting transfer yields the explicit bound (16).Source: §C.3 · pp. 62–64

C.4 Uniform concentration

• With probability at least 1 − δ, localized excess-loss concentration holds simultaneously over the velocity class.The proof concentrates over the finite representatives and transfers the bound back to the class.Source: §C.4 · p. 64
• The ERM comparison yields an oracle inequality by comparing the estimator with any deterministic field in the class.The proof applies the uniform concentration result to both fields and takes the infimum over comparators.Source: §C.4 · p. 64

D Auxiliary Results for Distributional Convergence

• The appendix collects the sampling analysis's auxiliary growth estimate alongside the Euler discretization result.These are its two stated purposes.Source: D Auxiliary Results for Distributional… · p. 64, §D.1 · p. 65, §D.2 · p. 66

D.1 Linear-growth bound

• The appendix establishes linear growth for velocity fields in the specified class.The radius-two clipping in the class definition supplies the stated growth bounds.Source: §D.1 · p. 65

D.2 Euler discretization

• The Euler analysis couples the continuous learned trajectory and the Euler scheme from the same initial condition.The proof uses the growth estimates and Grönwall's inequality to establish the discretization bound (21).Source: §D.2 · pp. 65–66