betaFeedback

WorksComputer Science

LIFT: Latent Flow Matching for Probabilistic Cell Fate Dynamics with Single Cell Snapshots

Abstract

Lineage tracing links cell states through clonal ancestry, but does not generally identify unique source-descendant pairings. Learning from such data therefore requires a model of conditional outcomes, rather than only the terminal population distribution. We introduce LIFT, a latent flow framework for probabilistic cell-state generation from lineage-informed couplings. Variational inference encodes alternative outcomes in a latent space, while a conditional prior enables sampling from the initial state alone. A Mean Flow decoder predicts future cell state directly, requiring one decoder evaluation per sampling, which makes it convenient to predict cell fate through large scale Monte-Carlo sampling. This construction combines a representation of one-to-many relationships with efficient sampling and an analysis of conditional prediction error. Experiments on synthetic and real datasets demonstrate LIFT's ability to recover both population distribution and cell fate probabilities.

Funding: L.Z. was supported by the National Natural Science Foundation of China (No. 12225102, T2321001, and 12288101) and the National Key Research and Development Program of China 2024YFA0919500.

Introduction

Understanding how an early cell state constrains its future fate is a central problem in developmental biology and cellular reprogramming. Single-cell RNA sequencing provides detailed molecular snapshots, but its destructive nature generally prevents repeated measurements of the same cell. Trajectory inference must therefore reconstruct developmental relationships from population snapshots and additional assumptions [13]. Lineage-tracing technologies provide an experimental source of such information. Methods such as CellTagging and LARRY recover heritable barcodes together with transcriptomic states, linking clonal ancestry to subsequent outcomes [1] [22]. These observations motivate models that use lineage information to predict not only a likely terminal distribution, but also the probabilities of cellular fates.

A shared barcode does not, however, generally determine a unique pairing between individual cells observed at different times. When several source and terminal cells belong to the same clone, their molecular correspondence remains only partially specified. We represent the true endpoint relationship by an unknown coupling π⋆\pi_{\star} and distinguish it from the lineage-informed coupling π\pi constructed for training. The prediction target is the conditional law

K⋆(x,A)=Pπ⋆(Y∈A∣X=x),K_{\star}(x,A)=\mathbb{P}_{\pi_{\star}}(Y\in A\mid X=x),

where xx is the observed initial state and AA is a region of terminal state space. Different source-descendant relationships can have identical marginals but imply different fate probabilities. Recovering this law requires more than matching the endpoint populations.

Probabilistic models of cellular differentiation range from fate mapping on observed states to continuous descriptions of population dynamics. Palantir estimates terminal fate probabilities from transitions on the transcriptomic manifold [15], while CellRank 2 integrates multiple sources of directional and temporal information through Markov state modeling [21]. Waddington-OT estimates probabilistic couplings between successive populations while accounting for growth and death [14], and PRESCIENT learns stochastic cell-state dynamics from expression time series [24]. These approaches already represent heterogeneous outcomes and, in several cases, provide fate probabilities. Without explicit lineage constraints, the inferred fate relationships depend on assumptions or constraints used by each method, but they need not coincide with those supported by lineage observations.

Explicit incorporation of lineage information provides additional constraints on these relationships. LineageOT uses lineage trees to inform ancestral-state estimation and transport coupling [3]; CoSpar reconstructs transition maps from sparse clonal observations using sparsity and local coherence [20]; and moslin combines expression similarity with lineage geometry through fused Gromov–Wasserstein transport [9]. LineageVAE further uses variational latent dynamics to reconstruct historical states toward unobserved common progenitors [11]. These studies establish lineage observations as valuable information for both transition inference and state reconstruction. Nevertheless, an endpoint transition map does not uniquely specify intermediate paths. Estimating a lineage-informed coupling and constructing a continuous generator that retains its conditional structure are thus separate, complementary tasks.

Generative modeling provides tools for constructing continuous conditional models. Flow matching learns time-dependent velocity fields without numerical integration during training [10, 17], and Mean Flow addresses sampling cost by learning interval-average velocities for one-step generation [4], but a deterministic flow initialized solely at the observed cell state produces only one outcome from that state. Augmented Bridge Matching addresses coupling preservation by retaining the initial sample in the conditioning [2], while Variational Rectified Flow Matching uses latent variables to represent multimodal velocity directions [5]. For lineage-informed applications, the remaining task is to combine one-to-many conditional representation with direct time-dependent sampling. This combination motivates the LIFT framework developed below.

We introduce LIFT, a latent flow framework that combines conditional representation, efficient generation, and error control for this task. Building on variational inference and conditional latent-variable modeling [7] [16], an encoder infers latent codes from training pairs, while a conditional prior predicts their distribution from the initial state alone. We formulate Flow Matching and Mean Flow reconstruction surrogates and use the Mean Flow decoder for direct generation. Its average velocity parameterization [4] allows only one decoder evaluation per sampling, and keeping the latent code fixed across time defines a continuous generated path. The coupling can incorporate geometric or clonal information without changing the sampling architecture.

Our theoretical analysis makes the conditional prediction objective explicit. We characterize the best deterministic approximation by its irreducible conditional-variance error and establish reconstruction bounds for both FM and Mean Flow decoders. Under the stated regularity and prior-scale assumptions, the total conditional Wasserstein error is bounded by the discrepancy between π\pi and π⋆\pi_\star, reconstruction under encoder samples, and mismatch between the aggregated posterior and the sampling prior. We further translate this bound into fate-probability estimates, including the effects of classification boundaries and finite Monte Carlo sampling. These results concern a fitted model and do not identify the true coupling or a unique intermediate biological trajectory.

The experiments follow the same conditional prediction perspective. A synthetic branching system tests recovery of both population distributions and branching probabilities. Comparisons across numerical sampling budgets assess the balance between one-step efficiency and prediction accuracy. Finally, mouse hematopoiesis and direct reprogramming demonstrate the use of lineage information for cell fate prediction, with improved probability scores relative to LIFT trained with geometric couplings.

We organize this paper as follows. § 2 formulates the reconstruction problem, § 3 develops LIFT, and § 4 establishes its error estimates. § 5 presents the experiments, followed by § 6 on conclusions and future work.

Problem Formulation and Conditional Fate Reconstruction

We consider trajectory reconstruction from barcode-resolved cell populations observed at two time points. The endpoint measures are given, but the correspondence between individual source and terminal states remains unknown. We distinguish the true lineage-compatible coupling, denoted by π⋆\pi_\star, from the coupling π\pi used for training. Our objective is to recover the conditional transition structure of π⋆\pi_\star.

Barcode-resolved observations and the target dynamics

Consider two observation times, rescaled to 0 and 1, and MM distinct barcodes. For each barcode i=1,2,⋯ ,Mi = 1, 2, \cdots, M, let μ0i,μ1i\mu_0^i,\mu_1^i be the observed nonnegative finite measures on Rd\mathbb{R}^d describing its source and terminal cell populations. Assume that these measures have finite second moments. The total population measures are

μ0=∑i=1Mμ0i,μ1=∑i=1Mμ1i,(1)\mu_0 = \sum_{i=1}^{M} \mu_0^i,\qquad\mu_1 = \sum_{i=1}^{M} \mu_1^i, \tag*{(1)}

where we assume μ0i(Rd)=μ1i(Rd)=wi\mu_0^i(\mathbb{R}^d)=\mu_1^i(\mathbb{R}^d)=w_i, ∑wi=1\sum w_i=1, so that μ0,μ1\mu_0,\mu_1 are probability measures, and π⋆\pi_\star becomes a probability measure as well.

The true endpoint relationship is represented by an unknown coupling π⋆\pi_\star with marginals μ0\mu_0 and μ1\mu_1. Since barcode identity is preserved, this coupling permits no transport between different barcodes. More precisely, it admits a decomposition

π⋆=∑i=1Mπ⋆i,π⋆i∈Π(μ0i,μ1i),(2)\pi_\star= \sum_{i=1}^{M}\pi_\star^i,\qquad\pi_\star^i\in\Pi(\mu_0^i,\mu_1^i), \tag*{(2)}

where Π(μ0i,μ1i)\Pi(\mu_0^i,\mu_1^i) denotes the finite nonnegative measures on Rd×Rd\mathbb{R}^d\times\mathbb{R}^d with the indicated marginals. Equation (2) is a block-diagonal constraint with respect to barcode labels, as all blocks with distinct labels vanish.

Define the set of lineage compatible couplings by

Clin={∑i=1Mγi:γi∈Π(μ0i,μ1i)}⊆Π(μ0,μ1).(3)\mathcal{C}_{\mathrm{lin}}=\left\{\sum_{i=1}^{M}\gamma^i:\gamma^i\in\Pi(\mu_0^i,\mu_1^i)\right\}\subseteq\Pi(\mu_0,\mu_1). \tag*{(3)}

The observations specify this constraint set, but do not generally identify its true member π⋆\pi_\star. In practice, training sampling uses a constructed coupling π∈Π(μ0,μ1)\pi\in\Pi(\mu_0,\mu_1), chosen from Clin\mathcal{C}_{\mathrm{lin}} when the barcode constraints are enforced.

Because the two couplings have the same marginals, we can write their conditional decompositions as follows.

π⋆(dx,dy)=μ0(dx)K⋆(x,dy),π(dx,dy)=μ0(dx)Kπ(x,dy).(4)\begin{aligned} \pi_\star(dx,dy)&=\mu_0(dx)K_\star(x,dy),\\ \pi(dx,dy)&=\mu_0(dx)K_\pi(x,dy). \tag*{(4)} \end{aligned}

Here, K⋆(x,⋅)K_\star(x,\cdot) is the true distribution of terminal states conditioned on the observed molecular state xx, whereas Kπ(x,⋅)K_\pi(x,\cdot) is the corresponding conditional distribution induced by the training coupling. For a measurable partition {B1,…,BK}\{B_1,\ldots,B_K\} of the terminal state space into distinct fates, the true fate probabilities are

pk⋆(x)=K⋆(x,Bk).(5)p_k^\star(x)=K_\star(x,B_k). \tag*{(5)}

To reconstruct continuous dynamics, we seek a probability law on Ω=C([0,1];Rd)\Omega=C([0,1];\mathbb{R}^d) whose initial-terminal joint distribution is π⋆\pi_\star. Writing et(γ)=γ(t)e_t(\gamma)=\gamma(t) for the evaluation map, the ideal endpoint requirement is

(e0,e1)#P=π⋆.(6)(e_0,e_1)_\#\mathbb{P}=\pi_\star. \tag*{(6)}

The endpoint coupling does not uniquely determine intermediate paths; their construction therefore requires additional dynamical assumptions.

For probability measures P,QP,Q with finite second moments, we can define the Wasserstein distance

W22(P,Q)=inf⁡λ∈Π(P,Q)∫∥a−b∥2λ(da,db),W_2^2(P,Q)=\inf_{\lambda\in\Pi(P,Q)}\int\lVert a-b\rVert^2\lambda(\mathrm{d}a,\mathrm{d}b),

where Π(P,Q)\Pi(P,Q) contains all pairings with marginals PP and QQ, to evaluate the distance between two measures. For a learned path law P^\widehat{P} with initial marginal μ0\mu_0, define

K^0,t(x,A)=P^(et∈A∣e0=x),K^1=K^0,1.\widehat{K}_{0,t}(x,A)=\widehat{P}(e_t\in A\mid e_0=x),\qquad\widehat{K}_1=\widehat{K}_{0,1}.

Assuming that the generated terminal distribution has a finite second moment, we measure the true conditional reconstruction error by

E⋆2(K^1)=∫W22(K⋆(x,⋅),K^1(x,⋅)) μ0(dx)(7)\mathcal{E}_{\star}^2(\widehat{K}_1)=\int W_2^2(K_{\star}(x,\cdot),\widehat{K}_1(x,\cdot))\,\mu_0(\mathrm{d}x) \tag*{(7)}

The Wasserstein triangle inequality and Minkowski’s inequality yield

E⋆(K^1)≤εcpl+Eπ(K^1).\mathcal{E}_{\star}(\widehat{K}_1)\leq\varepsilon_{\mathrm{cpl}}+\mathcal{E}_{\pi}(\widehat{K}_1).

where

Eπ2(K^1)=∫W22(Kπ(x,⋅),K^1(x,⋅)) μ0(dx),εcpl2=∫W22(K⋆(x,⋅),Kπ(x,⋅)) μ0(dx)(8)\mathcal{E}_{\pi}^2(\widehat{K}_1)=\int W_2^2(K_{\pi}(x,\cdot),\widehat{K}_1(x,\cdot))\,\mu_0(\mathrm{d}x),\qquad\varepsilon_{\mathrm{cpl}}^2=\int W_2^2(K_{\star}(x,\cdot),K_{\pi}(x,\cdot))\,\mu_0(\mathrm{d}x) \tag*{(8)}

represent two sources of E⋆\mathcal{E}_{\star}. This discrepancy compares future distributions while holding the initial state fixed. In particular, matching the terminal marginal μ1\mu_1 does not ensure recovery of the true conditional relationship.

Before analyzing training objectives, we examine a representational limitation that persists even if the true coupling were available.

Limitations of deterministic mappings

A deterministic state generator produces an endpoint Y^=T(X)\widehat{Y}=T(X), so its terminal conditional law is K^1(x,⋅)=δT(x)\widehat{K}_1(x,\cdot)=\delta_{T(x)}. This includes deterministic ODE models with unique solutions when the initial condition is fully specified by the observed state xx. Their ability to recover the target coupling is limited by the conditional variability under π⋆\pi_{\star}.

Let

m⋆(x)=Eπ⋆[Y∣X=x],Σ⋆(x)=Cov⁡π⋆(Y∣X=x)m_{\star}(x)=\mathbb{E}_{\pi_{\star}}[Y\mid X=x],\qquad\Sigma_{\star}(x)=\operatorname{Cov}_{\pi_{\star}}(Y\mid X=x)

denote the true conditional mean and covariance.

Proposition 2.1 (Irreducible conditional error of deterministic maps). For every measurable T∈L2(μ0;Rd)T\in L^2(\mu_0;\mathbb{R}^d),

E⋆2(T):=∫W22(K⋆(x,⋅),δT(x)) μ0(dx)\mathcal{E}_{\star}^2(T):=\int W_2^2(K_{\star}(x,\cdot),\delta_{T(x)})\,\mu_0(\mathrm{d}x)
=∫tr⁡Σ⋆(x) μ0(dx)+∫∥m⋆(x)−T(x)∥2 μ0(dx).(9)=\int\operatorname{tr}\Sigma_{\star}(x)\,\mu_0(\mathrm{d}x)+\int\lVert m_{\star}(x)-T(x)\rVert^2\,\mu_0(\mathrm{d}x). \tag*{(9)}

Consequently,

inf⁡T∈L2(μ0;Rd)E⋆2(T)=∫tr⁡Σ⋆(x) μ0(dx),(10)\inf_{T\in L^2(\mu_0;\mathbb{R}^d)} \mathcal{E}^2_{\star}(T)=\int\operatorname{tr}\Sigma_{\star}(x)\,\mu_0(\mathrm{d}x), \tag*{(10)}

with the minimum attained by T=m⋆T=m_{\star}.

Moreover, an exact deterministic realization

π⋆=(Id⁡,T)#μ0\pi_{\star}=(\operatorname{Id},T)_{\#}\mu_0

exists if and only if K⋆(x,⋅)K_{\star}(x,\cdot) is a Dirac measure for μ0\mu_0-almost every xx.

We provide proof of the proposition in SM1. The proposition separates deterministic reconstruction error into the true conditional variability and the error in approximating the true conditional mean. Improving a deterministic model can reduce the second component, but not the first.

Crucially, this limitation is conditional rather than marginal. A deterministic flow can produce a terminal population by sending different initial states to different regions. However, it produces only one endpoint from each fixed initial state. Its fate probabilities under the prescribed partition are therefore

p^k(x)=1Bk(T(x)),\widehat{p}_k(x)=\mathbf{1}_{B_k}(T(x)),

which cannot equal a true probability 0<pk⋆(x)<10<p^{\star}_k(x)<1 without additional randomization.

Motivated by these observations, we introduce a latent representation,

Z∼pη(⋅∣x),Y^=Fθ(x,Z),Z\sim p_{\eta}(\cdot\mid x),\qquad\widehat{Y}=F_{\theta}(x,Z),

whose terminal conditional law is

K^1(x,⋅)=Fθ(x,⋅)#pη(⋅∣x).\widehat{K}_1(x,\cdot)=F_{\theta}(x,\cdot)_{\#}p_{\eta}(\cdot\mid x).

Unlike a deterministic state map, this representation allows multiple future outcomes from the same observed state. The following sections develop this representation into continuous dynamics and relate its training error to the conditional reconstruction objective.

Variational Learning of Latent Cell Dynamics

We now construct a probabilistic dynamical model from the training coupling π\pi. A latent variable represents alternative outcomes from the same observed initial state, while a time-dependent decoder generates the corresponding cell-state trajectories. We first derive a variational reconstruction objective, then introduce flow matching and Mean Flow as simulation-free training strategies. Throughout this section, training expectations are taken under π\pi, whereas the true conditional reconstruction error remains defined relative to π⋆\pi_{\star} as in (7).

Variational formulation

Let Z∈RdzZ\in\mathbb{R}^{d_z} be a latent variable. The generative model consists of a conditional prior pη(dz∣x)p_{\eta}(\mathrm{d}z\mid x) and a decoder Fθ(x,z)F_{\theta}(x,z), giving

Z∼pη(⋅∣x),Y^=Fθ(x,Z).(11)Z\sim p_{\eta}(\cdot\mid x),\qquad\widehat{Y}=F_{\theta}(x,Z). \tag*{(11)}

Its terminal conditional distribution is

K^1(x,⋅)=Fθ(x,⋅)#pη(⋅∣x).(12)\widehat{K}_1(x,\cdot)=F_\theta(x,\cdot)_{\#}p_\eta(\cdot\mid x). \tag*{(12)}

During training, an encoder qϕ(dz∣x,y)q_\phi(\mathrm{d}z\mid x,y) uses both endpoints to infer latent representations compatible with the observed pair. During inference, the terminal state is unavailable, and latent variables are drawn from pη(⋅∣x)p_\eta(\cdot\mid x).

We parameterize the encoder and conditional prior by diagonal Gaussian distributions,

qϕ(⋅∣x,y)=N(mϕ(x,y),diag⁡(sϕ2(x,y))),pη(⋅∣x)=N(mη(x),diag⁡(sη2(x))),(13)\begin{aligned} q_\phi(\cdot\mid x,y)&=\mathcal{N}\left(m_\phi(x,y),\operatorname{diag}(s_\phi^2(x,y))\right),\\ p_\eta(\cdot\mid x)&=\mathcal{N}\left(m_\eta(x),\operatorname{diag}(s_\eta^2(x))\right), \tag*{(13)} \end{aligned}

with strictly positive standard deviations. The means and standard deviations are outputs of neural networks. Define the training joint distribution

Qϕ(dx,dy,dz)=π(dx,dy)qϕ(dz∣x,y).(14)Q_\phi(\mathrm{d}x,\mathrm{d}y,\mathrm{d}z)=\pi(\mathrm{d}x,\mathrm{d}y)q_\phi(\mathrm{d}z\mid x,y). \tag*{(14)}

To derive a variational objective, introduce the auxiliary observation model

pθτ(y∣x,z)=N(y;Fθ(x,z),τ2Id),τ>0,(15)p_\theta^\tau(y\mid x,z)=\mathcal{N}\left(y;F_\theta(x,z),\tau^2 I_d\right),\qquad\tau>0, \tag*{(15)}

where τ\tau is fixed. Its conditional likelihood is

pθ,ητ(y∣x)=∫pθτ(y∣x,z)pη(z∣x) dz.p_{\theta,\eta}^\tau(y\mid x)=\int p_\theta^\tau(y\mid x,z)p_\eta(z\mid x)\,\mathrm{d}z.

Following variational inference [7] [16], Jensen’s inequality gives

log⁡pθ,ητ(y∣x)=log⁡Eqϕ(⋅∣x,y)[pθτ(y∣x,Z)pη(Z∣x)qϕ(Z∣x,y)]≥Eqϕ(⋅∣x,y)log⁡pθτ(y∣x,Z)−DKL(qϕ(⋅∣x,y)∥pη(⋅∣x)).(16)\begin{aligned} \log p_{\theta,\eta}^\tau(y\mid x) &=\log\mathbb{E}_{q_\phi(\cdot\mid x,y)}\left[\frac{p_\theta^\tau(y\mid x,Z)p_\eta(Z\mid x)}{q_\phi(Z\mid x,y)}\right]\\ &\geq\mathbb{E}_{q_\phi(\cdot\mid x,y)}\log p_\theta^\tau(y\mid x,Z)-D_{\mathrm{KL}}\left(q_\phi(\cdot\mid x,y)\mathbin{\|}p_\eta(\cdot\mid x)\right). \tag*{(16)} \end{aligned}

Define the reconstruction error and prior-matching penalty by

Dπ(θ,ϕ)=EQϕ∥Y−Fθ(X,Z)∥2,RKL(ϕ,η)=EπDKL(qϕ(⋅∣X,Y)∥pη(⋅∣X)).(17)\begin{aligned} D_\pi(\theta,\phi)&=\mathbb{E}_{Q_\phi}\left\|Y-F_\theta(X,Z)\right\|^2,\\ R_{\mathrm{KL}}(\phi,\eta)&=\mathbb{E}_\pi D_{\mathrm{KL}}\left(q_\phi(\cdot\mid X,Y)\mathbin{\|}p_\eta(\cdot\mid X)\right). \tag*{(17)} \end{aligned}

Substituting the Gaussian observation model into (16) and averaging over π\pi yields

−Eπlog⁡pθ,ητ(Y∣X)≤d2log⁡(2πτ2)+Dπ(θ,ϕ)2τ2+RKL(ϕ,η).(18)-\mathbb{E}_\pi\log p_{\theta,\eta}^\tau(Y\mid X)\leq\frac{d}{2}\log(2\pi\tau^2)+\frac{D_\pi(\theta,\phi)}{2\tau^2}+R_{\mathrm{KL}}(\phi,\eta). \tag*{(18)}

Dropping the parameter-independent constant gives the negative evidence lower bound objective

LVI=12τ2Dπ+RKL.(19)\mathcal{L}_{\mathrm{VI}}=\frac{1}{2\tau^2}D_\pi+R_{\mathrm{KL}}. \tag*{(19)}

The Gaussian model in (15) provides a likelihood-based interpretation of the reconstruction objective. The generator considered below remains (11), without additional observation noise.

Direct evaluation of DπD_\pi can be expensive when FθF_\theta is defined through a differential equation. We therefore seek simulation-free objectives that control this reconstruction error without solving the learned dynamics during training.

Flow matching and Mean Flow decoders

Both training strategies use endpoint pairs drawn from π\pi and latent variables drawn from the encoder. For (X,Y,Z)∼Qϕ(X,Y,Z) \sim Q_{\phi}, define the reference interpolation

Xt=(1−t)X+tY,V=Y−X.(20)X_t = (1-t)X + tY,\qquad V = Y - X. \tag*{(20)}

The latent variable is held fixed along each reference path. The interpolation specifies a training path between the endpoints, and is not assumed to be an observed biological trajectory.

Flow matching. Flow matching trains a time-dependent velocity field by regression against velocities of prescribed reference paths, avoiding numerical integration during training [10, 17]. In our latent formulation, the decoder is defined by

ddtϕθ,tz(x)=vθ(ϕθ,tz(x),z,t),ϕθ,0z(x)=x,(21)\frac{\mathrm{d}}{\mathrm{d}t}\phi^z_{\theta,t}(x) = v_{\theta}\bigl(\phi^z_{\theta,t}(x), z, t\bigr),\qquad\phi^z_{\theta,0}(x)=x, \tag*{(21)}

with terminal map

FθFM(x,z)=ϕθ,1z(x).F^{\mathrm{FM}}_{\theta}(x,z)=\phi^z_{\theta,1}(x).

The conditional flow-matching objective is

LFM(θ,ϕ)=E(X,Y,Z)∼Qϕ Et∼U[0,1]∥vθ(Xt,Z,t)−V∥2.(22)\mathcal{L}_{\mathrm{FM}}(\theta,\phi)=\mathbb{E}_{(X,Y,Z)\sim Q_{\phi}} \, \mathbb{E}_{t\sim U[0,1]}\left\|v_{\theta}(X_t,Z,t)-V\right\|^2. \tag*{(22)}

Its inputs and regression targets are available directly from (20). For a fixed encoder, the unrestricted squared-loss minimizer is

vˉϕ(ξ,z,t)=EQϕ[V∣Xt=ξ,Z=z].\bar{v}_{\phi}(\xi,z,t)=\mathbb{E}_{Q_{\phi}}[V\mid X_t=\xi,Z=z].

Conditioning on ZZ allows the regression to distinguish different outcomes that may overlap in the observed state space.

The FM loss is evaluated along the prescribed reference paths, whereas generation follows the learned ODE. Under suitable regularity assumptions, the former controls the endpoint reconstruction error of the latter. See Theorem 4.1 in § 4.

Mean Flow. Mean Flow replaces instantaneous velocity with an interval-average velocity, allowing direct prediction of finite-time displacements [4]. Using a forward-time convention, consider a trajectory satisfying x˙s=v(xs,z,s)\dot{x}_s=v(x_s,z,s) and define

u(xr,z;r,t)=1t−r∫rtv(xs,z,s) ds=xt−xrt−r,r<t.(23)u(x_r,z;r,t)=\frac{1}{t-r}\int_r^t v(x_s,z,s)\,\mathrm{d}s=\frac{x_t-x_r}{t-r},\qquad r<t. \tag*{(23)}

Consequently,

xt=xr+(t−r)u(xr,z;r,t).x_t=x_r+(t-r)u(x_r,z;r,t).

Differentiating along the trajectory with respect to the starting time rr gives

u(xr,z;r,t)=v(xr,z,r)+(t−r)[∂ru+Dxu v(xr,z,r)],(24)u(x_r,z;r,t)=v(x_r,z,r)+(t-r)\left[\partial_r u+D_xu\,v(x_r,z,r)\right], \tag*{(24)}

where all derivatives on the right are evaluated at (xr,z;r,t)(x_r,z;r,t).

We parameterize the average velocity by uθu_\theta and use the reference velocity VV to construct the training target. Define the directional derivative

Druθ=∂ruθ(Xr,Z;r,t)+Dxuθ(Xr,Z;r,t)V(25)D_r u_\theta= \partial_r u_\theta(X_r,Z;r,t) + D_x u_\theta(X_r,Z;r,t)V \tag*{(25)}

and the target

Tθ=V+(t−r)Druθ.\mathcal{T}_\theta= V + (t-r)D_r u_\theta.

The Mean Flow loss is

LMF(θ,ϕ)=E(X,Y,Z)∼Qϕt∼U[0,1], r∣t∼U[0,t]∥uθ(Xr,Z;r,t)−sg⁡(Tθ)∥2(26)\mathcal{L}_{\mathrm{MF}}(\theta,\phi) = \mathbb{E}_{(X,Y,Z)\sim Q_\phi\atop t\sim U[0,1],\ r|t\sim U[0,t]}\left\|u_\theta(X_r,Z;r,t)-\operatorname{sg}(\mathcal{T}_\theta)\right\|^2 \tag*{(26)}

where sg⁡\operatorname{sg} denotes the stop-gradient operation. The directional derivative is computed by a Jacobian vector product, without constructing a full Jacobian. Detaching the target avoids differentiating through this derivative when updating the model.

The numerical value of (26) is the squared residual of the reference-path identity:

LMF=E∥uθ(Xr,Z;r,t)−V−(t−r)Druθ∥2.\mathcal{L}_{\mathrm{MF}} = \mathbb{E}\left\|u_\theta(X_r,Z;r,t)-V-(t-r)D_r u_\theta\right\|^2.

This observation allows its value to be related to reconstruction error independently of the stop-gradient update rule.

The learned conditional path and endpoint map are

Gθ,t(x,z)=x+t uθ(x,z;0,t),FθMF(x,z)=Gθ,1(x,z).(27)G_{\theta,t}(x,z)=x+t\,u_\theta(x,z;0,t),\qquad F^{\mathrm{MF}}_\theta(x,z)=G_{\theta,1}(x,z). \tag*{(27)}

We use the Mean Flow loss as a simulation-free surrogate for the variational reconstruction term. Its control of the one-step endpoint error is established in Theorem 4.2. Thus, the resulting training objective is motivated by the variational formulation, but is not algebraically identical to the negative evidence lower bound.

Training and sampling

We use the Mean Flow decoder for direct conditional generation. In addition to the prior-matching penalty, we regularize the encoder against a standard Gaussian:

R0(ϕ)=EπDKL(qϕ(⋅∣X,Y)∥N(0,Idz)).(28)\mathcal{R}_0(\phi)=\mathbb{E}_{\pi}D_{\mathrm{KL}}\left(q_\phi(\cdot\mid X,Y)\middle\|\mathcal{N}(0,I_{d_z})\right). \tag*{(28)}

The full training objective is

LLIFT(θ,ϕ,η)=C LMF(θ,ϕ)+RKL(ϕ,η)+β R0(ϕ),(29)\mathcal{L}_{\mathrm{LIFT}}(\theta,\phi,\eta)=C\,\mathcal{L}_{\mathrm{MF}}(\theta,\phi)+\mathcal{R}_{\mathrm{KL}}(\phi,\eta)+\beta\,\mathcal{R}_0(\phi), \tag*{(29)}

where C>0C>0 and β≥0\beta\geq0. The KL terms are evaluated analytically for the Gaussian parameterization, while the Mean Flow term is estimated from sampled endpoint pairs, latent variables, and times.

Latent samples are obtained by reparameterization:

Z=mϕ(X,Y)+sϕ(X,Y)⊙ϵ,ϵ∼N(0,Idz).(30)Z=m_\phi(X,Y)+s_\phi(X,Y)\odot\epsilon,\qquad\epsilon\sim\mathcal{N}(0,I_{d_z}). \tag*{(30)}

Algorithm 3.1 LIFT training

Require: Training coupling π\pi, batch size BB, weights CC, β\beta.

  1. Initialize encoder ϕ\phi, conditional prior η\eta, and Mean Flow decoder θ\theta.

  2. while the training budget is not exhausted do

  3. Sample {(xb,yb)}b=1B\{(x_b,y_b)\}_{b=1}^{B} independently from π\pi.

  4. for b=1,…,Bb=1,\ldots,B do

  5. Evaluate qb=qϕ(⋅∣xb,yb)q_b=q_\phi(\cdot\mid x_b,y_b) and pb=pη(⋅∣xb)p_b=p_\eta(\cdot\mid x_b).

  6. Sample ϵb∼N(0,Idz)\epsilon_b\sim\mathcal{N}(0,I_{d_z}) and set zb=mϕ(xb,yb)+sϕ(xb,yb)⊙ϵbz_b=m_\phi(x_b,y_b)+s_\phi(x_b,y_b)\odot\epsilon_b.

  7. Sample tb∼U[0,1]t_b\sim U[0,1] and rb∼U[0,tb]r_b\sim U[0,t_b].

  8. Set xrb=(1−rb)xb+rbybx_{r_b}=(1-r_b)x_b+r_by_b and Vb=yb−xbV_b=y_b-x_b.

  9. Evaluate Ub=uθ(xrb,zb;rb,tb)U_b=u_\theta(x_{r_b},z_b;r_b,t_b).

  10. Compute Jb=∂ruθ+DxuθVbJ_b=\partial_ru_\theta+D_xu_\theta V_b at (xrb,zb;rb,tb)(x_{r_b},z_b;r_b,t_b) using a Jacobian–vector product.

  11. Set Tb=sg⁡(Vb+(tb−rb)Jb)T_b=\operatorname{sg}(V_b+(t_b-r_b)J_b).

  12. Compute ℓb=C∥Ub−Tb∥2+DKL(qb∥pb)+βDKL(qb∥N(0,Idz))\ell_b=C\lVert U_b-T_b\rVert^2+D_{\mathrm{KL}}(q_b\lVert p_b)+\beta D_{\mathrm{KL}}(q_b\lVert\mathcal{N}(0,I_{d_z})).

  13. end for

  14. Update (θ,ϕ,η)(\theta,\phi,\eta) using B−1∑b=1BℓbB^{-1}\sum_{b=1}^{B}\ell_b and the specified stop-gradient rule.

  15. end while

  16. return Trained parameters (θ,ϕ,η)(\theta,\phi,\eta).

which permits joint updates of the encoder, prior, and decoder. Gradients through ZZ are retained in the direct decoder evaluation; the Mean Flow regression target is detached as in (26).

Endpoint pairs are sampled according to the prescribed coupling π\pi. When π=∑iπi\pi=\sum_i\pi^i is lineage-compatible, this can be implemented by sampling barcode ii with probability wiw_i and then sampling a pair from πi/wi\pi^i/w_i, for wi>0w_i>0. Thus, barcode information enters through the training coupling without requiring barcode labels as network inputs.

At inference, only the conditional prior and Mean Flow decoder are required. Given an initial state xx, draw Z∼pη(⋅∣x)Z\sim p_\eta(\cdot\mid x) and evaluate

X^t=x+tuθ(x,Z;0,t).(31)\widehat{X}_t=x+tu_\theta(x,Z;0,t). \tag*{(31)}

The same latent sample is used at all queried times to represent one realization of a path. A different latent sample produces another possible realization from the same initial state.

If uθu_\theta is continuous in time, the map t↦X^tt\mapsto\widehat{X}_t is continuous and satisfies X^0=x\widehat{X}_0=x. This construction defines a conditional path law. It does not require an independently trained instantaneous velocity field or numerical ODE integration.

After evaluating the conditional prior, sampling requires one decoder evaluation per realization and requested time point. For a terminal fate region BkB_k, its model probability can be estimated by

p^k(N)(x)=1N∑n=1N1Bk(X^1(n)).(32)\widehat{p}^{(N)}_k(x)=\frac{1}{N}\sum_{n=1}^{N}\mathbf{1}_{B_k}\left(\widehat{X}^{(n)}_1\right). \tag*{(32)}

This estimates K^1(x,Bk)\widehat{K}_1(x,B_k); increasing NN reduces Monte Carlo error but does not remove coupling or model approximation errors.

The resulting model is trained using the available coupling π\pi and sampled through the learned conditional prior. Section 4 relates its training losses to reconstruction under the unknown coupling π⋆\pi_{\star} and to the accuracy of the resulting fate-probability estimates.

Error Analysis of Conditional Dynamics and Fate Prediction

We analyze how the training objectives control conditional reconstruction and fate prediction. The analysis proceeds in three steps. First, the FM and MF losses control reconstruction under the training coupling π\pi. Second, reconstruction and latent distribution matching are related to the true conditional error under π⋆\pi_{\star}. Finally, this error is translated into fate-probability error and combined with the uncertainty of finite Monte Carlo sampling.

Throughout this section, the model parameters are fixed, and all training losses denote expectations under the prescribed training distribution rather than individual minibatch estimates. The couplings π\pi and π⋆\pi_{\star} have the same marginals μ0\mu_0 and μ1\mu_1. We retain the notation

Qϕ(dx,dy,dz)=π(dx,dy)qϕ(dz∣x,y)Q_{\phi}(dx,dy,dz) = \pi(dx,dy)q_{\phi}(dz \mid x,y)

from Section 3. All proofs of the results in this section are provided in SM2.

Reconstruction guarantees for flow matching and Mean Flow

For either decoder, the training reconstruction error is

Dπ=EQϕ∥Y−Fθ(X,Z)∥2.(33)D_{\pi} = \mathbb{E}_{Q_{\phi}}\lVert Y - F_{\theta}(X,Z)\rVert^2. \tag*{(33)}

Recall that the training reference path is

Xt=(1−t)X+tY,V=Y−X.X_t = (1-t)X+tY,\qquad V=Y-X.

The following results connect the regression losses evaluated along this reference path to the endpoint produced by the learned decoder.

Flow matching. For the FM decoder, let

ddtX^t=vθ(X^t,Z,t),X^0=X,\frac{d}{dt}\widehat{X}_t = v_{\theta}(\widehat{X}_t,Z,t),\qquad\widehat{X}_0=X,

and set FθFM(X,Z)=X^1F^{\mathrm{FM}}_\theta(X,Z)=\widehat{X}_1. Its training loss is

LFM=EQϕ∫01∥vθ(Xt,Z,t)−V∥2 dt.\mathcal{L}_{\mathrm{FM}}=\mathbb{E}_{Q_\phi}\int_0^1\left\|v_\theta(X_t,Z,t)-V\right\|^2\,\mathrm{d}t.

Although this loss is evaluated on the reference interpolation, the reconstruction error concerns the solution of the learned ODE. A stability estimate bridges this difference.

Theorem 4.1 (Reconstruction control by flow matching). Assume that vθv_\theta is measurable in time and globally LxL_x-Lipschitz in its state argument, uniformly in zz and tt. Assume also that t↦vθ(0,z,t)t\mapsto v_\theta(0,z,t) is integrable on [0,1][0,1] for almost every latent variable under the training law, so that the generated ODE has a unique solution. If LFM<∞\mathcal{L}_{\mathrm{FM}}<\infty, then

DπFM:=EQϕ∥Y−FθFM(X,Z)∥2≤CLxLFM,(34)D^{\mathrm{FM}}_\pi:=\mathbb{E}_{Q_\phi}\left\|Y-F^{\mathrm{FM}}_\theta(X,Z)\right\|^2\le C_{L_x}\mathcal{L}_{\mathrm{FM}}, \tag*{(34)}

where

CLx={e2Lx−12Lx,Lx>0,1,Lx=0.(35)C_{L_x}= \begin{cases} \frac{e^{2L_x}-1}{2L_x}, & L_x>0,\\ 1, & L_x=0. \end{cases} \tag*{(35)}

The estimate accounts for the accumulation of local regression errors along the generated trajectory. It concerns the exact ODE solution; a numerical implementation introduces an additional integration error.

Mean Flow. For the Mean Flow decoder,

FθMF(x,z)=x+uθ(x,z;0,1).F^{\mathrm{MF}}_\theta(x,z)=x+u_\theta(x,z;0,1).

Define the residual along the reference path by

Rr,t=uθ(Xr,Z;r,t)−V−(t−r)[∂ruθ(Xr,Z;r,t)+Dxuθ(Xr,Z;r,t)V].(36)\mathcal{R}_{r,t}=u_\theta(X_r,Z;r,t)-V-(t-r)\left[\partial_r u_\theta(X_r,Z;r,t)+D_xu_\theta(X_r,Z;r,t)V\right]. \tag*{(36)}

The numerical value of the Mean Flow loss is

LMF=E(X,Y,Z)∼Qϕt∼U[0,1], r∣t∼U[0,t]∥Rr,t∥2.(37)\mathcal{L}_{\mathrm{MF}}=\mathbb{E}_{\substack{(X,Y,Z)\sim Q_\phi\\ t\sim U[0,1],\ r|t\sim U[0,t]}}\left\|\mathcal{R}_{r,t}\right\|^2. \tag*{(37)}

The stop-gradient operation used in training changes the parameter update but not this residual value.

Theorem 4.2 (Reconstruction control by Mean Flow). Assume that, for QϕQ_\phi-almost every (X,Y,Z)(X,Y,Z) and every t∈(0,1]t\in(0,1], the map

r↦uθ(Xr,Z;r,t)r\mapsto u_\theta(X_r,Z;r,t)

is absolutely continuous on [0,t][0,t], with derivative

∂ruθ(Xr,Z;r,t)+Dxuθ(Xr,Z;r,t)V\partial_r u_\theta(X_r,Z;r,t)+D_xu_\theta(X_r,Z;r,t)V

almost everywhere.

Assume further that t↦uθ(X,Z;0,t)t \mapsto u_\theta(X,Z;0,t) is Lipschitz as a map into L2(Qϕ;Rd)L^2(Q_\phi;\mathbb{R}^d):

∥uθ(X,Z;0,t)−uθ(X,Z;0,s)∥L2(Qϕ)≤LT∣t−s∣,s,t∈[0,1].(38)\left\|u_\theta(X,Z;0,t)-u_\theta(X,Z;0,s)\right\|_{L^2(Q_\phi)} \le L_T|t-s|,\qquad s,t\in[0,1]. \tag*{(38)}

If LMF<∞L_{\mathrm{MF}}<\infty, then

DπMF:=EQϕ∥Y−FθMF(X,Z)∥2≤max⁡{3LMF,(3LTLMF)2/3}.(39)D_\pi^{\mathrm{MF}}:=\mathbb{E}_{Q_\phi}\left\|Y-F_\theta^{\mathrm{MF}}(X,Z)\right\|^2\le\max\left\{3L_{\mathrm{MF}},(3L_TL_{\mathrm{MF}})^{2/3}\right\}. \tag*{(39)}

For a unified notation, define

ΨFM(a)=CLxa,ΨMF(a)=max⁡{3a,(3LTa)2/3}.(40)\Psi_{\mathrm{FM}}(a)=C_{L_x}a,\qquad\Psi_{\mathrm{MF}}(a)=\max\left\{3a,(3L_Ta)^{2/3}\right\}. \tag*{(40)}

Then, for either decoder dec∈{FM,MF}\mathrm{dec}\in\{\mathrm{FM},\mathrm{MF}\},

Dπdec≤Ψdec(Ldec).(41)D_\pi^{\mathrm{dec}}\le\Psi_{\mathrm{dec}}(L_{\mathrm{dec}}). \tag*{(41)}

These estimates control endpoint reconstruction under π\pi. They do not by themselves establish fidelity to the unknown coupling π⋆\pi_\star, nor do they identify the true intermediate biological paths. The next result addresses the former question.

Total conditional reconstruction error

There are two additional discrepancies between training reconstruction and the desired conditional generation. First, the training coupling π\pi may differ from π⋆\pi_\star. Second, training uses the encoder qϕ(⋅∣x,y)q_\phi(\cdot\mid x,y), whereas sampling uses the conditional prior pη(⋅∣x)p_\eta(\cdot\mid x).

Recall the conditional decompositions

π⋆(dx,dy)=μ0(dx)K⋆(x,dy),π(dx,dy)=μ0(dx)Kπ(x,dy),\pi_\star(dx,dy)=\mu_0(dx)K_\star(x,dy),\qquad\pi(dx,dy)=\mu_0(dx)K_\pi(x,dy),

and write the conditional coupling discrepancy as

εcpl2=∫W22(K⋆(x,⋅),Kπ(x,⋅)) μ0(dx).(42)\varepsilon_{\mathrm{cpl}}^2=\int W_2^2\left(K_\star(x,\cdot),K_\pi(x,\cdot)\right)\,\mu_0(dx). \tag*{(42)}

To describe the training latent distribution conditioned only on the initial state, define the aggregated posterior

q‾ϕ(dz∣x)=∫qϕ(dz∣x,y)Kπ(x,dy).(43)\overline{q}_\phi(dz\mid x)=\int q_\phi(dz\mid x,y)K_\pi(x,dy). \tag*{(43)}

Its discrepancy from the sampling prior is

Δz2=∫W22(q‾ϕ(⋅∣x),pη(⋅∣x)) μ0(dx).(44)\Delta_z^2=\int W_2^2\left(\overline{q}_\phi(\cdot\mid x),p_\eta(\cdot\mid x)\right)\,\mu_0(dx). \tag*{(44)}

The generated terminal kernel is

K^1(x,⋅)=Fθ(x,⋅)#pη(⋅∣x).\widehat{K}_1(x,\cdot)=F_\theta(x,\cdot)_{\#}p_\eta(\cdot\mid x).

Its true conditional error is

ε⋆2(K^1)=∫W22(K⋆(x,⋅),K^1(x,⋅)) μ0(dx).\varepsilon_\star^2(\widehat{K}_1)=\int W_2^2\left(K_\star(x,\cdot),\widehat{K}_1(x,\cdot)\right)\,\mu_0(dx).

Theorem 4.3 (Total conditional reconstruction error). Assume that

∥Fθ(x,z)−Fθ(x,z′)∥≤LZ∥z−z′∥\lVert F_\theta(x,z)-F_\theta(x,z')\rVert\le L_Z\lVert z-z'\rVert

for μ0\mu_0-almost every xx and all z,z′z,z'. Assume that the required second moments and the quantities below are finite. Then

E⋆(K^1)≤εcpl+Dπ+LZΔZ.(45)\mathcal{E}_\star(\widehat{K}_1) \le\varepsilon_{\mathrm{cpl}}+\sqrt{D_\pi}+L_Z\Delta_Z. \tag*{(45)}

Suppose additionally that the encoder and prior are diagonal Gaussians as in (13), and that their prior standard deviations satisfy

0<sη,j(x)≤S0 < s_{\eta,j}(x) \le S

for every latent coordinate jj and μ0\mu_0-almost every xx. Then

ΔZ2≤2S2RKL,(46)\Delta_Z^2 \le2S^2\mathcal{R}_{\mathrm{KL}}, \tag*{(46)}

where

RKL=EπDKL(qϕ(⋅∣X,Y)∥pη(⋅∣X)).\mathcal{R}_{\mathrm{KL}}=\mathbb{E}_\pi D_{\mathrm{KL}}\left(q_\phi(\cdot\mid X,Y)\mathbin\Vert p_\eta(\cdot\mid X)\right).

Consequently, under the assumptions of the corresponding reconstruction theorem,

E⋆(K^1)≤εcpl+Ψdec(Ldec)+2LZSRKL.(47)\mathcal{E}_\star(\widehat{K}_1) \le\varepsilon_{\mathrm{cpl}}+\sqrt{\Psi_{\mathrm{dec}}(\mathcal{L}_{\mathrm{dec}})}+\sqrt{2}L_ZS\sqrt{\mathcal{R}_{\mathrm{KL}}}. \tag*{(47)}

The three terms in (45) have distinct interpretations. The coupling discrepancy measures the difference between the training relationships and the unknown target. The reconstruction term measures how accurately the decoder reproduces training endpoints from encoder samples. The final term measures the effect of replacing the training latent distribution with the prior used for generation.

These are conditional error guarantees for a fitted model. For later use, denote the upper bound in (47) by

Bdec=εcpl+Ψdec(Ldec)+2LZSRKL.(48)\mathcal{B}_{\mathrm{dec}}=\varepsilon_{\mathrm{cpl}}+\sqrt{\Psi_{\mathrm{dec}}(\mathcal{L}_{\mathrm{dec}})}+\sqrt{2}L_ZS\sqrt{\mathcal{R}_{\mathrm{KL}}}. \tag*{(48)}

Fate-probability estimation

Let {B1,…,BK}\{B_1,\ldots,B_K\} be the fixed partition of terminal state space into fates. The true and model fate probabilities are

pk⋆(x)=K⋆(x,Bk),p^k(x)=K^1(x,Bk).(49)p_k^\star(x)=K_\star(x,B_k),\qquad\widehat{p}_k(x)=\widehat{K}_1(x,B_k). \tag*{(49)}

Estimating pk⋆(x)p_k^\star(x) by generated samples involves two different errors: the discrepancy between p^k(x)\widehat{p}_k(x) and pk⋆(x)p_k^\star(x), and the finite-sampling error in estimating p^k(x)\widehat{p}_k(x).

From conditional distributions to fate probabilities. A small displacement can change a hard fate label when the terminal state lies near a classification boundary. Define

Γ=⋃k=1K∂Bk,Γρ={y:dist⁡(y,Γ)≤ρ},ρ>0.\Gamma=\bigcup_{k=1}^{K}\partial B_k,\qquad\Gamma_\rho=\{y:\operatorname{dist}(y,\Gamma)\le\rho\},\qquad\rho>0.

The next bound separates this boundary effect from the endpoint reconstruction error.

Proposition 4.4 (Stability of fate probabilities). For every ρ>0\rho> 0,

max⁡1≤k≤K∥pk⋆−p^k∥L1(μ0)≤μ1(Γρ)+E⋆2(K^1)ρ2≤μ1(Γρ)+Bdec2ρ2,(50)\max_{1\le k\le K}\left\|p_k^\star-\widehat{p}_k\right\|_{L^1(\mu_0)} \le\mu_1(\Gamma_\rho)+\frac{\mathcal{E}_\star^2(\widehat{K}_1)}{\rho^2} \le\mu_1(\Gamma_\rho)+\frac{\mathcal{B}_{\mathrm{dec}}^2}{\rho^2}, \tag*{(50)}

where the second inequality holds under the assumptions of Theorem 4.3.

The estimate follows the geometry of a label change: a coupled pair of endpoints can have different fates only if the true endpoint lies near a boundary or the displacement between the endpoints is sufficiently large. The radius ρ\rho balances these two possibilities. In particular, if μ1(Γ)=0\mu_1(\Gamma)=0, vanishing conditional reconstruction error implies vanishing average fate-probability error. Furthermore, we combine the results above with standard results of Monte-Carlo sampling[6], we can obtain the following end-to-end fate probability error estimation.

Corollary 4.5 (End-to-end fate-probability error). Suppose

μ0=∑a=1nωaδxa,ωa>0,∑a=1nωa=1.\mu_0=\sum_{a=1}^{n}\omega_a\delta_{x_a},\qquad\omega_a>0,\qquad\sum_{a=1}^{n}\omega_a=1.

Under the assumptions of Theorem 4.3, with probability at least 1−δ1-\delta over the Monte Carlo samples,

max⁡1≤k≤K∑a=1nωa∣p^k(N)(xa)−pk⋆(xa)∣≤inf⁡ρ>0{μ1(Γρ)+Bdec2ρ2}+log⁡(2nK/δ)2N.(51)\max_{1\le k\le K}\sum_{a=1}^{n}\omega_a\left|\widehat{p}_k^{(N)}(x_a)-p_k^\star(x_a)\right| \le\inf_{\rho>0}\left\{\mu_1(\Gamma_\rho)+\frac{\mathcal{B}_{\mathrm{dec}}^2}{\rho^2}\right\}+\sqrt{\frac{\log(2nK/\delta)}{2N}}. \tag*{(51)}

Experimental Validation and Lineage-Informed Applications

We evaluate LIFT on synthetic and real-world datasets. We first use a synthetic branching benchmark to assess its ability to reconstruct population distributions and branching probabilities. We then examine whether one-step generation combines computational efficiency with accurate conditional prediction. Finally, we evaluate how LIFT incorporates lineage information to predict cell fate probabilities in two real datasets. Across all experiments, we compare LIFT with baselines including conditional flow matching (CFM), Mean Flow (MF) and score matching (SM, here we use an implementation of SF2M[19]) to analyze the ability of LIFT. Detailed experimental settings, model-component ablations, hyperparameter sensitivity analyses, and additional results are provided in Supplementary Section SM3.

Recovering conditional fate distributions in a synthetic dataset

Dataset construction. We sample 1,000 source states xi∼N(0,I2)x_i\sim\mathcal{N}(0,I_2) and assign a branch by

Sij=sign⁡(v⊤xi+αϵij),v=(1,2)⊤,ϵij∼N(0,1),j=1,2,⋯ ,5.(52)S_i^j=\operatorname{sign}(v^\top x_i+\alpha\epsilon_i^j),\qquad v=(1,2)^\top,\qquad\epsilon_i^j\sim\mathcal{N}(0,1),\qquad j=1,2,\cdots,5. \tag*{(52)}

For each source state xix_i, five targets yijy_i^j are generated conditional on its sampled branch SijS_i^j. Specifically, we construct yij=xi+Sij⋅3v+0.3⋅ηijy_i^j=x_i+S_i^j\cdot3v+0.3\cdot\eta_i^j, ηij∼N(0,I2)\eta_i^j\sim\mathcal{N}(0,I_2). Thus, these five targets share the same barcode label, but may have different fate labels. For α>0\alpha> 0, the population probability of the positive branch is

pαbr(x)=Φ(vTxα),(53)p_{\alpha}^{\mathrm{br}}(x) = \Phi\left(\frac{v^{\mathsf{T}}x}{\alpha}\right), \tag*{(53)}

where Φ\Phi is the standard Gaussian distribution function. This analytic reference permits evaluation beyond the particular branch labels realized in the finite training sample.

The principal experiment uses α∈{0.5,1,2}\alpha\in\{0.5, 1, 2\}. Branch probabilities are evaluated at nqn_q uniformly spaced query states along the segment from (−1,−1)(-1,-1) to (1,1)(1,1). We define the probability error convention,

PE=∑a=1nq∣p^+(xa)−pαbr(xa)∣.(54)\mathrm{PE} = \sum_{a=1}^{n_q} \left|\widehat{p}_{+}(x_a) - p_{\alpha}^{\mathrm{br}}(x_a)\right|. \tag*{(54)}

Here we choose nq=21n_q = 21. Population level accuracy is measured by W2W_2 between generated and reference terminal samples.

LIFT recovers both distribution and branching probabilities. Table 1 shows that LIFT can achieve the lowest branching probability error across all tested noise levels while maintaining competitive population reconstruction. In contrast, accurate marginal reconstruction by CFM does not translate into accurate branching probabilities, illustrating the distinction between these two objectives. Note that the CFM and MF probability errors coincide with the minimum error attainable by hard branch assignments on the evaluation grid, because deterministic methods can only provide 0-1 predictions. Figure 1(b) further shows that LIFT successfully captures the conditional branching probabilities, whereas the deterministic baselines produce hard branch assignments.

Methodα=0.5\alpha= 0.5α=1\alpha= 1α=2\alpha= 2
MethodW2W_2PEW2W_2PEW2W_2PE
CFM0.2381.3700.2752.6780.3055.002
Mean Flow0.4701.3701.0112.6781.1015.002
SM0.4630.2820.4781.5000.4603.580
LIFT0.2660.1210.2680.4570.2800.809

Table 1. Results on the synthetic branching benchmark. Lower values are better. PE is the sum in (54).

Conditional reconstruction in the branching system

Figure 1. Conditional reconstruction in the branching system. (a) Source and terminal populations. (b) Predicted branch probabilities along the query segment. (c) Generated endpoint distributions at a fixed initial state as α\alpha varies.

Conditional distributions at a fixed initial state. Figure 1(c) examines the generated endpoint distributions with the initial state held fixed. At low noise, the predicted outcomes are concentrated in one branch. As α\alpha increases, probability mass is distributed across both branches, producing a bimodal conditional distribution. This qualitative transition illustrates that LIFT can capture changes in branching uncertainty.

Efficient one-step conditional generation

We assess the accuracy-cost trade-off on the synthetic branching benchmark by varying the number of function evaluations (NFE) for

CFM and SM. We compare these iterative configurations with one-step MF and LIFT, evaluating population reconstruction, branching probability error, and inference time at several noise levels.

Table 2: Effect of NFE on the synthetic branching benchmark. Inference times are averaged over 100 timing measurements. Lower values are better for W2W_2, PE, and inference time. The best values in each metric column are highlighted in bold.

LIFT balances generation efficiency and conditional accuracy. Table 2 shows that LIFT achieves the lowest branching probability error across all tested noise levels while sampling faster than every iterative configuration considered. Its population reconstruction remains competitive, although CFM with larger integration budgets attains a smaller W2W_2 in the lowest-noise setting. Increasing NFE improves the marginal reconstruction of the iterative baselines but does not consistently improve their branching probabilities; in particular, the reported CFM probability error remains high as it can give only 0-1 prediction. Although unaugmented MF is slightly faster than LIFT, it has larger errors in both metrics. These results show that LIFT combines efficient one-step sampling with accurate branching probabilities. Specifically, lower inference time makes it convenient to predict fate probabilities through large-scale Monte-Carlo sampling.

MethodNFETime (ms)α=0.5\alpha= 0.5 W2W_2α=0.5\alpha= 0.5 PEα=1\alpha= 1 W2W_2α=1\alpha= 1 PEα=2\alpha= 2 W2W_2α=2\alpha= 2 PE
CFM1,000401.090.2381.3700.2752.6780.3055.002
10040.260.2381.3700.2822.6780.3055.002
103.870.2701.3700.3672.6780.3695.002
SM1,000855.930.4630.2820.4781.5000.4603.580
10086.740.5310.3820.5481.5560.5583.619
109.650.6280.6230.6561.5010.7193.834
MF10.560.4701.3701.0112.6781.1015.002
LIFT10.960.2660.1210.2680.4570.2800.809

Table 2. Effect of NFE on the synthetic branching benchmark. Inference times are averaged over 100 timing measurements. Lower values are better for W2W_2, PE, and inference time. The best values in each metric column are highlighted in bold.

Lineage-informed cell fate prediction in real datasets

We evaluate LIFT on two lineage-traced scRNA-seq datasets: mouse hematopoiesis measured by LARRY [22] and direct fibroblast reprogramming measured by CellTagging [1]. The hematopoietic task predicts monocyte (Mo) and neutrophil (Neu) outcomes from Day 2 to Day 6, while the reprogramming task predicts induced endoderm progenitor (iEP) and dead-end outcomes from Day 15 to Day 28. Both datasets use two-dimensional SPRING state representations. We compare geometric and lineage-informed couplings, denoted by Geo and Lin, respectively. The lineage-informed construction uses uniform pairing within each barcode for hematopoiesis and within-clone OT for reprogramming. Population reconstruction is measured by W2W_2, while fate prediction is assessed by mean probability error (MeanPE) and Pearson correlation rr against lineage-derived reference probabilities.

Table 3: Results on the two lineage-traced datasets. Geo and Lin indicate geometric and lineage-informed coupling variants, respectively. Here SM relies on entropic optimal transport coupling as [19], so SM does not have different couplings. Lower W2W_2 and MeanPE and higher Pearson correlation rr are better. The best value in each column is highlighted in bold.

LIFT improves fate prediction with lineage information. Table 3 shows that LIFT-Lin can achieve the lowest MeanPE and highest correlation on both datasets. Compared with LIFT-Geo, incorporating lineage information improves fate-probability prediction even though the population Wasserstein distance does not decrease. Supplying lineage-informed couplings to deterministic CFM or MF does not produce comparable probability accuracy. These results indicate that accurate fate prediction requires pairing lineage-informed couplings with a stochastic generator capable of representing one-to-many transitions.

MethodHematopoiesis W2W_2Hematopoiesis MeanPEHematopoiesis rrReprogramming W2W_2Reprogramming MeanPEReprogramming rr
CFM–Geo0.3010.3360.6571.0130.2880.729
CFM–Lin0.5390.3660.7283.3930.2680.729
MF–Geo0.3700.3430.6431.8460.2610.746
MF–Lin0.5490.3490.6893.9460.3630.701
SM0.1900.2960.7351.1910.2130.763
LIFT–Geo0.2470.2230.7050.9360.2110.727
LIFT–Lin0.2690.1190.8030.9590.1910.795

Table 3. Results on the two lineage-traced datasets. Geo and Lin indicate geometric and lineage-informed coupling variants, respectively. Here SM relies on entropic optimal transport coupling as [19], so SM does not have different couplings. Lower W2W_2 and MeanPE and higher Pearson correlation rr are better. The best value in each column is highlighted in bold.

LIFT captures graded fate bias in hematopoiesis. Figure 2 presents mouse hematopoiesis as a representative example. In the highlighted region, the lineage-derived reference contains mixed Mo and Neu outcomes with a preference toward Mo. Hard assignments replace this mixture with a sharp partition, whereas LIFT retains intermediate probabilities with a modest Mo bias. The resulting probability map therefore captures graded lineage-associated fate bias.

Representative fate prediction in mouse hematopoiesis

Figure 2. Representative fate prediction in mouse hematopoiesis. (a) Lineage-derived reference fate information. (b) Hard fate assignments. (c) Mo fate probabilities predicted by LIFT. The highlighted region contains mixed Mo and Neu clonal outcomes with a bias toward Mo. LIFT represents this region through intermediate fate probabilities rather than exclusively binary assignments.

This qualitative agreement complements the quantitative results across both datasets.

Discussion and Conclusion

Lineage tracing links molecular states to clonal ancestry, but does not generally reveal unique cell-to-cell pairings across destructively sampled populations. This motivates a conditional-generation problem: the same observed state may be compatible with multiple future outcomes, whereas a deterministic state map assigns it a single endpoint. Consequently, reproducing the terminal population distribution alone does not ensure accurate fate probabilities. A model must also represent the one-to-many relationships supported by the available lineage information.

LIFT addresses this mismatch through a variational latent representation of conditional variability. An encoder learns latent distributions from endpoint pairs, a conditional prior predicts them from the initial state, and a Mean Flow decoder generates alternative outcomes without numerical integration. Under the stated regularity and prior-scale assumptions, we bound the true conditional distribution error by the discrepancy in the training coupling, decoder reconstruction error, and latent training-sampling mismatch. The analysis further connects these errors to fate-probability estimation, providing a theoretical basis for the conditional prediction objective.

The experiments show that LIFT combines accurate branching-probability prediction with competitive population distribution reconstruction and efficient one-step sampling. In mouse hematopoiesis and direct reprogramming, lineage-informed training improves fate prediction relative to the geometric-coupling variant. The hematopoietic example further illustrates how graded predictions represent mixed clonal outcomes instead of imposing hard assignments. Together, these results support the use of LIFT for efficient prediction of differentiation and reprogramming probabilities from molecular states and lineage information. The current framework nevertheless has several limitations. Its generative dynamics do not explicitly model unbalanced effects such as proliferation and cell death, and its predictions remain dependent on the supplied coupling and chosen state representation. Endpoint constraints also do not uniquely identify intermediate biological paths or the sources of conditional variability. Future work will extend the framework to unbalanced dynamics, incorporate additional observation times and biological constraints. These extensions could improve both the biological fidelity of generated trajectories and the reliability of fate predictions.

Appendix A. Mathematical and Computational Background

This appendix introduces the concepts connecting barcode-resolved observations, conditional generation, and fate-probability estimation.

Probabilistic modeling of cell dynamics. A probability measure μt\mu_t describes the relative abundance of cell states at time tt. Population evolution can therefore be modeled as a continuous family of distributions connecting μ0\mu_0 and μ1\mu_1, rather than as a single representative trajectory. Waddington-OT infers probabilistic couplings between observed populations, accounting for proliferation and death [14]. TrajectoryNet learns continuous paths using neural ordinary differential equations and optimal-transport regularization [18], while PRESCIENT models population evolution through stochastic dynamics [24]. These approaches connect snapshot observations through assumed dynamics. A coupling additionally specifies how initial and terminal states are paired; lineage information constrains these relationships [3]. The conditional law K⋆(x,⋅)K_\star(x,\cdot) then describes the possible outcomes of a fixed initial state, which are not determined by the endpoint population measures alone.

Variational inference and variational autoencoders. Latent-variable models generate observations through auxiliary variables whose values are not directly measured. Inferring their posterior distribution is often computationally difficult. Variational inference replaces this posterior with a tractable approximation and optimizes a lower bound on the data log likelihood. Variational autoencoders (VAEs) implement this idea using neural encoder and decoder networks [7, 8]. Their objective balances reconstruction against a Kullback–Leibler (KL) divergence penalty, which compares the encoder distribution with the prior. In conditional generation [16], the model also receives an input xx. LIFT uses qϕ(z∣x,y)q_\phi(z\mid x,y) during training and samples pη(z∣x)p_\eta(z\mid x) at inference. Different latent samples allow the same initial state to generate different outcomes. The latent variable represents conditional variability, not an automatically identified biological mechanism.

Flow matching and Mean Flow. Flow matching trains a velocity field by regression against velocities of easily sampled reference paths [10, 17]. For the linear path Xt=(1−t)X+tYX_t=(1-t)X+tY, the reference velocity is Y−XY-X. Training therefore avoids solving the learned differential equation, although generation with an FM decoder still requires numerical integration. Mean Flow instead learns velocity averaged over a time interval [4]. For r<tr<t, its average velocity along a trajectory satisfies

u(xr,z;r,t)=xt−xrt−r,xt=xr+(t−r)u(xr,z;r,t).u(x_r,z;r,t)=\frac{x_t-x_r}{t-r}, \qquad x_t=x_r+(t-r)u(x_r,z;r,t).

Here, the average is over time along a trajectory, not over alternative fates. Learning this quantity allows LIFT to predict each requested state directly, with a latent code held fixed along a sampled path.

Wasserstein distance. Wasserstein distance compares distributions by the displacement required to move probability mass from one to the other [12]. For probability measures P,QP, Q with finite second moments,

W22(P,Q)=inf⁡λ∈Π(P,Q)∫∥a−b∥2λ(da,db),W_2^2(P,Q) = \inf_{\lambda\in\Pi(P,Q)} \int\lVert a-b \rVert^2 \lambda(\mathrm{d}a,\mathrm{d}b),

where Π(P,Q)\Pi(P,Q) contains all pairings with marginals PP and QQ. On finite cell populations, these pairings are nonnegative matrices with prescribed row and column sums. Unlike comparing population means, this distance measures differences in the locations and spread of probability mass. Our conditional error applies it separately at each initial state and averages the squared distances, distinguishing correct future-state distributions from terminal population matching.

Further reading. Biological observations and trajectory inference. Biddy et al. [1] introduce the CellTagging reprogramming application; Weinreb et al. [22] connect hematopoietic state and fate using LARRY. Saelens et al. [13] compare trajectory-inference methods and their assumptions.

Lineage-informed endpoint relationships. LineageOT [3] explains how lineage trees inform transport couplings, while CoSpar [20] combines sparse clonal observations with state-space structure to infer transition maps.

Latent generation and flow-based computation. Kingma and Welling [7] and Sohn et al. [16] develop variational and conditional generative models. Kingma and Welling also provide a tutorial on VAEs and variational training [8]. Lipman et al. [10] introduce flow-matching training, and Geng et al. [4] explain the average-velocity formulation underlying Mean Flow.

Optimal transport: Peyré and Cuturi provide a computational introduction to couplings, Wasserstein distances, and their applications [12].

Finite-sample probability estimates. Hoeffding [6] provides the concentration inequality used to bound the sampling error of the estimated fate probabilities.

Declarations

Authorship and Contributorship. All authors have made substantial intellectual contributions to the study conception, execution, and design of the work. All authors have read and approved the final manuscript. In addition, the following contributions occurred: Conceptualization: B. Yang; Methodology: B. Yang, J. Ying, and Y. Wang; Formal analysis and investigation: B. Yang, J. Ying, and Y. Wang; Writing – original draft preparation: B. Yang; Writing – review and editing: P. Zhou and L. Zhang; Funding acquisition: L. Zhang; Resources: L. Zhang; Supervision: P. Zhou and L. Zhang.

Conflicts of Interest. The authors declare there are no conflicts of interest

Data & Code Availability. The datasets analyzed in this study are derived from previously published lineage-tracing studies, including the LARRY hematopoiesis dataset [22] and the CellTagging direct reprogramming dataset[1]. The original datasets are available from the sources cited in the corresponding publications.

The code used to generate the results presented in this study are publicly available in https://github.com/van-genshin-wander/LIFT.

Ethics. This study involves computational analyses of previously published and publicly available single-cell lineage-tracing datasets. No new experiments involving human participants or animals were conducted as part of this study.

Funding. L. Zhang was supported by the National Natural Science Foundation of China (Nos. 12225102, T2321001, and 12288101) and the National Key Research and Development Program of China (No. 2024YFA0919500).

Use of AI Tools. AI tools were used only to assist with language translation, editing, and polishing during the preparation of this manuscript. All authors have checked and verified the accuracy of all parts of this work and assume responsibility for all content.

SUPPLEMENTARY MATERIALS: LIFT: Latent Flow Matching for Probabilistic Cell Fate Dynamics with Single Cell Snapshots

Bowen Yang; Junda Ying; Yuxuan Wang; Peijie Zhou; Lei Zhang

Proof of Proposition2.1

Proof. There is only one coupling between K⋆(x,⋅)K_{\star}(x,\cdot) and δT(x)\delta_{T(x)}. Hence, by definition,

W22(K⋆(x,⋅),δT(x))=∫∥y−T(x)∥2K⋆(x,dy).W_{2}^{2}\left(K_{\star}(x,\cdot),\delta_{T(x)}\right)=\int\left\|y-T(x)\right\|^{2}K_{\star}(x,\mathrm{d}y).

Integrating over the source marginal yields

E⋆2(T)=Eπ⋆∥Y−T(X)∥2.\mathcal{E}_{\star}^{2}(T)=\mathbb{E}_{\pi_{\star}}\left\|Y-T(X)\right\|^{2}.

Decompose

Y−T(X)=Y−m⋆(X)+m⋆(X)−T(X).Y-T(X)=Y-m_{\star}(X)+m_{\star}(X)-T(X).

The cross term vanishes because

Eπ⋆[Y−m⋆(X)∣X]=0.\mathbb{E}_{\pi_{\star}}\left[Y-m_{\star}(X)\mid X\right]=0.

Therefore,

Eπ⋆∥Y−T(X)∥2=Eπ⋆∥Y−m⋆(X)∥2+Eπ⋆∥m⋆(X)−T(X)∥2.\mathbb{E}_{\pi_{\star}}\left\|Y-T(X)\right\|^{2}=\mathbb{E}_{\pi_{\star}}\left\|Y-m_{\star}(X)\right\|^{2}+\mathbb{E}_{\pi_{\star}}\left\|m_{\star}(X)-T(X)\right\|^{2}.

The first term equals ∫tr⁡Σ⋆(x) μ0(dx)\int\operatorname{tr}\Sigma_{\star}(x)\,\mu_{0}(\mathrm{d}x), proving (9). The second term is nonnegative and vanishes for T=m⋆T=m_{\star}, which proves (10).

Finally, a deterministic realization has conditional law δT(x)\delta_{T(x)} almost everywhere. Conversely, if K⋆(x,⋅)K_{\star}(x,\cdot) is a Dirac measure almost everywhere, its point of concentration is m⋆(x)m_{\star}(x), and

π⋆=(Id⁡,m⋆)#μ0.\pi_{\star}=(\operatorname{Id},m_{\star})_{\#}\mu_{0}.

Proofs of the Error Estimates

We provide the proofs of the results in Section 4. Throughout the appendix, the model parameters are fixed, and

Qϕ(dx,dy,dz)=π(dx,dy)qϕ(dz∣x,y).Q_{\phi}(\mathrm{d}x,\mathrm{d}y,\mathrm{d}z)=\pi(\mathrm{d}x,\mathrm{d}y)q_{\phi}(\mathrm{d}z\mid x,y).

All statements about conditional distributions are understood to hold for μ0\mu_{0}-almost every initial state. We write

Xt=(1−t)X+tY,V=Y−X,X_{t}=(1-t)X+tY,\qquad V=Y-X,

for the reference path and its velocity. Norms without a subscript are Euclidean norms. The finite second moments of the endpoint measures imply that V∈L2(Qϕ;Rd)V\in L^{2}(Q_{\phi};\mathbb{R}^{d}).

Proof of Theorem 4.1

Proof. Fix a triple (X,Y,Z)(X,Y,Z) for which the generated ODE has a solution and the reference-path regression residual is square-integrable in time. These properties hold QϕQ_{\phi}-almost surely under the theorem’s assumptions. Let

X^˙t=vθ(X^t,Z,t),X^0=X,\dot{\widehat{X}}_t = v_{\theta}(\widehat{X}_t,Z,t), \qquad\widehat{X}_0 = X,

and define

et=X^t−Xt,at=vθ(Xt,Z,t)−V.e_t = \widehat{X}_t - X_t, \qquad a_t = v_{\theta}(X_t,Z,t) - V.

The difference between the generated and reference paths satisfies

e˙t=vθ(X^t,Z,t)−vθ(Xt,Z,t)+at,e0=0.\dot{e}_t = v_{\theta}(\widehat{X}_t,Z,t) - v_{\theta}(X_t,Z,t) + a_t, \qquad e_0 = 0.

The state-space Lipschitz assumption gives the integral inequality

∥et∥≤∫0t(LX∥es∥+∥as∥) ds.\lVert e_t\rVert\le\int_0^t \left(L_X\lVert e_s\rVert+ \lVert a_s\rVert\right)\,\mathrm{d}s.

Gronwall’s inequality therefore yields

∥et∥≤∫0teLX(t−s)∥as∥ ds(55)\lVert e_t\rVert\le\int_0^t e^{L_X(t-s)}\lVert a_s\rVert\,\mathrm{d}s \tag*{(55)}

At t=1t=1, the Cauchy–Schwarz inequality implies

∥X^1−Y∥2≤(∫01eLX(1−s)∥as∥ ds)2≤(∫01e2LX(1−s) ds)(∫01∥as∥2 ds).\begin{aligned} \lVert\widehat{X}_1-Y\rVert^2 \le\left(\int_0^1 e^{L_X(1-s)}\lVert a_s\rVert\,\mathrm{d}s\right)^2 \\ &\le\left(\int_0^1 e^{2L_X(1-s)}\,\mathrm{d}s\right)\left(\int_0^1 \lVert a_s\rVert^2\,\mathrm{d}s\right). \end{aligned}

The first factor equals

∫01e2LX(1−s) ds={e2LX−12LX,LX>0,1,LX=0.\int_0^1 e^{2L_X(1-s)}\,\mathrm{d}s = \begin{cases} \frac{e^{2L_X}-1}{2L_X}, & L_X>0,\\ 1, & L_X=0. \end{cases}

Taking expectation under QϕQ_{\phi} gives

DπFM≤CLXEQϕ∫01∥vθ(Xs,Z,s)−V∥2 ds=CLXLFM.D_{\pi}^{\mathrm{FM}} \le C_{L_X}\mathbb{E}_{Q_{\phi}}\int_0^1 \lVert v_{\theta}(X_s,Z,s)-V\rVert^2\,\mathrm{d}s = C_{L_X}\mathcal{L}_{\mathrm{FM}}.

Proof of Theorem 4.2

The proof first relates the Mean Flow residual to a direct prediction error at each terminal time. The time-regularity assumption then converts an integrated error bound into control at t=1t=1.

Lemma SM2.1 (Integrated Mean Flow residual). Under the absolute-continuity assumption of Theorem 4.2, define

Gθ,t(X,Z)=X+t uθ(X,Z;0,t).G_{\theta,t}(X,Z)=X+t\,u_\theta(X,Z;0,t).

Then, for every t∈(0,1]t\in(0,1],

Gθ,t(X,Z)−Xt=∫0tRr,t drQϕ-almost surely.(56)G_{\theta,t}(X,Z)-X_t=\int_0^t \mathcal{R}_{r,t}\,\mathrm{d}r \qquad Q_\phi\text{-almost surely}. \tag*{(56)}

In particular, if

A(t)=EQϕ[1t∫0t∥Rr,t∥2 dr],A(t)=\mathbb{E}_{Q_\phi}\left[\frac{1}{t}\int_0^t\lVert\mathcal{R}_{r,t}\rVert^2\,\mathrm{d}r\right],

then

EQϕ∥uθ(X,Z;0,t)−V∥2≤A(t),(57)\mathbb{E}_{Q_\phi}\lVert u_\theta(X,Z;0,t)-V\rVert^2\leq A(t), \tag*{(57)}

and

EQϕ∥Gθ,t(X,Z)−Xt∥2≤t2A(t).(58)\mathbb{E}_{Q_\phi}\lVert G_{\theta,t}(X,Z)-X_t\rVert^2\leq t^2 A(t). \tag*{(58)}

Proof. Fix t∈(0,1]t\in(0,1] and a triple for which the assumed absolute continuity holds. Write

Ut(r)=uθ(Xr,Z;r,t),Ht(r)=Xr+(t−r)Ut(r).U_t(r)=u_\theta(X_r,Z;r,t),\qquad H_t(r)=X_r+(t-r)U_t(r).

The chain rule along the reference path gives

Ut′(r)=∂ruθ(Xr,Z;r,t)+Dxuθ(Xr,Z;r,t)VU_t'(r)=\partial_r u_\theta(X_r,Z;r,t)+D_xu_\theta(X_r,Z;r,t)V

for almost every rr. Consequently,

Ht′(r)=V−Ut(r)+(t−r)Ut′(r)=−Rr,t.H_t'(r)=V-U_t(r)+(t-r)U_t'(r)=-\mathcal{R}_{r,t}.

Since HtH_t is absolutely continuous and

Ht(0)=Gθ,t(X,Z),Ht(t)=Xt,H_t(0)=G_{\theta,t}(X,Z),\qquad H_t(t)=X_t,

the fundamental theorem of calculus proves (SM2.2). Using Xt=X+tVX_t=X+tV and dividing by tt yields

uθ(X,Z;0,t)−V=1t∫0tRr,t dr.u_\theta(X,Z;0,t)-V=\frac{1}{t}\int_0^t\mathcal{R}_{r,t}\,\mathrm{d}r.

The Cauchy–Schwarz inequality gives

∥uθ(X,Z;0,t)−V∥2≤1t∫0t∥Rr,t∥2 dr.\lVert u_\theta(X,Z;0,t)-V\rVert^2\leq\frac{1}{t}\int_0^t\lVert\mathcal{R}_{r,t}\rVert^2\,\mathrm{d}r.

Taking expectation proves (SM2.3). Multiplication by t2t^2 gives (SM2.4). These inequalities remain valid with an infinite right-hand side at exceptional times.

Proof of Theorem 4.2. Let

H=L2(Qϕ;Rd),g(t)=uθ(X,Z;0,t)−V∈H,J=LMF.\mathcal{H}=L^2(Q_\phi;\mathbb{R}^d), \qquad g(t)=u_\theta(X,Z;0,t)-V\in\mathcal{H}, \qquad J=\mathcal{L}_{\mathrm{MF}}.

Because tt is uniform on [0,1][0,1] and r∣tr\mid t is uniform on [0,t][0,t], Tonelli’s theorem and Lemma SM2.1 imply

∫01∥g(t)∥H2 dt≤∫01A(t) dt=J.(59)\int_0^1 \lVert g(t)\rVert_{\mathcal{H}}^2\,\mathrm{d}t\leq\int_0^1 A(t)\,\mathrm{d}t=J. \tag*{(59)}

The value assigned to the conditional time distribution at t=0t=0 has no effect on this integral. The time-regularity assumption gives

∥g(t)−g(s)∥H≤LT∣t−s∣.\lVert g(t)-g(s)\rVert_{\mathcal{H}}\leq L_T\lvert t-s\rvert.

Set

a=∥g(1)∥H=DπMF.a=\lVert g(1)\rVert_{\mathcal{H}}=\sqrt{D_{\pi}^{\mathrm{MF}}}.

The reverse triangle inequality yields

∥g(t)∥H≥max⁡{0,a−LT(1−t)}.(60)\lVert g(t)\rVert_{\mathcal{H}}\geq\max\{0,a-L_T(1-t)\}. \tag*{(60)}

If a=0a=0, the conclusion is immediate. If LT=0L_T=0, then gg is constant, and (SM2.5) gives a2≤Ja^2\leq J, which is stronger than the claimed estimate. It remains to consider a>0a>0 and LT>0L_T>0.

If a<LTa<L_T, integration over the interval where the lower bound in (SM2.6) is positive gives

J≥∫1−a/LT1(a−LT(1−t))2 dt=a33LT.J\geq\int_{1-a/L_T}^{1}\left(a-L_T(1-t)\right)^2\,\mathrm{d}t =\frac{a^3}{3L_T}.

Hence a2≤(3LTJ)2/3a^2\leq(3L_TJ)^{2/3}.

If a≥LTa\geq L_T, the lower bound is nonnegative throughout [0,1][0,1], and

J≥∫01(a−LT(1−t))2 dt=a2−aLT+LT23≥a23.J\geq\int_0^1\left(a-L_T(1-t)\right)^2\,\mathrm{d}t =a^2-aL_T+\frac{L_T^2}{3}\geq\frac{a^2}{3}.

The last inequality follows from

a2−aLT+LT23−a23=(a−LT)(2a−LT)3≥0.a^2-aL_T+\frac{L_T^2}{3}-\frac{a^2}{3} =\frac{(a-L_T)(2a-L_T)}{3}\geq0.

Thus a2≤3Ja^2\leq3J in this case. Combining the cases proves

DπMF≤max⁡{3LMF,(3LTLMF)2/3}.D_{\pi}^{\mathrm{MF}}\leq\max\{3\mathcal{L}_{\mathrm{MF}},(3L_T\mathcal{L}_{\mathrm{MF}})^{2/3}\}.

Sharpness of the small-loss exponent. The exponent 2/32/3 is attained by a simple decoder satisfying the stated assumptions. Let X=Y=0X=Y=0 and, for fixed L>0L>0 and 0<a<L0<a<L, take a scalar decoder independent of its state, latent, and starting-time arguments:

uθ(ξ,z;r,t)=(a−L(1−t))+,b+=max⁡{b,0}.u_\theta(\xi,z;r,t)=\left(a-L(1-t)\right)_{+},\qquad b_{+}=\max\{b,0\}.

Its reference-path derivative with respect to rr vanishes, so Rr,t=uθ(t)\mathcal{R}_{r,t}=u_\theta(t). Therefore,

LMF=a33L,DπMF=a2=(3LLMF)2/3.\mathcal{L}_{\mathrm{MF}}=\frac{a^3}{3L},\qquad D_{\pi}^{\mathrm{MF}}=a^2=(3L\mathcal{L}_{\mathrm{MF}})^{2/3}.

As a↓0a\downarrow0 with LL fixed, no uniformly stronger small-loss power can replace 2/32/3 under these assumptions alone.

Proof of Theorem 4.3

We first record an elementary Gaussian comparison. The proof uses an explicit coupling and will therefore also apply after mixing encoder distributions over terminal states.

Lemma SM2.2 (Gaussian coupling and relative entropy). Let

q=N(mq,diag⁡(sq2)),p=N(mp,diag⁡(sp2))q=\mathcal{N}(m_q,\operatorname{diag}(s_q^2)),\qquad p=\mathcal{N}(m_p,\operatorname{diag}(s_p^2))

be diagonal Gaussians on Rdz\mathbb{R}^{d_z} with strictly positive standard deviations. If sp,j≤Ss_{p,j}\le S for every coordinate, then

W22(q,p)≤∥mq−mp∥2+∥sq−sp∥2W_2^2(q,p)\le\lVert m_q-m_p\rVert^2+\lVert s_q-s_p\rVert^2
≤2S2DKL(q∥p).(61)\le2S^2D_{\mathrm{KL}}(q\mathbin\Vert p). \tag*{(61)}

Proof. For a common standard Gaussian vector ϵ∼N(0,Idz)\epsilon\sim\mathcal{N}(0,I_{d_z}), set

Z=mq+sq⊙ϵ,Z′=mp+sp⊙ϵ.Z=m_q+s_q\odot\epsilon,\qquad Z'=m_p+s_p\odot\epsilon.

These variables have marginals qq and pp, and their coupling cost is

E∥Z−Z′∥2=∥mq−mp∥2+∥sq−sp∥2.\mathbb{E}\lVert Z-Z'\rVert^2=\lVert m_q-m_p\rVert^2+\lVert s_q-s_p\rVert^2.

This proves the first inequality.

For the second, set αj=sq,j/sp,j>0\alpha_j=s_{q,j}/s_{p,j}>0. The Gaussian relative entropy is

2DKL(q∥p)=∑j=1dz[(mq,j−mp,j)2sp,j2+αj2−1−2log⁡αj].2D_{\mathrm{KL}}(q\mathbin\Vert p)=\sum_{j=1}^{d_z}\left[\frac{(m_{q,j}-m_{p,j})^2}{s_{p,j}^2}+\alpha_j^2-1-2\log\alpha_j\right].

For every α>0\alpha>0,

α2−1−2log⁡α−(α−1)2=2(α−1−log⁡α)≥0.\alpha^2-1-2\log\alpha-(\alpha-1)^2=2(\alpha-1-\log\alpha)\ge0.

Consequently, for each coordinate,

(mq,j−mp,j)2+(sq,j−sp,j)2≤sp,j2[(mq,j−mp,j)2sp,j2+αj2−1−2log⁡αj]≤S2[(mq,j−mp,j)2sp,j2+αj2−1−2log⁡αj].\begin{aligned} (m_{q,j}-m_{p,j})^2+(s_{q,j}-s_{p,j})^2 \\ &\le s_{p,j}^2\left[\frac{(m_{q,j}-m_{p,j})^2}{s_{p,j}^2}+\alpha_j^2-1-2\log\alpha_j\right] \\ &\le S^2\left[\frac{(m_{q,j}-m_{p,j})^2}{s_{p,j}^2}+\alpha_j^2-1-2\log\alpha_j\right]. \end{aligned}

Summing over coordinates proves the result.

Proof of Theorem 4.3. We separate reconstruction under encoder samples from the change of latent distribution at inference.

Step 1: Reconstruction under the aggregated posterior. Define

K~1(x,⋅)=Fθ(x,⋅)#qˉϕ(⋅∣x)\widetilde{K}_1(x,\cdot)=F_\theta(x,\cdot)_{\#}\bar{q}_\phi(\cdot\mid x)

and the conditional reconstruction cost

dπ(x)=∬∥y−Fθ(x,z)∥2qϕ(dz∣x,y)Kπ(x,dy).d_\pi(x)=\iint\lVert y-F_\theta(x,z)\rVert^2q_\phi(dz\mid x,y)K_\pi(x,dy).

Then ∫dπ(x) μ0(dx)=Dπ\int d_\pi(x)\,\mu_0(dx)=D_\pi. For fixed xx, sample Y∼Kπ(x,⋅)Y\sim K_\pi(x,\cdot) and then Z∼qϕ(⋅∣x,Y)Z\sim q_\phi(\cdot\mid x,Y). The pair (Y,Fθ(x,Z))(Y,F_\theta(x,Z)) is a coupling of Kπ(x,⋅)K_\pi(x,\cdot) and K~1(x,⋅)\widetilde{K}_1(x,\cdot), because the marginal law of ZZ is qˉϕ(⋅∣x)\bar{q}_\phi(\cdot\mid x). It follows that

W2(Kπ(x,⋅),K~1(x,⋅))≤dπ(x).(62)W_2\left(K_\pi(x,\cdot),\widetilde{K}_1(x,\cdot)\right)\le\sqrt{d_\pi(x)}. \tag*{(62)}

Step 2: Replacing the aggregated posterior by the prior. Fix xx and take any coupling of qˉϕ(⋅∣x)\bar{q}_\phi(\cdot\mid x) and pη(⋅∣x)p_\eta(\cdot\mid x). Pushing both components through Fθ(x,⋅)F_\theta(x,\cdot) gives a coupling of K~1(x,⋅)\widetilde{K}_1(x,\cdot) and K^1(x,⋅)\widehat{K}_1(x,\cdot). By the latent-variable Lipschitz assumption, its squared cost is at most LZ2L_Z^2 times the original cost. Taking the infimum over latent couplings yields

W2(K~1(x,⋅),K^1(x,⋅))≤LZW2(qˉϕ(⋅∣x),pη(⋅∣x)).(63)W_2\left(\widetilde{K}_1(x,\cdot),\widehat{K}_1(x,\cdot)\right)\le L_ZW_2\left(\bar{q}_\phi(\cdot\mid x),p_\eta(\cdot\mid x)\right). \tag*{(63)}

The Wasserstein triangle inequality, followed by (SM2.8) and (SM2.9), gives

W2(K⋆(x,⋅),K^1(x,⋅))≤W2(K⋆(x,⋅),Kπ(x,⋅))+dπ(x)+LZW2(qˉϕ(⋅∣x),pη(⋅∣x)).\begin{aligned} W_2\left(K_\star(x,\cdot),\widehat{K}_1(x,\cdot)\right) &\le W_2\left(K_\star(x,\cdot),K_\pi(x,\cdot)\right)+\sqrt{d_\pi(x)}\\ &\quad+L_ZW_2\left(\bar{q}_\phi(\cdot\mid x),p_\eta(\cdot\mid x)\right). \end{aligned}

Taking the L2(μ0)L^2(\mu_0) norm and applying Minkowski’s inequality proves

E⋆(K^1)≤εcpl+Dπ+LZΔZ.(64)\mathcal{E}_\star(\widehat{K}_1)\le\varepsilon_{\mathrm{cpl}}+\sqrt{D_\pi}+L_Z\Delta_Z. \tag*{(64)}

In particular, this is a triangle-inequality estimate, not an orthogonal decomposition of squared errors.

Step 3: Gaussian control of the latent discrepancy. Fix xx and sample Y∼Kπ(x,⋅)Y\sim K_\pi(x,\cdot) and an independent ϵ∼N(0,Idz)\epsilon\sim\mathcal{N}(0,I_{d_z}). Define

Z=mϕ(x,Y)+sϕ(x,Y)⊙ϵ,Z′=mη(x)+sη(x)⊙ϵ.\begin{aligned} Z&=m_\phi(x,Y)+s_\phi(x,Y)\odot\epsilon,\\ Z'&=m_\eta(x)+s_\eta(x)\odot\epsilon. \end{aligned}

The marginal law of ZZ is the aggregated posterior, and the marginal law of Z′Z' is the prior. Hence,

W22(qˉϕ(⋅∣x),pη(⋅∣x))≤∫[∥mϕ(x,y)−mη(x)∥2+∥sϕ(x,y)−sη(x)∥2]Kπ(x,dy)≤2S2∫DKL(qϕ(⋅∣x,y)∥pη(⋅∣x))Kπ(x,dy),\begin{aligned} W_2^2\left(\bar{q}_\phi(\cdot\mid x),p_\eta(\cdot\mid x)\right) &\le\int\left[\lVert m_\phi(x,y)-m_\eta(x)\rVert^2+\lVert s_\phi(x,y)-s_\eta(x)\rVert^2\right]K_\pi(x,dy)\\ &\le2S^2\int D_{\mathrm{KL}}\left(q_\phi(\cdot\mid x,y)\|p_\eta(\cdot\mid x)\right)K_\pi(x,dy), \end{aligned}

@bibliography where the second inequality follows from Lemma SM2.2. Integration with respect to μ0\mu_0 gives

ΔZ2≤2S2RKL.\Delta_Z^2 \le2S^2\mathcal{R}_{\mathrm{KL}}.

This argument does not assume that the aggregated posterior is Gaussian; it explicitly couples the mixture to the Gaussian prior.

Finally, substitute this estimate and

Dπ≤Ψdec(Ldec)D_\pi\le\Psi_{\mathrm{dec}}(L_{\mathrm{dec}})

from the appropriate reconstruction theorem into (SM2.10). This yields

E⋆(K^1)≤ϵcpl+Ψdec(Ldec)+2LZSRKL,\mathcal{E}_\star(\widehat{K}_1) \le\epsilon_{\mathrm{cpl}} + \sqrt{\Psi_{\mathrm{dec}}(L_{\mathrm{dec}})} + \sqrt{2}L_ZS\sqrt{\mathcal{R}_{\mathrm{KL}}},

as claimed.

Proof of Proposition 4.4

Proof. Fix a fate kk and a radius ρ>0\rho> 0. Recall

Γ=⋃j=1K∂Bj,Γρ={y:dist⁡(y,Γ)≤ρ}.\Gamma= \bigcup_{j=1}^{K} \partial B_j,\qquad\Gamma_\rho= \{y : \operatorname{dist}(y,\Gamma) \le\rho\}.

For any y,y^∈Rdy,\widehat{y} \in\mathbb{R}^d,

∣1Bk(y)−1Bk(y^)∣≤1Γρ(y)+1{∥y−y^∥>ρ}.(65)\left|\mathbf{1}_{B_k}(y) - \mathbf{1}_{B_k}(\widehat{y})\right| \le\mathbf{1}_{\Gamma_\rho}(y) + \mathbf{1}_{\{\lVert y-\widehat{y}\rVert>\rho\}}. \tag*{(65)}

Indeed, if y∉Γρy \notin\Gamma_\rho and ∥y−y^∥≤ρ\lVert y-\widehat{y}\rVert\le\rho, the line segment joining the two points does not intersect ∂Bk\partial B_k. Membership in BkB_k cannot change along this segment, so the two indicators agree. No smoothness of the boundary is required for this observation.

For fixed xx, let λ\lambda be any coupling of K⋆(x,⋅)K_\star(x,\cdot) and K^1(x,⋅)\widehat{K}_1(x,\cdot). Using (SM2.11),

∣pk⋆(x)−p^k(x)∣≤∫∣1Bk(y)−1Bk(y^)∣λ(dy,dy^)≤K⋆(x,Γρ)+λ{∥y−y^∥>ρ}≤K⋆(x,Γρ)+1ρ2∫∥y−y^∥2λ(dy,dy^).\begin{aligned} \left|p_k^\star(x)-\widehat{p}_k(x)\right| \le\int\left|\mathbf{1}_{B_k}(y)-\mathbf{1}_{B_k}(\widehat{y})\right|\lambda(\mathrm{d}y,\mathrm{d}\widehat{y}) \\ &\le K_\star(x,\Gamma_\rho)+\lambda\{\lVert y-\widehat{y}\rVert>\rho\} \\ &\le K_\star(x,\Gamma_\rho)+\frac{1}{\rho^2}\int\lVert y-\widehat{y}\rVert^2\lambda(\mathrm{d}y,\mathrm{d}\widehat{y}). \end{aligned}

Taking the infimum over λ\lambda gives the pointwise estimate

∣pk⋆(x)−p^k(x)∣≤K⋆(x,Γρ)+W22(K⋆(x,⋅),K^1(x,⋅))ρ2.(66)\left|p_k^\star(x)-\widehat{p}_k(x)\right| \le K_\star(x,\Gamma_\rho)+\frac{W_2^2\left(K_\star(x,\cdot),\widehat{K}_1(x,\cdot)\right)}{\rho^2}. \tag*{(66)}

This pointwise argument does not require selecting optimal couplings measurably as a function of xx.

Integrating against μ0\mu_0 and using the true terminal marginal,

∫K⋆(x,Γρ) μ0(dx)=μ1(Γρ),\int K_\star(x,\Gamma_\rho)\,\mu_0(\mathrm{d}x)=\mu_1(\Gamma_\rho),

we obtain

∥p^k⋆−p^k∥L1(μ0)≤μ1(Γρ)+E⋆2(K^1)ρ2.\lVert\widehat{p}_k^\star-\widehat{p}_k\rVert_{L^1(\mu_0)}\leq\mu_1(\Gamma_\rho)+\frac{\mathcal{E}_\star^2(\widehat{K}_1)}{\rho^2}.

The right-hand side is independent of kk, so the same estimate holds after taking the maximum over fates. Theorem 4.3 gives the further bound obtained by replacing E⋆\mathcal{E}_\star with Bdec\mathcal{B}_{\mathrm{dec}}. Finally, Γ\Gamma is closed because it is a finite union of boundaries. Therefore Γρ↓Γ\Gamma_\rho\downarrow\Gamma as ρ↓0\rho\downarrow0, and

μ1(Γρ)⟶μ1(Γ).\mu_1(\Gamma_\rho)\longrightarrow\mu_1(\Gamma).

If μ1(Γ)=0\mu_1(\Gamma)=0, first choose ρ\rho so that the boundary mass is small, and then let the conditional reconstruction error tend to zero with this ρ\rho fixed. This proves the stated convergence of the average fate probabilities.

Proof of Monte-Carlo error bound

Proposition SM2.3 (Monte Carlo estimation error). For each aa and kk, the estimator

p^k(N)(xa)=1N∑ℓ=1N1Bk(Y^a,ℓ).(67)\widehat{p}_k^{(N)}(x_a)=\frac{1}{N}\sum_{\ell=1}^{N}\mathbf{1}_{B_k}(\widehat{Y}_{a,\ell}). \tag*{(67)}

is unbiased for p^k(xa)\widehat{p}_k(x_a) and satisfies

EMC[∣p^k(N)(xa)−p^k(xa)∣2]=p^k(xa)(1−p^k(xa))N≤14N.(68)\mathbb{E}_{\mathrm{MC}}\left[\left|\widehat{p}_k^{(N)}(x_a)-\widehat{p}_k(x_a)\right|^2\right] =\frac{\widehat{p}_k(x_a)(1-\widehat{p}_k(x_a))}{N} \leq\frac{1}{4N}. \tag*{(68)}

Moreover, for any δ∈(0,1)\delta\in(0,1), with probability at least 1−δ1-\delta,

max⁡1≤a≤n1≤k≤K∣p^k(N)(xa)−p^k(xa)∣≤ϵMC,(69)\max_{\substack{1\leq a\leq n\\1\leq k\leq K}} \left|\widehat{p}_k^{(N)}(x_a)-\widehat{p}_k(x_a)\right| \leq\epsilon_{\mathrm{MC}}, \tag*{(69)}

where

ϵMC=log⁡(2nK/δ)2N.(70)\epsilon_{\mathrm{MC}}=\sqrt{\frac{\log(2nK/\delta)}{2N}}. \tag*{(70)}

Proof. Condition on the fitted model and the query states. For fixed aa and kk, set

Iℓ=1Bk(Y^a,ℓ),p=p^k(xa).I_\ell=\mathbf{1}_{B_k}(\widehat{Y}_{a,\ell}),\qquad p=\widehat{p}_k(x_a).

The variables I1,…,INI_1,\ldots,I_N are independent Bernoulli variables with success probability pp. Thus,

EMCp^k(N)(xa)=p,\mathbb{E}_{\mathrm{MC}}\widehat{p}_k^{(N)}(x_a)=p,

and independence gives

EMC∣p^k(N)(xa)−p∣2=Var⁡MC(1N∑ℓ=1NIℓ)=p(1−p)N≤14N.\mathbb{E}_{\mathrm{MC}}\left|\widehat{p}_k^{(N)}(x_a)-p\right|^2 =\operatorname{Var}_{\mathrm{MC}}\left(\frac{1}{N}\sum_{\ell=1}^{N}I_\ell\right) =\frac{p(1-p)}{N}\leq\frac{1}{4N}.

For completeness, we give the exponential-moment argument for the required concentration bound. For a Bernoulli variable II with success probability pp, define

H(λ)=log⁡Eeλ(I−p)=log⁡(1−p+peλ)−pλ.H(\lambda)=\log\mathbb{E}e^{\lambda(I-p)}=\log(1-p+pe^\lambda)-p\lambda.

Then H(0)=H′(0)=0H(0)=H'(0)=0, and

H′′(λ)=ϑλ(1−ϑλ)≤14,ϑλ=peλ1−p+peλ.H''(\lambda)=\vartheta_\lambda(1-\vartheta_\lambda)\le\frac{1}{4},\qquad\vartheta_\lambda=\frac{pe^\lambda}{1-p+pe^\lambda}.

Taylor’s formula therefore implies

Eeλ(I−p)≤eλ2/8for every λ∈R.\mathbb{E}e^{\lambda(I-p)}\le e^{\lambda^2/8}\qquad\text{for every }\lambda\in\mathbb{R}.

By independence and exponential Markov’s inequality, for ε>0\varepsilon>0 and λ>0\lambda>0,

PMC(1N∑ℓ=1NIℓ−p≥ε)≤exp⁡(−Nλε+Nλ28).\mathbb{P}_{\mathrm{MC}}\left(\frac{1}{N}\sum_{\ell=1}^{N} I_\ell-p\ge\varepsilon\right)\le\exp\left(-N\lambda\varepsilon+\frac{N\lambda^2}{8}\right).

Choosing λ=4ε\lambda=4\varepsilon gives the upper-tail bound e−2Nε2e^{-2N\varepsilon^2}. Applying the same argument to the negative centered sum gives

PMC(∣p^k(N)(xa)−p^k(xa)∣>ε)≤2e−2Nε2.(71)\mathbb{P}_{\mathrm{MC}}\left(\left|\hat{p}^{(N)}_k(x_a)-\hat{p}_k(x_a)\right|>\varepsilon\right)\le2e^{-2N\varepsilon^2}. \tag*{(71)}

This is the Bernoulli form of Hoeffding’s inequality.

A union bound over the nKnK pairs (a,k)(a,k) gives

PMC(max⁡a,k∣p^k(N)(xa)−p^k(xa)∣>ε)≤2nKe−2Nε2.\mathbb{P}_{\mathrm{MC}}\left(\max_{a,k}\left|\hat{p}^{(N)}_k(x_a)-\hat{p}_k(x_a)\right|>\varepsilon\right)\le2nK e^{-2N\varepsilon^2}.

Set

ε=εMC=log⁡(2nK/δ)2N.\varepsilon=\varepsilon_{\mathrm{MC}}=\sqrt{\frac{\log(2nK/\delta)}{2N}}.

The right-hand side then equals δ\delta, proving the claimed simultaneous bound. Independence between indicators of different fates is not needed for this union-bound step.

Proof of Corollary 4.5

Proof. Let Aδ\mathcal{A}_\delta be the event

max⁡a,k∣p^k(N)(xa)−p^k(xa)∣≤εMC.\max_{a,k}\left|\hat{p}^{(N)}_k(x_a)-\hat{p}_k(x_a)\right|\le\varepsilon_{\mathrm{MC}}.

By Proposition SM2.3, PMC(Aδ)≥1−δ\mathbb{P}_{\mathrm{MC}}(\mathcal{A}_\delta)\ge1-\delta. On this event, for every fate kk,

∑a=1nωa∣p^k(N)(xa)−pk⋆(xa)∣≤∑a=1nωa∣p^k(xa)−pk⋆(xa)∣+∑a=1nωa∣p^k(N)(xa)−p^k(xa)∣≤∥p^k−pk⋆∥L1(μ0)+εMC,\begin{aligned} \sum_{a=1}^{n}\omega_a\left|\hat{p}^{(N)}_k(x_a)-p_k^\star(x_a)\right| \\ &\le\sum_{a=1}^{n}\omega_a\left|\hat{p}_k(x_a)-p_k^\star(x_a)\right|+\sum_{a=1}^{n}\omega_a\left|\hat{p}^{(N)}_k(x_a)-\hat{p}_k(x_a)\right| \\ &\le\|\hat{p}_k-p_k^\star\|_{L^1(\mu_0)}+\varepsilon_{\mathrm{MC}}, \end{aligned}

Table SM1: Experimental datasets. Here we use low dimension representation of real datasets deduced from standard SPRING[SM23] method.

where we used μ0=∑aωaδxa\mu_0 = \sum_a \omega_a\delta_{x_a} and ∑aωa=1\sum_a \omega_a = 1. Proposition 4.4 now implies that, on the same event and for every ρ>0\rho> 0,

max⁡k∑a=1nωa∣p^k(N)(xa)−pk⋆(xa)∣≤μ1(Γρ)+Bdec2ρ2+ϵMC.\max_k \sum_{a=1}^{n} \omega_a \left|\hat{p}^{(N)}_k(x_a)-p_k^\star(x_a)\right| \le\mu_1(\Gamma_\rho)+\frac{\mathcal{B}_{\mathrm{dec}}^2}{\rho^2}+\epsilon_{\mathrm{MC}}.

Taking the infimum over ρ>0\rho> 0 and substituting the expression for ϵMC\epsilon_{\mathrm{MC}} proves

max⁡k∑a=1nωa∣p^k(N)(xa)−pk⋆(xa)∣≤inf⁡ρ>0{μ1(Γρ)+Bdec2ρ2}+log⁡(2nK/δ)2N.\max_k \sum_{a=1}^{n} \omega_a \left|\hat{p}^{(N)}_k(x_a)-p_k^\star(x_a)\right| \le\inf_{\rho>0}\left\{\mu_1(\Gamma_\rho)+\frac{\mathcal{B}_{\mathrm{dec}}^2}{\rho^2}\right\}+\sqrt{\frac{\log(2nK/\delta)}{2N}}.

The event Aδ\mathcal{A}_\delta does not depend on ρ\rho. Thus, taking this infimum introduces no additional probability penalty.

Additional Experimental Details and Results

This section provides the implementation settings, numerical-integration comparisons, hyperparameter sensitivity results, and additional hematopoietic reconstruction scores. We also present qualitative comparisons of the learned fields, latent representations, and baseline fate maps. Population reconstruction and fate-probability accuracy are reported separately throughout.

Experimental setup and evaluation protocol

Data and implementation. The controlled benchmark is a two-dimensional branching system with an analytically specified branch probability. The biological applications use barcode-resolved hematopoietic data from LARRY [SM22] and direct-reprogramming data from CellTagging [SM1]. Dataset sizes and the correspondence between experimental and rescaled times are recorded in Table SM1.

DatasetTime intervalSource cellsTarget cells
Synthetic branching0 →\to 11,0005,000
HematopoiesisDay 2→\toDay 611446477
ReprogrammingDay 15→\toDay 2818975945

Table SM1. Experimental datasets. Here we use low dimension representation of real datasets deduced from standard SPRING[SM23] method.

The reported LIFT implementation uses multilayer perceptrons with ReLU activations for the encoder, conditional prior, and decoder, with a two-dimensional latent variable in the principal experiments. Training uses Adam and the objective introduced in § 3.3. Our timing results were obtained on an Intel i7-13700 CPU. Dataset-specific network widths, optimization parameters, and sensitivity results are contained below.

Training couplings and comparison methods. We compare geometric endpoint couplings with lineage-informed couplings, denoted by πgeo\pi_{\mathrm{geo}} and πlin\pi_{\mathrm{lin}}. The latter use shared barcodes to restrict the admissible source–target relationships; the reprogramming experiment additionally uses OT within each retained clone. During LIFT sampling, only the molecular state is provided to the conditional prior. Barcode labels and terminal states are not inputs to the sampling network.

The primary comparisons include deterministic CFM and Mean Flow, the score-matching baseline denoted SM. We also compare latent decoders sharing the encoder–prior construction: a static MLP, an FM decoder, and the Mean Flow decoder used by LIFT. These experiments distinguish changes in the coupling from changes in the generator.

Evaluation and data separation. Population reconstruction is assessed by a Wasserstein distance in the evaluation state representation. Conditional evaluation uses analytic branch probabilities in the controlled benchmark and lineage derived outcome information in the biological applications. LIFT probabilities are estimated using 1,000 independent prior samples per initial state. The terminal fate assignment rule is shared across methods and is specified by using the convex hull as fate region. Throughout evaluated datasets, convex hulls of different labeled fates do not intersect so this assignment is well-defined. As for evaluation, we randomly separate cells from 20% barcodes as test set.

LIFT implementation

The encoder, conditional prior, and Mean Flow decoder are multilayer perceptrons with ReLU activations. The encoder receives the concatenated endpoint pair (x,y)(x, y) and parameterizes a diagonal Gaussian latent posterior. The conditional prior parameterizes a diagonal Gaussian using xx alone. The decoder receives the state, latent variable, starting time, and terminal time, and returns the interval-average velocity. The networks are trained jointly with Adam. Table SM2 lists the dataset-specific architecture and optimization settings; the real-data configuration is shared by the hematopoiesis and reprogramming experiments.

ParameterSyntheticReal data
Encoder widths32, 64, 3232, 128, 128, 32
Prior widths32, 64, 3232, 128, 128, 32
Decoder widths32, 64, 3232, 64, 128, 64, 32
Latent dimension dzd_z22
Regularization weight β\beta102
Reconstruction weight CC110
Learning rate10−310^{-3}10−410^{-4}
Batch size1,0001,024
Training iterations5,00020,000

Table SM2. LIFT architecture and training parameters. Widths specify hidden layers in order. The real-data column applies to both lineage-traced scRNA-seq applications. ReLU activations and Adam optimization are used throughout.

Training pairs are sampled from the prescribed coupling π\pi, and latent variables are sampled from qϕ(⋅∣x,y)q_{\phi}(\cdot\mid x,y) using reparameterization. The time variables are sampled as t∼U[0,1]t \sim U[0,1] and r∣t∼U[0,t]r \mid t \sim U[0,t]. The directional derivative in the Mean Flow target is computed by a Jacobian– vector product, and the target is detached during parameter updates. The KL divergences between diagonal Gaussians are evaluated analytically. These operations follow the training algorithm in the main article.

At inference, the encoder is not used. Given an initial state xx, we sample Z∼pη(⋅∣x)Z \sim p_{\eta}(\cdot\mid x) and evaluate

X^t=x+t uθ(x,Z;0,t).\widehat{X}_t = x + t\,u_{\theta}(x,Z;0,t).

The same latent realization is retained across requested times along one generated path. Fate probabilities are estimated from 1,000 independent latent samples per initial state. The real-data experiments use the two-dimensional SPRING representations described in the main article. Barcode information is incorporated through the coupling rather than as an input to the sampling network. Timing measurements are obtained on an Intel i7-13700 CPU.

Hyperparameter sensitivity

We vary the latent dimension dzd_z, reconstruction weight CC, and standard-Gaussian regularization weight β\beta on the synthetic branching benchmark. One parameter is varied at a time, with the others fixed at the default values (dz,C,β)=(2,1,10)(d_z, C, \beta) = (2, 1, 10). Table SM3 gives the complete results for α∈{0.5,1,2}\alpha\in\{0.5, 1, 2\}.

ParameterValueα=0.5\alpha= 0.5 W2α=0.5\alpha= 0.5 PEα=1\alpha= 1 W2α=1\alpha= 1 PEα=2\alpha= 2 W2α=2\alpha= 2 PE
Latent dimension dzd_z10.4170.1350.2880.5770.3021.212
20.2660.1210.2680.4570.2800.809
40.2770.1220.3060.3340.3791.095
Reconstruction weight CC0.50.2500.1790.2790.5370.2610.936
10.2660.1210.2680.4570.2800.809
20.2720.1990.2860.4460.2790.569
40.2900.2070.3350.4310.3670.670
Regularization weight β\beta10.3670.3620.2940.5380.3091.175
30.2590.1980.3030.3670.2790.840
50.2570.2160.2860.3820.2750.936
100.2660.1210.2680.4570.2800.809

Table SM3. Hyperparameter sensitivity on the synthetic branching benchmark. The default setting is (dz,C,β)=(2,1,10)(d_z, C, \beta) = (2, 1, 10). Lower values are better for W2W_2 and PE.

Table SM3: Hyperparameter sensitivity on the synthetic branching benchmark. The default setting is (dz,C,β)=(2,1,10)(d_z, C, \beta) = (2, 1, 10). Lower values are better for W2W_2 and PE.

These results support the default configuration as a balanced choice across the tested regimes, rather than a uniformly optimal setting.

Ablation on LIFT components

Separating representation from decoding. We next keep the encoder and conditional prior and change the decoder: a static latent MLP, a latent FM decoder, or the latent Mean Flow decoder. The FM decoder in this ablation includes ZZ and should not be confused with the unaugmented CFM baseline in Table 1.

The results in Table SM4 show a trade-off rather than uniform dominance. Latent FM gives the lowest population W2W_2 in all three settings, while latent Mean Flow gives the lowest

DecoderTimeα=0.5\alpha= 0.5α=1\alpha= 1α=2\alpha= 2
W2W_2PEW2W_2PEW2W_2PE
Latent FM43.160.2010.9280.2340.8050.2671.714
Latent MLP0.85160.2800.3500.4130.8130.3791.325
Latent Mean Flow0.96070.2590.1980.2680.4570.2800.809

Table SM4. Latent-decoder ablation. FM uses 100 Euler steps. Times are milliseconds per batch of 1,000 points on the reported CPU, averaged over 100 timing measurements. All three variants retain latent conditioning.

Table SM4: Latent-decoder ablation. FM uses 100 Euler steps. Times are milliseconds per batch of 1,000 points on the reported CPU, averaged over 100 timing measurements. All three variants retain latent conditioning.

PE and avoids iterative integration. The static MLP is slightly faster than Mean Flow but has larger probability errors. Thus, the advantage of the proposed decoder is not that one-step generation is unique to Mean Flow, but that it combines low sampling cost with accurate conditional fate prediction in these experiments.

Additional qualitative comparisons

Conditioned prediction fields. Figure SM1 displays the instantaneous velocity fields of CFM and the interval-average velocity outputs of the unaugmented Mean Flow model. Both exhibit a low-magnitude region separating the two branch-directed regimes. In Figure SM2, the location of this separating region changes with the latent code in LIFT. For a fixed code, the prediction remains deterministic, but different codes allow the same observed state to receive different predictions. This provides a qualitative illustration of latent conditioning; it is not a proof of nonintersection or of a topological property of the learned dynamics.

Figure SM1. Vector fields derived by standard conditional flow matching (a-c) and mean vector fields derived by standard mean flow (d-f).

Three vector-field plots for different latent variables

Figure SM2. Vector fields derived by LIFT with different latent variable zz.

Posterior and conditional-prior representations. Figure SM3 compares posterior means mϕ(x,y)m_{\phi}(x,y) and conditional-prior means mη(x)m_{\eta}(x) for the hematopoietic application. Endpoint pairs associated with Mo and Neu outcomes occupy different regions in the posterior representation. The encoder receives the terminal state, so this separation illustrates how endpoint information is encoded, rather than an independent prediction made from the initial state. The prior representation uses the initial state alone and also shows fate-associated structure. Fate probabilities, however, are obtained by sampling the full prior and applying the decoder, not by assigning a fate to the prior mean. These visualizations characterize learned representations without identifying a latent biological mechanism.

Two latent-representation plots for the hematopoietic experiment

Figure SM3. Latent representations in the hematopoietic experiment. (a) Posterior means mϕ(x,y)m_{\phi}(x,y), which use both endpoints. (b) Conditional-prior means mη(x)m_{\eta}(x), which use the initial state only. Colors indicate the Mo and Neu fate annotations associated with the displayed observations.

Baseline fate maps. Figure SM4 provide the complete baseline fate maps for the hematopoiesis applications. The maps compare geometric and lineage-informed couplings for CFM and Mean Flow, together with SM. They complement the numerical population and probability scores by showing where the spatial patterns of predicted fates differ. Gray regions in the baseline visualizations mark predictions outside the assigned fate regions, rather than an additional biological fate. The maps should be read together with the quantitative comparisons, since visually separated outcomes do not by themselves establish accurate conditional probabilities or irreversible fate commitment.

Figure SM4. Visualization of cell fate probabilities for other baseline methods on the hematopoiesis dataset. Colors match the main text.

References

  1. [1]B. A. Biddy, W. Kong, K. Kamimoto, C. Guo, S. E. Waye, T. Sun, and S. A. Morris, Single-cell mapping of lineage and identity in direct reprogramming, Nature, 564 (2018), pp. 219–224, https://doi.org/10.1038/s41586-018-0744-4.
  2. [2]V. De Bortoli, G.-H. Liu, T. Chen, E. A. Theodorou, and W. Nie, Augmented bridge matching, 2023, https://doi.org/10.48550/arXiv.2311.06978.
  3. [3]A. Forrow and G. Schiebinger, LineageOT is a unified framework for lineage tracing and trajectory inference, Nature Communications, 12 (2021), p. 4940, https://doi.org/10.1038/s41467-021-25133-1.
  4. [4]Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He, Mean flows for one-step generative modeling, in Advances in Neural Information Processing Systems, vol. 38, 2025, https://doi.org/10.52202/085713-2534.
  5. [5]P. Guo and A. Schwing, Variational rectified flow matching, in Proceedings of the 42nd International Conference on Machine Learning, vol. 267 of Proceedings of Machine Learning Research, PMLR, 2025, pp. 20921–20940, https://proceedings.mlr.press/v267/guo25i.html.arxiv.org/abs/2502.09616
  6. [6]W. Hoeffding, Probability inequalities for sums of bounded random variables, Journal of the American Statistical Association, 58 (1963), pp. 13–30, https://doi.org/10.1080/01621459.1963.10500830.
  7. [7]D. P. Kingma and M. Welling, Auto-encoding variational Bayes, in International Conference on Learning Representations, 2014, https://arxiv.org/abs/1312.6114.
  8. [8]D. P. Kingma and M. Welling, An introduction to variational autoencoders, Foundations and Trends in Machine Learning, 12 (2019), pp. 307–392, https://doi.org/10.1561/2200000056.
  9. [9]M. Lange, Z. Piran, M. Klein, B. Spanjaard, D. Klein, J. P. Junker, F. J. Theis, and M. Nitzan, Mapping lineage-traced cells across time points with moslin, Genome Biology, 25 (2024), p. 277, https://doi.org/10.1186/s13059-024-03422-4.

Paper details

vlnc:W_hwpu78pg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

AI tools

View statement
Looking for related papers…
Contents