🤖 machine learning

Learning Individual Dynamics from Sparse Cross-Sectional Snapshots

This paper introduces CADENCE, a probabilistic framework that uniquely recovers continuous individual trajectories from sparse cross-sectional snapshots by anchoring latent dynamics to static contexts, thereby achieving identifiability and performance comparable to dense longitudinal models without requiring sequential data.

Original authors: Christian Lagemann, Kai Lagemann, Steven L. Brunton, Sach Mukherjee

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Christian Lagemann, Kai Lagemann, Steven L. Brunton, Sach Mukherjee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to predict how a specific person will age, how a specific virus will spread, or how a specific machine will wear out. Usually, to do this, you need a "movie" of that person or object: you need to watch them every day for years to see the full story.

But in the real world, we often don't have movies. We only have snapshots. Maybe we see a patient once in a hospital, a machine once in a factory, or a virus sample once in a lab. We have thousands of these snapshots from different people at different times, but we don't know how any single one of them changes over time.

Trying to guess the future of a single person from just one photo is like trying to guess the plot of a movie by looking at a single frame. Mathematically, this is a broken puzzle; there are infinite ways to connect the dots, and most methods fail.

The Paper's Big Idea: CADENCE
The authors introduce a new method called CADENCE. It solves this puzzle by realizing that while we don't have movies for individuals, we do have a massive library of snapshots from people who are similar to each other.

Here is how CADENCE works, using a simple analogy:

1. The "Context" is the Key

Imagine you are trying to predict the future of a 20-year-old athlete. You only have one photo of them.

The Old Way: Look at the photo and guess. (Impossible).
The CADENCE Way: Look at the photo and ask, "Who else is like this?" You find a database of thousands of other 20-year-old athletes. Even though you don't have a movie of your athlete, you might have a movie of a different 20-year-old athlete who started at the same time and has the same stats.

CADENCE uses the "static context" (the person's age, genetics, or machine specs) to find these "twins" in the database. It assumes that if two people look the same at the start, they will follow the same "script" or "trajectory" as they age.

2. The "Script" (The Archetypes)

The paper argues that even though everyone is unique, they all follow a few basic "scripts" or patterns.

The Analogy: Think of a theater troupe. There are only a few main roles (the Hero, the Villain, the Comic Relief). Even though there are 1,000 actors, they all play one of these few roles.
How CADENCE uses this: It doesn't try to invent a unique script for every single person. Instead, it identifies which "role" (or Archetype) a person belongs to based on their context. Once it knows the role, it can predict the future of that role. If your patient fits the "Fast Recovery" archetype, the model predicts they will recover quickly, based on how other people in that group recovered.

3. The Two-Step Process

The method works in two distinct stages to avoid getting confused:

Stage 1: The Translator (Spatial Encoding)
The raw data (like a complex medical scan or a high-resolution photo) is messy and hard to compare. CADENCE first translates all these messy images into a clean, standardized "language" (a latent space). It's like converting thousands of different dialects into a single, perfect English so everyone can be compared fairly. This step removes the "noise" and ensures that two people who look similar in the raw data are recognized as similar here.
Stage 2: The Director (Temporal Dynamics)
Now that everyone is speaking the same language, the model acts as a director. It looks at the "context" (the actor's resume) and says, "Okay, you are playing the 'Fast Recovery' role." It then pulls up the "script" for that role and predicts how the actor will move in the future.
- The Magic: It does this even if it has never seen a full movie of your specific actor. It just needs to know the script for the role you are playing.

Why is this a breakthrough?

Previous methods had to choose between two bad options:

The "Movie" Method: Needed dense, long-term data for every single person. If you only had one photo, it failed.
The "Crowd" Method: Could handle one photo, but it only predicted the average behavior of the whole group. It lost the individual details.

CADENCE breaks this rule. It uses the "Crowd" to learn the "Scripts" (Archetypes) and then applies those scripts to the "Individual" using their context.

The Results

The authors tested this on everything from simple physics simulations to real-world biological data (like how blood cells turn into different types of cells).

They trained CADENCE only on sparse snapshots (one photo per person).
They compared it to other models that were trained on full movies (dense data).
The Result: CADENCE performed just as well as, or sometimes better than, the models that had the luxury of seeing the full movies.

The Catch (Limitations)

The paper is honest about its limits. This method only works if the "Context" (the static info you have) actually tells you which "Script" the person is following.

The Analogy: If you try to predict a person's future based on their height, but their future actually depends on a secret genetic mutation you didn't measure, the model will fail. It will just guess the average.
The paper calls this the "Context Observability Assumption." If the static data is missing the crucial clues, the model cannot do magic.

Summary

CADENCE is a tool that lets us predict the future of individuals even when we only have a single snapshot of them. It does this by realizing that individuals are like actors following a few standard scripts. By matching a person to their script based on their background, the model can "fill in the missing movie" without ever having seen the full film.

Technical Summary: Learning Individual Dynamics from Sparse Cross-Sectional Snapshots

Problem Formulation

The paper addresses a fundamental challenge in quantitative sciences: predicting the continuous-time evolution of individual units (e.g., aging patients, degrading systems) when only extremely sparse or purely cross-sectional data is available. In many real-world scenarios, dense longitudinal tracking is impossible due to cost, difficulty, or high attrition rates, leaving researchers with datasets where each unit is observed only once or a few times at irregular intervals.

Mathematically, inferring an individual continuous vector field from unpaired cross-sectional data is a severely ill-posed inverse problem. Existing methods face a dichotomy:

Sequence models (e.g., Latent ODEs) require dense longitudinal sequences to propagate gradients and learn temporal representations; they fail with 1–3 snapshots.
Cross-sectional methods (e.g., Optimal Transport, Flow Matching) map aggregate population distributions but "track the crowd," losing all individual identity and dynamics.
Contextual dynamics models condition parameters on static context but universally require dense sequential observations to fit the routing function.

The core question is whether it is possible to rigorously predict individual, continuous-time trajectories from isolated snapshots if the ensemble possesses contextual structure.

Methodology: CADENCE

The authors introduce CADENCE (Contextual Archetypes and Diffusion ENCodings for Dynamics Estimation), a probabilistic framework that recovers continuous individual trajectories by anchoring latent dynamics to static, individual-level contexts. The method relies on three structural assumptions grounded in foliation theory, nonlinear observability, and convex geometry:

Dynamical Foliation Assumption (DFA): The context space can be partitioned such that units within a partition (leaf) share the same dynamical parameters. This allows a late-stage unit's observation to supervise the future of a contextually matched early-stage unit.
Context Observability Assumption (COA): Static context contains sufficient information to identify a unit's dynamical regime (leaf membership), though not the specific parameters directly.
Manifold Regularity Assumption (MRA): The space of dynamical parameters lies on a compact submanifold and can be approximated as a convex combination of a finite set of "archetypal" experts.

Architectural Components

The architecture is a direct implementation of these theoretical constraints:

Score-Based Bijective PF-ODE (Spatial Encoder): To resolve the non-identifiability of the latent space (where infinitely many diffeomorphisms could explain the data), the spatial mapping is implemented as a score-based Probability Flow ODE trained via denoising score matching. This pins the latent space to a standard Gaussian, collapsing the symmetry group from the infinite-dimensional diffeomorphism group to a compact subgroup of isometries.
Soft Mixture-of-Experts (SMoE) Router: A neural network maps static context to routing weights ( $\alpha$ ), which form a convex combination of a shared basis of expert parameters ( $W$ ). This implements the MRA, allowing the model to represent continuous dynamical regimes as blends of fundamental archetypes.
FiLM Conditioning: The resulting individual parameter vector conditions a shared Neural ODE vector field via Feature-wise Linear Modulation.

Training Strategy

To avoid the computational bottleneck of end-to-end backpropagation through cascaded ODEs ( $O(Steps_{spatial} \times Steps_{temporal})$ ), CADENCE employs a decoupled two-stage training scheme:

Stage 1 (Spatial Encoding): The spatial encoder is trained via denoising score matching on $(x, c)$ pairs and then frozen. All observations are encoded offline into latent states.
Stage 2 (Temporal Dynamics): The temporal modules are trained in the low-dimensional latent space.
- If sparse follow-up data exists, a Semi-longitudinal objective (latent MSE) is used.
- For purely cross-sectional data, a Cross-sectional distributional objective is used. This minimizes the Maximum Mean Discrepancy (MMD) between the model's predicted conditional marginals (grouped by expert) and a reference marginal, enforcing ensemble-level consistency without requiring matched temporal pairs.

Key Contributions

Identifiability Theory: The paper provides the first rigorous identifiability guarantees for single-timepoint trajectory inference. It proves that under the three structural assumptions and the spatial prior, the full generative system (spatial encoder, ODE vector field, and routing function) is identifiable up to a harmless isometric transformation.
New Learning Paradigm: It formally bridges the gap between cross-sectional optimal transport and dense sequential modeling, demonstrating that individual continuous-time trajectories can be inferred from $\le 3$ observations per unit.
Decoupled Architecture & Performance: The proposed two-stage training reduces computational cost significantly. Empirically, CADENCE, trained strictly on sparse snapshots, matches or exceeds the performance of state-of-the-art sequential models trained on dense, full-trajectory data.

Results

The authors evaluated CADENCE on a suite of seven benchmarks ranging from controlled dynamical systems (Lotka-Volterra, Van der Pol) to high-dimensional biomedicine (gene expression, single-cell genomics).

Quantitative Performance: Across benchmarks, CADENCE achieved the lowest Mean Absolute Error (MAE) in low-dimensional settings and remained highly competitive in high-dimensional settings. Crucially, it maintained excellent Sliced Wasserstein Distance (SW2) scores, indicating it correctly captured the population distribution, whereas sequential models trained on dense data often exhibited severe distributional drift when tested on sparse data.
Subgroup Recovery: In experiments with heterogeneous dynamics (e.g., BM1, BM6), CADENCE successfully recovered true subgroup structures in an unsupervised manner (e.g., 94% routing accuracy in BM1, 85% in BM6), whereas baseline models either collapsed to a single mean or diverged.
Real-World Application (BM7): On the LARRY haematopoiesis dataset, CADENCE achieved state-of-the-art fate prediction accuracy (67.8% vs. 58.5% for the best baseline) and superior population-level distributional distance, demonstrating scalability to primary, real-world dynamical systems.
Ablation: Removing the SMoE router caused severe performance degradation, validating the necessity of context-driven routing. Similarly, spatial compression via the PF-ODE was shown to be essential for stability in high-dimensional settings.

Significance and Claims

The paper claims to break the longstanding dichotomy between aggregate mass transport and data-hungry sequential latent models. By rigorously converting sparse cross-sectional snapshots into effective longitudinal signals through contextual anchoring, CADENCE enables individualized, continuous-time forecasting in domains where full longitudinal tracking is impossible.

The authors emphasize that the architecture is not a collection of heuristic choices but a minimal implementation of specific geometric and dynamical constraints required for identifiability. They acknowledge limitations, noting that the validity of recovered trajectories depends on the structural assumptions holding true; specifically, if critical information dictating the dynamical archetype is missing from the static context (violating COA), the model may degrade to an ensemble-level transport flow. However, in settings with rich context information, the framework offers a principled solution to a previously intractable problem.