Imagine trying to map the entire global weather system using only a handful of sensors scattered across a few continents. Or tracking the structural integrity of a massive suspension bridge using just three accelerometers. This is the challenge of partial observability. High-dimensional systems—whether fluid flows, utility networks, or physical structures—possess thousands or even infinite degrees of freedom. Yet, physical, financial, and logistical constraints restrict us to sparse, localized measurements.
If a system could explore every possible configuration completely randomly, reconstructing its global state from sparse data would be mathematically impossible. The reason we can reconstruct the full state is that real-world systems are highly constrained: they exhibit low-rank structures, follow strict physical laws, or evolve along low-dimensional attractors.
By leveraging these hidden structures, modern frameworks can bridge the gap between sparse point measurements and full-field understanding.
Exploiting Low-Rank and Sparse Structures
The most fundamental reason sparse reconstruction works is that the true information density of a system is often much lower than its nominal dimensionality.
In signal processing and communication networks, data matrices often exhibit low-rank properties because different channels or nodes share underlying correlations. For instance, Meng et al. demonstrated that cognitive radio networks can identify unused "spectrum holes" using a fraction of potential sensing data. By treating the global spectrum state as a low-rank matrix, the authors used matrix completion to accurately fill in missing observations, drastically reducing communication overhead.
When the state of a system is sparse—meaning only a few components are active or non-zero at any given time—the mathematical guarantees of compressive sensing apply. Dai and Yuksel showed that the initial state of a linear system under sparsity constraints can be recovered with significantly fewer measurements than classical control theory dictates.
In more complex, nonlinear dynamical systems, this sparsity can be captured by building libraries of known behaviors. Brunton et al. combined dimensionality reduction with $\ell_1$-based compressive sensing to classify and reconstruct nonlinear bifurcation regimes. Because the system's dynamics are constrained to specific regimes, a small number of noisy point measurements is sufficient to identify the current global state and predict its future evolution.
Resolving Ambiguity via Spatiotemporal Dynamics
A core challenge of sparse spatial sampling is observational ambiguity: multiple distinct global states can look identical when viewed through a few fixed sensor locations. As noted by Feng et al., this lack of information makes it difficult for standard neural networks to accurately reconstruct physical fields.
To break this ambiguity, algorithms must leverage the element of time. A single spatial snapshot may be ambiguous, but a sequence of sparse measurements carries the history of how information propagates through the system.
This principle is deeply rooted in time-delay embedding theory. Botvinick-Greenhouse et al. extended classical embedding theorems to a measure-theoretic framework, showing that continuous mappings between distant data points can be established by looking at probability distributions over time, making state reconstruction highly robust to noise and sparsity.
[High-Dimensional Physical System]
│
▼ (Sparse Spatial Sampling)
[Observational Ambiguity] ───► (Multiple states look identical)
│
▼ (Introduce Temporal History / Differential Equations)
[Unique State Reconstruction]
Modern deep learning architectures exploit this spatiotemporal continuity by embedding differential equations directly into neural networks:
- Neural ODEs: Lai et al. integrated physics-based modal analysis with Neural Ordinary Differential Equations to monitor large-scale structures with thousands of degrees of freedom using only a handful of sensors.
- Direct Space Modeling: Rather than encoding sparse data into a computationally heavy high-dimensional space, Song et al. bypassed the encoding step entirely. The authors combined Vision Transformers with Gated Neural ODEs to model the underlying dynamics directly within the low-dimensional measurement space, forecasting complex fluid flows and weather patterns from sparse points.
- Mobile Trajectories: When sensors are not fixed but moving (such as weather balloons or wearable tech), Williams et al. demonstrated that shallow recurrent decoder networks can leverage these arbitrary trajectories to reconstruct full-state dynamics from localized data streams.
Physics-Informed Constraints and Generative Priors
When a system is highly chaotic or the measurements are extremely sparse, purely data-driven interpolation fails. Here, the laws of physics act as a mathematical regularizer, constraining the search space to physically plausible solutions.
For example, in deep-well drilling, placing sensors kilometers below the surface at the drill bit is impossible. To overcome this, Darrin utilized a physics-informed dual-architecture neural network to infer hidden downhole dynamics solely from surface-level data, relying on the physical coupling of the drill string to transmit the underlying state information upward.
Where deterministic physics equations are difficult to solve under partial observation, generative machine learning models can act as powerful priors. Huang et al. introduced a generative framework that uses diffusion models to solve partial differential equations (PDEs) under extreme partial observation. The model learns the underlying distribution of valid physical states, allowing it to fill in massive gaps in spatial data while strictly adhering to the governing physical laws.
Topology and Network State Estimation
In large-scale infrastructure networks, such as power grids or water distribution systems, the sparsity constraint is topological rather than purely spatial. Operators cannot place sensors at every junction or node, leading to significant information scarcity (Picallo et al.; Kestler and Schmid).
However, physical networks are governed by conservation laws (e.g., Kirchhoff's laws or hydraulic principles) that dictate how states at one node affect neighboring nodes. This makes Graph Neural Networks (GNNs) an ideal mechanism for state reconstruction.
[Sensor Node] ──(Physics-Informed Edge)──► [Unmonitored Node] ──► [Predicted State]
Ashraf et al. designed spatial Graph Convolutional Networks that propagate information from a small percentage of monitored nodes across the network topology to estimate missing pressure values at unmonitored junctions. To ensure these predictions do not violate physical realities, Artelt et al. introduced a physics-informed GNN framework that incorporates data normalization based on hydraulic equations. This allows the model to act as a high-speed surrogate simulator, inferring global network states from sparse measurements up to 200 times faster than traditional hydraulic solvers.
Key Foundational Work
Collaborative Spectrum Sensing from Sparse Observations Using Matrix Completion for Cognitive Radio Networks (2010) — contributed 1 concept elements to this direction.
Intelligent Control of Auxiliary Ship Systems (2002) — contributed 1 concept elements to this direction.
Optimal estimation with limited measurements (2010) — contributed 1 concept elements to this direction.