Neural controlled differential equations (Neural CDEs) evolve hidden states along continuous paths constructed by interpolating discrete observations. Sparse measurements leave the input between observation times undetermined, and admissible continuous paths consistent with the same data can produce different terminal states. We characterize this ambiguity for planar drivers with bounded speed. The set of compatible terminal states admits a leading-order zonotopic approximation determined by feasible signed areas and transported Lie brackets. Its projection onto a scalar prediction yields a computable, target-specific ambiguity certificate. We also study the number of measurements needed to reduce the remaining uncertainty. For nondegenerate planar CDEs, the optimal worst-case radius over deterministic adaptive sampling policies scales as N−1 when speed is bounded and N−2 when both speed and acceleration are bounded, where N is the number of observation intervals. Synthetic experiments assess the certificate using admissible continuous completions and examine the role of path regularity.
Figures & tables
Figure 1: Sampling-induced ambiguity in Neural CDEs. A discrete observation record S={(ti,xi)} admits multiple bounded-speed continuous paths X∈XL(S) (colored curves). The dashed line represents the piecewise-linear reference Xˉ . Only the first coordinate of the planar input is shown. Paths sharing the same observations may have different signed areas ai(X) between measurements. Theorem 4.1 approximates the terminal-state identified set by zˉT+K(S) , where zˉT=ZT(Xˉ) and K(S) is the zonotope generated by feasible signed areas and transported Lie brackets, defined in ( 4 ). The Hausdorff approximation error is O(S3+S22) . Theorem 4.2 gives a leading-order approximation to the scalar response identified set by the interval with endpoints y±=g(zˉT)±R2(S) ( δ=maxi(ti+1−ti) ), where the target-specific radius R2(S) is defined in ( 6 ). Theorems 5.1 and 5.2 establish minimax information-radius rates Θ(N−1) and Θ(N−2) under bounded speed and bounded speed plus acceleration, respectively. Here N is the number of final observation intervals, and the minimax radius RN is defined in ( 3 ). The rates hold under the nondegeneracy assumptions of the respective theorems. The plotted curves are schematic.
Figure 2: Certificate accuracy and path regularity. (a) Predicted diameter 2R2(S) against the nonlinear output spread between admissible circular-arc completions. The dashed line indicates equality. (b) Median leading certificate diameters under bounded-speed and bounded-acceleration assumptions as the maximum observation gap decreases. All widths are normalized by the corresponding target standard deviation.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Observations
2R2(S)
Darc
Median rel. error
5
5.775
5.401
5.49%
9
2.769
2.684
2.55%
17
1.358
1.334
1.79%
33
0.678
0.668
1.69%
Appendix
Table 1: Predicted diameter and feasible circular-arc spread (target SD), with per-case relative discrepancy.
Observations
Speed only
Acceleration aware
Searched spread
5
1.7844
0.4591
0.09966
9
0.8547
0.0869
0.00935
17
0.4174
0.0183
0.00090
33
0.2061
0.0042
0.00010
Appendix
Table 2: Median leading geometric diameters and numerically screened completion spread (target SD).
Policy
Radius AUC
Final / initial
One-step look-ahead oracle
3.626
0.391
Largest local ambiguity
3.633
0.387
Predicted midpoint reduction
3.633
0.387
Area budget only
3.749
0.401
Longest gap
3.823
0.400
Random gap
4.544
0.603
Appendix
Table 3: Median normalized certificate-radius area under six acquisitions (lower is better), and median final radius relative to its initial value.
5 observations
9 observations
Target
Model
Target
Model
Target
Signed area
0.2712
0.2903
0.0686
0.0705
Nonlinear CDE
0.2034
0.2154
0.0517
0.0557
Endpoint only
0.0058
0.0000
0.0012
0.0000
Appendix
Table 4: Median learned-model spread and true-target spread on the same selected completion pairs, in training-target standard deviations.
Neural controlled differential equations model continuous-time dynamics driven by observed data paths, especially for irregular and asynchronous sequences.
Neural Controlled Differential Equations (Neural CDEs) provide a powerful continuous-time framework for sequence modeling, yet the roughness of the driving control path often restricts their efficiency. Standard splines introduce high-frequency variations that force adaptive solvers to take excessively small steps, driving up the Number of Function Evaluations (NFE). We propose a novel approach to Neural CDE path construction that replaces exact interpolation with Kernel and Gaussian Process (GP) smoothing, enabling explicit control over trajectory regularity. To recover details lost during smoothing, we propose an attention-based Multi-View CDE (MV-CDE) and its convolutional extension (MVC-CDE), which employ learnable queries to inform path reconstruction. This framework allows the model to distribute representational capacity across multiple trajectories, each capturing distinct temporal patterns. Empirical results demonstrate that our method, MVC-CDE with GP, achieves state-of-the-art accuracy while significantly reducing NFEs and total inference time compared to spline-based baselines.
Many real-world systems evolve continuously, yet most machine learning models interpret time series as discrete sequences. Continuous-time approaches instead treat time series as samples from an underlying input path, a formulation that naturally accommodates irregularly sampled or oversampled data. Among these, Neural Controlled Differential Equations (NCDEs) are a maximally expressive class of models that parametrise a vector field using a neural network and evolve their hidden state by solving a dynamical system driven by the input path. NCDEs typically use a non-linear vector field, so their expressive power and continuous-time flexibility come at the cost of a forward pass that is both computationally expensive and inherently sequential, limiting their scalability and practical applicability. This thesis advances the training and scalability of NCDEs through three complementary contributions. First, building on neural rough differential equations, Log-NCDEs apply the Log-ODE method to efficiently approximate an NCDE's solution during training, improving both computational speed and empirical performance. Second, Linear NCDEs replace the non-linear vector field with a linear one, enabling closed-form solutions and parallel-in-time computation without sacrificing theoretical expressivity. Third, Structured Linear NCDEs use structured linear vector fields to further enhance efficiency while maintaining theoretical expressiveness and empirical performance. Collectively, these methods reduce the time per training step for an NCDE by up to three orders of magnitude while achieving state-of-the-art performance across diverse time series benchmarks.
Neural ODEs learn dynamics from trajectory losses, but their adjoint gradients lack the regressor-times-parameter-error structure on which Lyapunov analyses of online adaptation rest, so training on streaming data comes without stability guarantees. We show that this structure is in fact present: the adjoint gradient decomposes exactly into a positive semi-definite trajectory operator acting on the parameter error plus a nonlinear perturbation with explicit, horizon-dependent bounds. A quadratic Lyapunov function then certifies online Neural ODE training over sliding windows under computable gain and horizon conditions, and the same certificate extends to stored data: its drift branch recovers concurrent learning, and its trajectory branch yields NODE-CL, a stored-segment Gauss-Newton method built on batched forward sensitivities that needs no state-derivative estimates. On four DeepMind Control Suite domains, NODE-CL attains the lowest median prediction error on three under velocity measurement noise, where observer-based concurrent learning degrades by up to 8x; with clean measurements it is best on the pendulum and within a factor of 1.6 of the best stored-data baseline on the cartpole and reacher.
Omkar Sudhir Patil
Division of Electrical and Computer Engineering Louisiana State University Baton Rouge, LA 70803, USA