Identifiable Decomposition of Submovements in Human Hand Trajectories

Submovement-Identifiable Decomposition (Sub-ID) proposes a novel method to decompose human hand trajectories into discrete submovements by using spatiotemporal kernel correlation as an identifiability criterion, allowing it to recover ground-truth parameter distributions in complex, long-horizon movements where existing methods fail.

The Gist

Sub-ID embeds a spatiotemporal kernel correlation between primitive pairs as an identifiability criterion within its adaptive-ridge regularization, biasing the optimiser toward low-correlation solutions to recover ground-truth parameter distributions at primitive counts where exact methods are intractable.

Core Methodology and Framework

The Sub-ID framework is a hierarchical decomposition method designed to balance heuristic initialization with global optimization, operating in both the velocity and spatial domains. The pipeline consists of three main stages:

1. Detect (speed): Heuristic Peak-Based Initialization, which identifies dominant speed transients through peak detection to provide initial estimates for onset times and durations.

2. Grow (speed): Greedy Residual Refinement, an iterative process that minimizes the residual error between the observed velocity profile and the current basis reconstruction by adding new candidate primitives based on a score maximization criterion.

3. Refine (position): Adaptive-Ridge Optimization, which performs the final optimization in the spatial domain to reconstruct the trajectory by minimizing positional error using Ridge Regression to prevent redundant scaling vectors and multicollinearity.

Identifiability Criterion and Regularization

The core innovation is the use of a spatiotemporal kernel correlation, defined as ρi, j = cos(ψi, j)· ρ(∆t), where ρ(∆t) is the off-diagonal entry of the Gram matrix above. This criterion formalizes when a pair of primitives is distinguishable; low values confirm primitives are distinguishable, while values approaching 1 signal information-theoretic impossibility (the singular regime). Sub-ID constrains tractability through heuristic initialization and adaptive-ridge regularization that biases the optimiser toward low-ρi, j solutions. When two primitives are separated by less than a minimum onset interval, their basis functions become nearly collinear, and the inverse problem loses its unique solution.

Performance on Synthetic Data

The method was validated on synthetic trajectories generated with known ground truth (GT) distributions at three primitive counts (K = 3, 8, and 15). Sub-ID consistently replicates the original distributions across all scenarios, whereas baseline methods exhibit significant parameter shifts or convergence failure as temporal density increases. The kernel correlation of the recovered decomposition tracks the ground truth: even at K = 15, critical overlaps (> 0.8) barely reach 3.00%, and the maximum correlation peaks at ρ = 0.45, well clear of the singular regime where identifiability is lost.

Generalization and Real Data Application

Sub-ID demonstrates success in more complex settings:

1. In heterogeneous (bimodal) distributions, it successfully identifies the dual nature of the signal by capturing clear separation in durations, amplitudes, and refractory periods without modes merging.

2. In real three-dimensional tasks (e.g., PushT 3D), Sub-ID extracts an average of K = 14.62 primitives with a mean overlap of just 0.0549 and keeps critical kernel correlations to a minimal 2.17%, ensuring each primitive resolves a distinct kinematic event without mathematical ambiguity, unlike SSSUMO which extracts approximately twice the primitives (K = 49.8) with maximum overlap of 1.0 due to collinearity and redundancy.

Conclusion on Motor Control Research

Sub-ID turns submovement decomposition from a descriptive fit into a measurement with a known operating range, enabling the extraction of physiologically grounded primitives for motor control research and imitation learning by identifying when decomposition is mathematically possible at all. The method scales to real three-dimensional, long-horizon movements where existing methods drift or over-segment. It provides a clear path forward by using the Sub-ID method to study how the nervous system assembles submovements into purposeful action.

Limitations and Future Directions

A caveat is that ρi, j depends on recovered direction vectors, which are themselves estimated outputs of the decomposition; thus, it should be treated as necessary but not sufficient evidence of identifiability until estimated directions are validated against synthetic ground truth. Future work may involve revising the coordinate frame for the final fit to better align with human perception and developing a model that treats measurement noise and motor noise together to close the gap between velocity-only methods and maximum-likelihood estimation. The method's success suggests that motor commands are identifiable in the investigated tasks, but further work is needed to rule out additional structures in other datasets.

Evaluation Metrics

Performance is quantified using kinematic reconstruction fidelity (Pos Error and Vel R2), structural complexity metrics (Primitive Rate K/s, Mean Overlap, and % Overlap > 0.

As a fastidious and diligent researcher, I have analyzed this paper, Identifiable Decomposition of Submovements in Human Hand Trajectories. The core innovation is the development of the Sub-ID method, which uses a spatiotemporal kernel correlation criterion to determine when submovement decomposition is mathematically possible.

Here are the specific improvements that can be made to AI systems based on this research:

)Improvement 1: Robust and Physiologically Grounded Motor Primitive Extraction

The current limitation of most existing methods (SSSUMO, Scattershot) is that they rely on heuristics or training biases, leading to spurious decompositions or loss of fidelity during long-horizon movements. Sub-ID provides a mathematically rigorous identifiability diagnostic using the spatiotemporal kernel correlation.

What the improved AI system can do:

1. Eliminate Spurious Decompositions: The system will only output submovements whose temporal and spatial relationships are statistically distinguishable, effectively filtering out artifacts caused by collinear bases or excessive temporal overlap (aliasing).

2. Extract Physiologically Plausible Primitives: By biasing the optimization toward low-correlation solutions, the AI can identify primitives that correspond to real physiological motor control units, rather than purely mathematical fits.

3. Handle Long-Horizon and High-Dimensional Data Robustly: The system will maintain millimetric precision (low positional error) even in complex 3D tasks (like PushT) where existing methods drift or over-segment into an excessive number of primitives.

)Improvement 2: Adaptive, Information-Theoretic Regularization for Model Selection

The Sub-ID framework incorporates adaptive ridge regularization based on the system's condition number, dynamically setting the penalty parameter to ensure numerical stability while preserving scientific value.

)Improvement 3: Domain-Agnostic Decomposition Across Spatial Dimensions

The paper explicitly decouples temporal parameter estimation from spatial direction estimation by solving for spatial scales independently via independent Ridge regressions along each dimension, while using a shared temporal basis matrix.

)Improvement 4: Enhanced Learning from Demonstration (LfD) and Skill Transfer

The framework provides a principled method to extract latent submovement parameters (duration, amplitude, asymmetry) directly from kinematic recordings, which can then be fed into advanced learning frameworks like Gaussian Mixture Models (GMMs), Gaussian Processes (GPs), or Dynamic Movement Primitives (DMPs).

)Improvement 5: Detection of Theoretical Limits and Failure Mode Diagnosis

The method explicitly detects when decomposition is mathematically impossible (the information-theoretic bound), providing a clear signal of failure rather than silent error.

Voluntary movements have long been hypothesised to be comprised of discrete primitives called submovements, as a descriptive model of human motor behaviour. However, existing methods scale poorly, and no principled method exists to determine whether a decomposition is informative. We propose a spatiotemporal kernel correlation between primitive pairs as an identifiability criterion. Submovement-Identifiable Decomposition (Sub-ID) embeds this criterion in its adaptive-ridge regularisation, biasing the optimiser toward low-correlation solutions. Identifiability is lost when primitives become collinear and recovered when they diverge spatially. On synthetic data, Sub-ID recovers ground-truth parameter distributions where existing methods fail; furthermore, when primitives overlap too heavily to be distinguished, the method explicitly detects this ambiguity rather than outputting misleading results. Sub-ID extracts submovements from real three-dimensional, long-horizon movements, a regime no prior method addresses. This method has the potential to identify physiologically grounded primitives for motor control research and imitation learning.