Unexpected complexity of everyday manual behaviors

The study investigates whether the human central nervous system simplifies hand control by restricting movements to a low-dimensional manifold composed of a limited set of synergies, as suggested by prior work using principal component analysis (PCA). Previous studies reported that a small number of principal components account for most of the variance in hand kinematics or muscle activations during tasks such as grasping, piano playing, and typing, leading to the inference that high-variance PCs reflect volitional control while low-variance PCs are noise. An alternative hypothesis is that hand control is intrinsically high-dimensional and that low-variance components capture subtle, volitionally controlled adjustments necessary for precise hand postures. This study aims to determine whether hand movements truly lie in a low-dimensional space by testing if low-variance PCs carry task- and condition-specific information during two manual behaviors: object grasping and American Sign Language (ASL) signing.

The concept of motor synergies proposes that coordinated muscle or joint patterns reduce control complexity. Evidence for hand synergies has come from PCA of kinematics and muscle activations showing that a few PCs explain most variance across tasks (e.g., grasping, typing, piano playing). However, debates persist regarding whether synergies reflect neural constraints or arise from biomechanics and task structure. Prior analyses often interpret low-variance PCs as noise. Some studies suggested grasping could be confined to about six dimensions, far fewer than the hand's many degrees of freedom. Optimal feedback control theory proposes low-dimensional control manifolds with noise shunted into orthogonal dimensions. Nonlinear dimensionality reduction methods have been proposed to uncover potentially curved, low-dimensional manifolds underlying kinematics, though previous work has sometimes found linear methods like PCA more parsimonious and reliable for hand kinematics. This study revisits these issues by explicitly testing the structure and informational content of low-variance dimensions across two tasks and by comparing linear and nonlinear dimensionality reduction approaches.

Participants and tasks: Eight right-handed adults (21–40 years) participated; all performed grasping and three with prior ASL knowledge performed ASL signing. Tasks were performed with the dominant right hand. Grasping: subjects started from a resting posture, then grasped, lifted, held (~1 s), and replaced an object, returning to the start position; five repetitions per condition. Twenty-five objects elicited 30 distinct grasps (some objects afforded multiple grasps, e.g., lightbulb by stem or bulb). ASL: subjects signed 26 letters and numbers 1–10, five repetitions per sign, starting from the same posture; no time limits. Measurement and preprocessing: Forty-one 4-mm hemispherical reflective markers were placed on the right hand/forearm (two per finger joint, two on ulnar, one on radial forearm). Fourteen infrared cameras (8 MP, 250 Hz hardware; data sampled at 100 Hz; Vicon MX-T Series) tracked 3D marker trajectories, which were labeled in Vicon Nexus. Inverse kinematics were computed in OpenSim using a musculoskeletal model (modified to include three rotational DOF for the first and fifth carpometacarpal joints), reconstructing 29 joint DOF (hand and three wrist DOF). Analyses focused on intervals from movement onset to 100 ms prior to object contact (grasping) and up to full ASL posture (signing). PCA and subspace similarity: PCA was applied separately to each subject’s kinematics to quantify variance explained by PCs. Cross-projection similarity was computed to compare task and subject subspaces: for N leading PCs of group A, variance explained in A (V1) was compared to variance explained when projecting A onto N leading PCs of group B (V2); similarity was the average of V2/V1 computed in both directions. Classification analyses: To assess condition specificity, linear discriminant analysis (LDA) classified objects or ASL signs from instantaneous joint angles taken 100 ms before contact (grasp) or at achieved ASL posture. Using full kinematics provided an upper bound. To test low-variance dimensions, PCs were progressively removed from highest to lowest variance and LDA was trained on the remaining PCs using leave-one-out cross-validation at the trial level (≈5 trials/condition). Nonlinear dimensionality reduction: Two nonlinear methods were applied: Isomap (MATLAB implementation; 29 nearest neighbors; variance per dimension from eigenvalues) and nonlinear PCA (NLPCA; autoencoder-based; MATLAB implementation; variance from reconstruction). Due to computational cost, kinematics were downsampled from 100 Hz to 20 Hz for nonlinear analyses; PCA was also run on downsampled data for comparison. Classification performance was reassessed while progressively removing high-variance dimensions for Isomap and NLPCA, analogous to the PCA procedure. Conditional noise simulations: To test whether condition-dependent noise could explain classification from low-variance PCs, denoised kinematics were created by selecting one trial per object and replicating it to eliminate within-condition variability, then reconstructing with only the first 10 PCs. Condition-dependent noise was added by sampling from zero-mean multivariate Gaussian distributions with object-specific covariance matrices generated by shuffling joint angle order and recomputing covariances per object; noise was rescaled so that overall classification performance matched the original (~95.5%). PCs and classification analyses were then repeated while sequentially removing PCs. The procedure was repeated five times with different random seeds. Correlations of PC scores across trials within and across objects were computed to assess repeatability of structure in individual PCs.

• Dimensionality by PCA: 3–5 PCs accounted for approximately 80% of variance; 8–11 PCs accounted for about 95% of variance in hand kinematics across tasks and subjects, replicating prior reports.
• Subspace similarity: Despite different behaviors, grasping and ASL yielded highly similar kinematic subspaces. For subjects performing both tasks (N=3), the leading 10 grasp PCs explained ~85% of variance in ASL kinematics and vice versa.
• Structure in low-variance PCs: Even PCs explaining less than 1% of variance (e.g., around the 20th PC) showed coordinated, condition-specific joint patterns. Trajectories along all PCs were more consistent within conditions than across conditions.
• Classification from reduced subspaces: Object and ASL sign identity could be classified well above chance even after removing most high-variance PCs; high performance persisted when the remaining PCs collectively accounted for less than 1% of total variance. Using LDA, accurate classification was achievable with only a handful of high-variance PCs, but importantly remained above chance using predominantly low-variance PCs.
• Nonlinear methods: PCA provided the most parsimonious representation by variance explained compared to Isomap and NLPCA. All three methods maintained substantial classification accuracy (>50%) even after removing the first 20 dimensions, which together accounted for more than 90% of variance, indicating that task-relevant information resides in low-variance dimensions not attributable to simple nonlinearity captured by Isomap or NLPCA.
• Conditional noise control: Simulated kinematics with condition-dependent noise confined true signal to 10 PCs. Although low-variance PCs in the simulations yielded slightly above-chance classification, performance declined much faster with PC removal than in real data, and low-variance PC scores lacked within-condition correlation structure (near zero), unlike measured kinematics. Thus, condition-dependent noise cannot account for the observed structured information in low-variance PCs.

The results challenge the view that low-variance principal components of hand kinematics are mere motor or measurement noise. Both grasping and ASL postures occupy subspaces that are highly similar across tasks and subjects, and low-variance dimensions contain reproducible, condition-specific structure supporting classification, even when those dimensions account for less than 1% of variance. Nonlinear manifold hypotheses, if present, were not supported by Isomap or NLPCA, which did not yield more efficient representations or concentrate task-relevant information into fewer dimensions. Simulations showed that condition-dependent noise can only modestly elevate classification from low-variance PCs and fails to reproduce the within-condition consistency observed in real kinematics. Collectively, these findings suggest that everyday hand control uses a high-dimensional manifold with structured contributions from many dimensions, including those of low variance. Rather than being restricted to a low-dimensional synergy space, hand postures may reflect a bounded repertoire within a high-dimensional subspace, consistent with known biomechanical constraints and emerging views of neural manifold constraints in motor cortex.

This study demonstrates that although a small number of principal components explain most variance in hand kinematics, low-variance components also carry reliable, task-dependent information, indicating volitional control over many dimensions. The similarity of subspaces across grasping and ASL, the persistence of classification accuracy after removing high-variance dimensions, and the failure of nonlinear reductions and conditional-noise simulations to account for these effects converge on the conclusion that hand control is higher dimensional than previously appreciated. Future work should: (1) examine additional manual behaviors, especially those involving object contact, to understand how contact shapes kinematic manifolds; (2) relate kinematic dimensionality to underlying neural population activity and constraints in motor cortex; (3) explore other manifold-learning techniques and larger datasets to test for alternative nonlinear structures; and (4) investigate implications for prosthetic and robotic hand control leveraging high-dimensional strategies.

• Task scope: Movements were analyzed prior to object contact (grasping) and at achieved ASL posture; contact-induced postures and force interactions were not included and may alter manifold structure.
• Sample size and generalizability: Only eight participants (three for ASL), potentially limiting generalization across populations and skill levels.
• Methodological assumptions: Dimensionality reduction relied on PCA and two nonlinear methods (Isomap, NLPCA); a low-dimensional nonlinear manifold not captured by these techniques may still exist. Nonlinear analyses required downsampling, which could affect fine temporal structure.
• Measurement constraints: Motion capture and inverse kinematics modeling (including model modifications) may introduce biases or errors. Classification used LDA, which assumes linear separability and Gaussian class structure.
• Simulation design: Conditional-noise simulations, while informative, make specific assumptions (e.g., Gaussian noise, covariance construction) that may not capture all forms of biological variability.