Non-negative Tucker (hierarchical) muscle-synergy decomposition

A hierarchical tensor synergy model: factorises the muscle x time x trial EMG tensor into non-negative spatial, temporal and trial factor matrices plus a CORE tensor that weights every combination of them. Unlike PARAFAC (muscleSynergyTensor()), the core is not diagonal, so a spatial synergy can pair with several temporal synergies – capturing interactions PARAFAC cannot.

Usage

tuckerSynergy(
  trials,
  ranks,
  max_iter = 300L,
  tol = 1e-06,
  n_restart = 3L,
  seed = NULL
)

Arguments

trials: A [muscle, time, trial] array or a list of muscle x time matrices.
ranks: Length-3 factor ranks c(spatial, temporal, trial).
max_iter, tol: Iterations and convergence tolerance.
n_restart: Random restarts (default 3).
seed: Optional RNG seed.

Value

a tucker_synergy: spatial (muscle x P), temporal (time x Q), trial (trial x R), core (P x Q x R), vaf, iterations.

References

Kim YD, Choi S (2007) non-negative Tucker; Tucker LR (1966).

Examples

set.seed(1)
A <- matrix(stats::runif(6*2),6,2); B <- matrix(stats::runif(40*2),40,2); C <- matrix(stats::runif(8*2),8,2)
G <- array(stats::runif(2*2*2), c(2,2,2))
X <- reconstructTucker(G, A, B, C)
tuckerSynergy(X, ranks = c(2, 2, 2), n_restart = 1)$vaf
#> [1] 0.9994916