Anna Zhou

Laguerre–Volterra
models

Represent input history with a small set of filters, then model nonlinear dynamics through products of their outputs.

Explore the Laguerre basis

At zero pole, each mode is a unit impulse at its own lag. As the pole approaches one, the first mode decays more slowly.
Mode 0Mode 1Mode 2Mode 3

At a = 0, mode k is a unit impulse at lag k. Increase a to see the first mode retain memory over more samples.

Laguerre memory

The basis sequence ℓk[m] is the causal impulse response of:

Lk(z) = √(1 − a²)/(1 − az−1) · [(z−1 − a)/(1 − az−1)]k,   |a| < 1.

Mode zero is a normalized exponential, ℓ0[m] = √(1 − a²)am. The modes are orthonormal over the full lag axis. A finite lag window can truncate their tails.

Volterra interactions

Let vt = [u[t], …, u[t − M + 1]]T. A second-order Volterra model combines linear convolution with products of delayed samples:

ŷ[t] = h0 + h1Tvt + vtTH2vt.

The kernel entry H2[i,j] weights the interaction between lags i and j. A cross-term coefficient of 0.30 corresponds to two symmetric entries of 0.15. Higher orders use products of more delayed samples.

Laguerre–Volterra expansion

Filter the history with Bkm = ℓk[m] to obtain xt = Bvt. Approximate the quadratic kernel as H2 ≈ BTQB:

ŷ[t] = b + cTxt + xtTQxt.

With fixed filters, the nonlinear model is linear in its readout coefficients. Using K basis functions replaces M raw lags per input channel with K filter outputs.

Example in robotics

For AntMaze dynamics, apply the filters to each of the eight action channels. This produces 8K action-history features. A vector-valued readout can combine them with state features to predict the next state increment. Quadratic products can capture interactions across joint commands and temporal patterns.

References