Interactive demo · Nonparametric Statistical Inference · Chapter 9

Wavelet thresholding

The data Yi = r(i/n) + σεi are put through the orthonormal discrete wavelet transform. Each detail coefficient is the true coefficient plus N(0, σ2) noise, σ is estimated from the finest level by the median absolute deviation, and the coefficients are thresholded: small ones are set to zero, large ones kept (hard) or pulled towards zero by λ (soft). The universal threshold λ = σ̂√(2 log n) kills pure noise with high probability; SureShrink picks one threshold per level by minimising Stein's unbiased risk estimate. The lower panel shows every detail coefficient, level by level, with the threshold: the function is sparse in the wavelet basis, so few coefficients survive.

Function

Sample size n

Noise level σ

Wavelet

Rule

Threshold multiplier (universal rules)

Try these

Periodised orthonormal transform from level J = log2 n down to level j0 = 5 (32 scaling coefficients, never thresholded). The coefficients are those of the data vector, i.e. √n times the book's Djk (footnote 2 of Section 9.3), so the noise has standard deviation σ at every level. The symmlet 8 filter is the one printed in the book. Average squared error = (1/n)Σ(r̂(xi) − r(xi))2.

In the lecture notes