Global Truncation Versus Spatial Thresholding

Two Ways to Control Complexity

A series estimator can reduce noise by discarding fine-scale terms. A global truncation rule makes one decision for an entire scale: keep every coefficient through level \(J\), and discard every coefficient above it. Wavelet thresholding instead makes a decision coefficient by coefficient. Because wavelets are localized in space, this allows an estimator to keep a prominent feature in one region while suppressing small coefficients elsewhere, even at the same scale.

The distinction is about how estimation complexity is controlled, not about whether one method is always preferable. When signal energy is distributed smoothly across locations and scales, a common cutoff can be effective. When a function has a small number of localized features, a global cutoff may retain many noisy coefficients just to preserve those features. But thresholding can also discard genuine small coefficients or retain noise, so its performance depends on the signal and the threshold.

For clarity, take levels \(j=0,1,\ldots\), and let \(N_j\) be the actual number of coefficients at level \(j\). In an ideal periodic basis, one often has \(N_j=2^j\); for other bases, including boundary-corrected bases, use the actual counts. Assume the basis is orthonormal and the target has finite squared norm, so \(\sum_{j,k}\theta_{jk}^2<\infty\). Risk below is integrated squared error, which equals squared coefficient error by orthonormality.

Risk of a Global Cutoff

A cutoff incurs variance on every retained coefficient and squared bias on every omitted coefficient. Thus its risk depends on the total energy at each scale, not on where within a scale that energy lies.

Proof. For \(j\leq J\), the coefficient error is \(Y_{jk}-\theta_{jk}=\tau Z_{jk}\), with expected square \(\tau^2\). For \(j>J\), the estimate is zero, so the squared error is \(\theta_{jk}^2\). Parseval's identity gives the squared function error as the sum of these squared coefficient errors. Taking expectations and summing gives the displayed expression. The sums are well-defined: the retained sum is finite for finite \(J\), and the omitted sum is finite because the target is square-integrable. \(\square\)

This identity gives the basic bias–variance tradeoff. For example, suppose \(N_j\leq C_N2^j\) and the coefficient energy \(S_j=\sum_k\theta_{jk}^2\) satisfies \(S_j\leq C_\theta 2^{-2sj}\) for some \(s>0\). Then \(\sum_{j\leq J}N_j\leq 2C_N2^J\), and \(\sum_{j>J}S_j\leq C_\theta 2^{-2s(J+1)}/(1-2^{-2s})\). Consequently, for constants \(A,B\) independent of \(J\), \(R(J)\leq A\tau^2 2^J+B2^{-2sJ}\). The two terms are balanced when \(2^J\) is of order \(\tau^{-2/(2s+1)}\), giving risk of order \(\tau^{4s/(2s+1)}\). If \(\tau^2=\sigma^2/n\), this is order \(n^{-2s/(2s+1)}\), up to constants. This calculation explains why a global cutoff can work well when energy declines regularly with scale.

Spatial Thresholding and Its Risk

A wavelet coefficient is attached to both a scale and a location. Thresholding uses the observed magnitude at that particular location: a large coefficient is kept, while a small one is set to zero. It therefore does not impose the all-or-nothing scale decision made by a global cutoff. Its risk is nonlinear in the signal coefficient, however, and is not simply a fixed variance term plus a fixed tail bias.

Derivation. The estimator retains \(Y\) on the events \(Z>a\) and \(Z<b\). Since \(\lambda\geq0\), we have \(b\leq a\), so these two events are disjoint, and the coefficient is discarded when \(b\leq Z\leq a\). On either retained event, the estimation error is \(Y-\theta=\tau Z\); on the discarded event, it is \(-\theta\). Therefore the risk is \(\tau^2\{E[Z^2\mathbf{1}(Z>a)]+E[Z^2\mathbf{1}(Z<b)]\}+\theta^2\{\Phi(a)-\Phi(b)\}\). Integration using \(\phi'(z)=-z\phi(z)\) gives \(\int_a^\infty z^2\phi(z)\,dz=a\phi(a)+1-\Phi(a)\) and \(\int_{-\infty}^b z^2\phi(z)\,dz=\Phi(b)-b\phi(b)\), proving the formula. \(\square\)

For a finite collection of orthonormal coefficients, the total thresholding risk is the sum of these single-coefficient risks, plus the squared energy of any unobserved or deliberately omitted coefficients. This sum makes spatial distribution consequential: coefficients at the same level can have very different individual risks under thresholding, even though global truncation treats them identically.

What the Comparison Establishes

Global truncation and spatial thresholding respond to different features of the coefficient array. Truncation responds to scale: if a fine level is retained, all its coefficients are included, whether large or negligible. Thresholding responds to individual observed magnitudes: coefficients at one scale may be treated differently. Accordingly, truncation risk is governed by the variance cost of all retained coordinates and the signal energy in omitted levels; thresholding risk accumulates coordinate-specific errors.

This is why an isolated local feature can be expensive for a global cutoff: retaining it may require retaining many other coefficients at the same and coarser scales. Conversely, if a signal's coefficients are numerous but individually moderate, a threshold can remove genuine signal along with noise. A favorable calculation for one localized signal is not a uniform dominance result. The best choice depends on the coefficient structure and on how the cutoff or threshold is selected.