Block Thresholding of Wavelet Coefficients

Why Make a Decision About a Block?

Wavelet representations can describe a function with many small coefficients and a smaller number of important ones. Termwise thresholding makes a separate retain-or-delete decision for every coefficient. That can work well when the signal is sparse, but it ignores a useful form of structure: coefficients at nearby locations and a common resolution may carry information about the same local feature. A block decision pools their observed energy before deciding whether to retain them.

Pooling is not automatically better. It can make a collection of individually modest but jointly informative coefficients easier to detect; it can also dilute a single strong coefficient by requiring the whole block to pass a threshold. To make these trade-offs precise, begin with the Gaussian sequence model, in which the coefficient noise and the block risk can be analyzed directly.

A common threshold is proportional to the total noise energy expected in the block. For a constant \(\lambda>0\), define

Here \(\mathbf{1}\{\cdot\}\) is the indicator of the event in braces, and \(\|v\|^2=\sum_{i\in B}v_i^2\). This is a group decision: if the block passes, even a small observed coefficient is retained; if it fails, even the largest observed coefficient is deleted. In a wavelet basis, blocks are often formed from neighboring locations at a fixed resolution. The sequence-model calculations below apply directly when the coefficient errors are independent Gaussian variables with common variance \(\tau^2\).

Worked Decisions and Their Scale

Risk of a Block Decision

Use squared-error loss for a block, \(\|\widehat{\theta}_B-\theta_B\|^2\), and let \(r=\|\theta_B\|^2\) denote its true signal energy. The next identity separates the two possible errors: noise retained when a block is kept, and signal discarded when it is deleted.

where \(Z_B=(Z_i:i\in B)\). In particular, the risk is bounded above by \(L\tau^2+r\).

Proof. On \(A\), the estimator equals \(Y_B\), so its error is \(Y_B-\theta_B=\tau Z_B\). On \(A^c\), it equals zero, so its error is \(-\theta_B\), with squared norm \(r\). Splitting the expectation over these disjoint events gives the identity. Since \(\mathbf{1}_A\leq1\) and \(E\|\tau Z_B\|^2=L\tau^2\), the first term is at most \(L\tau^2\), while the second is at most \(r\). \(\square\)

An idealized blockwise oracle would know \(\theta_B\) and choose in advance whether to retain the whole block or delete it. Retaining it has expected squared-error risk \(L\tau^2\); deleting it has risk \(r\). Thus the oracle risk is \(\min(r,L\tau^2)\). The data-driven threshold does not know which choice is better, and its extra difficulty is deciding from noisy observations. In particular, the preceding upper bound is a basic decomposition, not a claim that the threshold estimator always attains the oracle risk.

Controlling False Retention of a Null Block

A useful guarantee is available when the entire block is null. It shows explicitly how block length and threshold multiplier affect the expected cost of retaining noise.

Proof. Under the null, set \(S=\|Y_B\|^2/\tau^2\). Then \(S\) is chi-squared with \(L\) degrees of freedom and \[ E\|\widehat{\theta}_B\|^2 =\tau^2 E[S\mathbf{1}\{S>\lambda L\}]. \] For any \(s>0\), on \(S>\lambda L\) we have \(1\leq e^{s(S-\lambda L)}\). Therefore \[ E[S\mathbf{1}\{S>\lambda L\}] \leq e^{-s\lambda L}E[Se^{sS}]. \] The chi-squared moment-generating function is \(E[e^{sS}]=(1-2s)^{-L/2}\) for \(0<s<1/2\). Differentiating this expression gives \[ E[Se^{sS}]=L(1-2s)^{-L/2-1}. \] Choose \(s=(1-1/\lambda)/2\), which lies in \((0,1/2)\) because \(\lambda>1\). Then \(1-2s=1/\lambda\), and \[ e^{-s\lambda L}E[Se^{sS}] = e^{-(\lambda-1)L/2}L\lambda^{L/2+1} = L\lambda\exp\!\left\{-\frac{L}{2}(\lambda-1-\log\lambda)\right\}. \] Multiplying by \(\tau^2\) proves the bound. \(\square\)

When Grouping Helps—and When It Hurts

For a block, the observed energy aggregates contributions across coordinates. If many coefficients in a block are moderately sized, each may be hard to distinguish from noise alone, while their combined energy can cross a group threshold. For a signal with \(L\) equal coefficients \(a\), the signal energy is \(La^2\); the noise energy has mean \(L\tau^2\). This accumulation is the basic reason a grouped rule can detect a spatially distributed feature that termwise rules miss. It is not a guarantee of lower risk for every signal: the decision still depends on the random observed energy and on the chosen threshold.

The corresponding weakness is most apparent for a sparse block. A group can fail because the energy is insufficient, even when one coefficient is strong enough to be retained by a termwise rule. A group can also pass because several noisy coordinates collectively produce a large energy; once it passes, noise in every coordinate is retained. Block construction and threshold calibration therefore encode assumptions about how signal energy is distributed.

Practical Interpretation and Common Mistakes

In wavelet regression, an orthonormal basis makes squared \(L^2\) error correspond to a sum of squared coefficient errors. Consequently, block risk calculations can be added across disjoint blocks when the coefficient model and loss have the required orthonormal structure. The Gaussian sequence results above rely on independent, equal-variance Gaussian coefficient errors; if the errors are correlated or have unequal variances, the null energy distribution and the calibration must be reconsidered. Blocks should be chosen to reflect plausible local dependence or clustering in the signal, rather than solely to simplify notation.

• Confusing the energy threshold with a norm threshold. The rule compares \(\|Y_B\|^2\) with \(\lambda L\tau^2\). Equivalently, it compares \(\|Y_B\|\) with \(\sqrt{\lambda L}\tau\), not with \(\lambda L\tau^2\).
• Assuming grouping uniformly improves risk. Grouping can help dense or clustered signals and hurt isolated signals. Its benefit depends on the signal pattern, noise, block length, and threshold.
• Interpreting the null bound as a selection probability. The theorem bounds expected retained squared noise, \(E[\|\widehat{\theta}_B\|^2]\) under a null block. It does not directly report \(P(A)\).
• Using the Gaussian calibration under a different noise model without adjustment. The chi-squared calculation uses independent Gaussian coordinates with common variance. Dependence, heteroskedasticity, or non-Gaussian errors change the relevant tail behavior.