11.2 Wavelet and Scaling Coefficients

Wavelet and Scaling Coefficients

Wavelet and scaling coefficients are the two outputs you get from the Discrete Wavelet Transform. Scaling coefficients capture the low-frequency, smooth trends in a signal, while wavelet coefficients capture the high-frequency details. Together, they give you a complete multi-resolution picture of the signal, and understanding what they represent is essential for tasks like compression, denoising, and feature extraction.

Computing these coefficients involves passing the signal through matched pairs of low-pass and high-pass filters, then downsampling. The wavelet family you choose and the number of decomposition levels you use will shape the coefficients you get and how well they represent different signal features.

Wavelet Coefficients: Meaning and Significance

Physical Interpretation of Wavelet Coefficients

Wavelet coefficients measure the correlation between your signal and the wavelet function at specific scales and time positions. A coefficient at a given scale and translation tells you how closely that portion of the signal resembles the wavelet shape at that resolution.

• Magnitude indicates the strength of the match between the signal and the wavelet at that scale and location.
  • Large coefficients point to strong wavelet-like features in the signal, such as sharp edges or transient events.
  • Small or near-zero coefficients mean the signal has little energy matching the wavelet at that point.
• Sign indicates the phase relationship between the signal and the wavelet.
  • Positive coefficients mean the signal locally matches the wavelet shape (positive correlation).
  • Negative coefficients mean the signal is locally inverted relative to the wavelet shape.

Significance of Wavelet Coefficient Location and Scale

The position of significant wavelet coefficients in the time-scale plane directly tells you where and at what resolution important signal features occur.

• Time location of a coefficient corresponds to where a feature sits in the signal (a spike, a discontinuity, a burst of activity).
• Scale relates to frequency content:
  • Finer scales capture high-frequency details like sharp transitions or rapid oscillations.
  • Coarser scales capture low-frequency behavior like slow drifts or baseline trends.

Scaling coefficients at each level represent a progressively coarser approximation of the signal. They capture the overall trend or average behavior and are useful for analyzing global characteristics like baseline wander or slow variations.

Computing Wavelet Coefficients

Discrete Wavelet Transform (DWT) Process

The DWT decomposes a signal into wavelet and scaling coefficients through a hierarchical filter bank structure. Here's how it works at each decomposition level:

1. Pass the input (the original signal at level 1, or the previous level's scaling coefficients at deeper levels) through a high-pass filter derived from the wavelet function. This produces the wavelet (detail) coefficients for that level.
2. Pass the same input through a low-pass filter derived from the scaling function. This produces the scaling (approximation) coefficients for that level.
3. Downsample both outputs by a factor of 2 (keep every other sample). This halves the number of coefficients at each level.
4. Repeat steps 1–3 on the new scaling coefficients to decompose further into the next coarser level.

Each iteration peels off one layer of high-frequency detail (wavelet coefficients) while passing the remaining low-frequency content (scaling coefficients) to the next stage.

Wavelet Function and Filter Choice

The wavelet family you select determines the filter properties and, in turn, the characteristics of the resulting coefficients.

• Different families offer different trade-offs in symmetry, smoothness, and vanishing moments. Common choices include Haar, Daubechies, Symlets, and Coiflets.
• Vanishing moments control how well a wavelet can represent polynomial trends. A wavelet with $N$ vanishing moments will produce zero coefficients for any polynomial of degree less than $N$. More vanishing moments mean better representation of smooth signal components, but longer filters.
• The number of decomposition levels depends on what frequency resolution you need and the signal length. Each level halves the resolution, and the maximum number of levels is constrained by the requirement that the signal length be divisible by $2^{L}$, where $L$ is the number of levels.

Wavelet Coefficients: Scale and Translation

Relationship Across Scales

Wavelet and scaling coefficients at different scales partition the signal's information into distinct frequency bands.

• Finer scales correspond to higher-frequency bands and capture short-duration variations.
• Coarser scales correspond to lower-frequency bands and capture longer-duration trends.

The decomposition is hierarchical: scaling coefficients at scale $j$ serve as the input for computing both the wavelet and scaling coefficients at the next coarser scale $j+1$. This cascading structure is what makes the DWT computationally efficient.

Because of the factor-of-2 downsampling at each level, the number of coefficients halves with each coarser scale. This reduces both computation and storage while preserving the information needed for perfect reconstruction.

Translation and Localization

At each scale, the wavelet and scaling functions are shifted (translated) across the signal to probe local features at different time positions.

• Wavelet coefficients at a given scale capture exactly the high-frequency detail that is lost when moving from that scale's resolution to the next coarser one. They provide a localized record of transient events, edges, and singularities.
• Scaling coefficients at a coarser scale can be perfectly recovered from the wavelet and scaling coefficients at the next finer scale using the inverse DWT (IDWT). The IDWT upsamples by 2, applies reconstruction filters, and sums the results. Repeating this across all levels recovers the original signal exactly.

Wavelet Coefficient Sparsity vs Energy Compaction

Sparsity of Wavelet Coefficients

Sparsity means that most wavelet coefficients are small or zero, with only a few carrying significant energy. This property is what makes wavelets so useful for compression and denoising: you can ignore the small coefficients without losing much information.

How sparse a representation you get depends on the signal:

• Smooth signals with few discontinuities (sinusoids, low-order polynomials) tend to produce very sparse wavelet representations. Most of their energy lands in just a handful of large coefficients.
• Signals with many abrupt changes, edges, or impulses (square waves, delta-like spikes) produce less sparse representations because more coefficients are needed to capture those localized features.

Energy Compaction Property

Energy compaction is closely related to sparsity. It refers to how effectively the transform concentrates the signal's total energy into a small subset of coefficients.

• Wavelets with more vanishing moments generally provide better energy compaction for smooth signals, since they can zero out polynomial components more efficiently.
• Daubechies wavelets are particularly well-known for strong energy compaction properties.

To exploit sparsity and energy compaction in practice, you can apply:

• Thresholding: set coefficients below a chosen magnitude to zero. This removes noise (denoising) and reduces the number of stored values (compression).
• Coefficient selection: retain only the largest coefficients and discard the rest, keeping a compact but accurate signal representation.

These techniques underpin real-world applications like JPEG 2000 image compression, audio denoising, and biomedical signal feature extraction.