Figure 1: Heisenberg box representing an atom φ_γ.

A time-frequency dictionary D={φγ}γ∈ΓD={φγ}γ∈Γ is composed of waveforms of unit norm ∥φγ∥=1∥φγ∥=1, which have a narrow localization in time and frequency. The time localization u of φ_γ and its spread around u, are defined by

Similarly, the frequency localization and spread of φ^γφ^γ are defined by

The Fourier Parseval formula

shows that 〈f,φγ〉〈f,φγ〉 depends mostly on the values f(t)f(t) and f^(ω)f^(ω), where φγ(t)φγ(t) and φ^γ(ω)φ^γ(ω) are nonnegligible , and hence for (t,ω)(t,ω) in a rectangle centered at (u,ξ)(u,ξ), of size σt,γ×σω,γσt,γ×σω,γ. This rectangle is illustrated by Figure 1 in this time-frequency plane (t,ω)(t,ω). It can be interpreted as a “quantum of information” over an elementary resolution cell. The uncertainty principle theorem proves (see Chapter 2) that this rectangle has a minimum surface that limits the joint time-frequency resolution:

Constructing a dictionary of time-frequency atoms can thus be thought of as covering the time-frequency plane with resolution cells having a time width σt,γσt,γ anda frequency width σω,γσω,γ which may vary but with a surface larger than one-half. Windowed Fourier and wavelet transforms are two important examples.

A windowed Fourier dictionary is constructed by translating in time and frequency a time window g(t)g(t), of unit norm ∥g∥=1∥g∥=1, centered at t=0t=0:

The atom gu,ξgu,ξ is translated by u in time and by ξ in frequency. The time-and-frequency spread of gu,ξgu,ξ is independent of u and ξ. This means that each atom gu,ξgu,ξ corresponds to a Heisenberg rectangle that has a size σt×σωσt×σω independent of its position (u,ξ)(u,ξ), as shown by Figure 2.

Figure 2: Time-frequency boxes (“Heisenberg rectangles”) representing the energy spread of two windowed Fourier atoms.

The windowed Fourier transform projects ff on each dictionary atom gu,ξgu,ξ:

It can be interpreted as a Fourier transform of ff at the frequency ξ, localized by the window g(t-u)g(t-u) in the neighborhood of u. This windowed Fourier transform is highly redundant and represents one-dimensional signals by a time-frequency image in (u,ξ)(u,ξ). It is thus necessary to understand how to select many fewer time-frequency coefficients that represent the signal efficiently.

When listening to music, we perceive sounds that have a frequency that varies in time. Chapter 4 shows that a spectral line of ff creates high-amplitude windowed Fourier coefficients Sf(u,ξ)Sf(u,ξ) at frequencies ξ(u)ξ(u) that depend on time u. These spectral components are detected and characterized by ridge points, which are local maxima in this time-frequency plane. Ridge points define a time-frequency approximation support λ of ff with a geometry that depends on the time-frequency evolution of the signal spectral components. Modifying the sound duration or audio transpositions are implemented by modifying the geometry of the ridge support in time frequency.

A windowed Fourier transform decomposes signals over waveforms that have the same time and frequency resolution. It is thus effective as long as the signal does not include structures having different time-frequency resolutions, some being very localized in time and others very localized in frequency.  Wavelets address this issue by changing the time and frequency resolution.

In reflection seismology, Morlet knew that the waveforms sent underground have a duration that is too long at high frequencies to separate the returns of fine, closely spaced geophysical layers. Such waveforms are called wavelets in geophysics. Instead of emitting pulses of equal duration, he thought of sending shorter waveforms at high frequencies. These waveforms were obtained by scaling the mother wavelet, hence the name of this transform. Although Grossmann was working in theoretical physics, he recognized in Morlet's approach some ideas that were close to his own work on coherent quantum states.

Nearly forty years after Gabor, Morlet and Grossmann reactivated a fundamental collaboration between theoretical physics and signal processing, which led to the formalization of the continuous wavelet transform (GrossmannM:84). These ideas were not totally new to mathematicians working in harmonic analysis, or to computer vision researchers studying multiscale image processing. It was thus only the beginning of a rapid catalysis that brought together scientists with very different backgrounds.

A wavelet dictionary is constructed from a mother wavelet ψ of zero average

which is dilated with a scale parameter s, and translated by u:

The continuous wavelet transform of ff at any scale s and position u is the projection of ff on the corresponding wavelet atom:

It represents one-dimensional signals by highly redundant time-scale images in (u,s)(u,s).

Multiscale Zooming

A wavelet dictionary is also adapted to analyze the scaling evolution of transients with zooming procedures across scales. Suppose now that ψ is real. Since it has a zero average, a wavelet coefficient Wf(u,s)Wf(u,s) measures the variation of ff in a neighborhood of u that has a size proportional to s. Sharp signal transitions create large-amplitude wavelet coefficients.

Figure 3: Heisenberg time-frequency boxes of two wavelets, ψu,sψu,s and ψu0,s0ψu0,s0. When the scale s decreases, the time support is reduced but the frequency spread increases and covers an interval that is shifted toward high frequencies.

Signal singularities have specific scaling invariance characterized by Lipschitz exponents. Chapter 6 relates the pointwise regularity of ff to the asymptotic decay of the wavelet transform amplitude |Wf(u,s)||Wf(u,s)| when s goes to zero. Singularities are detected by following the local maxima of the wavelet transform acrossscales.

In images, wavelet local maxima indicate the position of edges, which are sharp variations of image intensity. It defines scale–space approximation support of ff from which precise image approximations are reconstructed. At different scales, the geometry of this local maxima support provides contours of image structures of varying sizes. This multiscale edge detection is particularly effective for pattern recognition in computer vision (Canny:86).

The zooming capability of the wavelet transform not only locates isolated singular events, but can also characterize more complex multifractal signals having nonisolated singularities. Mandelbrot (Mandelbrot:82) was the first to recognize the existence of multifractals in most corners of nature. Scaling one part of a multifractal produces a signal that is statistically similar to the whole. This self-similarity appears in the continuous wavelet transform, which modifies the analyzing scale. From global measurements of the wavelet transform decay, Chapter 6 measures the singularity distribution of multifractals. This is particularly important in analyzing their properties and testing multifractal models in physics or in financial time series.

Orthonormal bases of time-frequency atoms remove all redundancy and define stable representations. A wavelet orthonormal basis is an example of the time-frequency basis obtained by scaling a wavelet ψ with dyadic scales s=2js=2j and translating it by 2jn2jn, which is written ψj,nψj,n. In the time-frequency plane, the Heisenberg resolution box of ψj,nψj,n is a dilation by 2j2j and translation by 2jn2jn of the Heisenberg box of ψ. A wavelet orthonormal is thus a subdictionary of the continuous wavelet transform dictionary, which yields a perfect tiling of the time-frequency plane illustrated in Figure 4.

Figure 4: The time-frequency boxes of a wavelet basis define a tiling of the time-frequency plane.

One can construct many other orthonormal bases of time-frequency atoms, corresponding to different tilings of the time-frequency plane. Wavelet packet and local cosine bases are two important examples constructed in Chapter 8, with time-frequency atoms that split the frequency and the time axis, respectively, in intervals of varying sizes.

Wavelet Packet Bases

Wavelet bases divide the frequency axis into intervals of 1 octave bandwidth. Coifman, Meyer, and Wickerhauser (CoifmanMW:92) have generalized this construction with bases that split the frequency axis in intervals of bandwidth that may be adjusted. Each frequency interval is covered by the Heisenberg time-frequency boxes of wavelet packet functions translated in time, in order to cover the whole plane, as shown by Figure 5.

As for wavelets, wavelet-packet coefficients are obtained with a filter bank of conjugate mirror filters that split the frequency axis in several frequency intervals. Different frequency segmentations correspond to different wavelet packet bases. For images, a filter bank divides the image frequency support in squares of dyadic sizes that can be adjusted.

Figure 5: A wavelet packet basis divides the frequency axis in separate intervals of varying sizes. A tiling is obtained by translating in time the wavelet packets covering each frequency interval.