Mexican hat wavelet

Key formulae: Mexican hat wavelet Gaussian lter The scaled wavelet Transform Compatibility condition Inverse transform Multiresolution Local reconstruction Parseval (Plancherel)
t2 1 F (t) = e 4 2

2 (, t) =

dF d2 F = d dt2

u2 (, t) =

u(t )2 (, t t ) dt
2 u2 = t u2

u(t) = 4

u2 (, t ) 2 (, t t ) d dt U = u2

u(t) =
0

1 2 u dx = 2

ds s

2 U 2 dx

The normalization constant cp si is absorbed in the denition of 2 ; the scale corresponds to a frequency = 1/2 ; logarithmic scale integration for inverse and Parseval are natural; and the compatibility equation facilitates analytical manipulations.

Starting from the normalized Gaussian lter

t2 1 F (t) = e 4 2

(1)

in which = a2 /2, its derivative with respect to (second time derivative) gives the Mexican hat wavelet: 2 (, t) = 2 t2 d2 F dF . = ( 2 )F (t) = d 4 dt2 (2)

It is a dierence of Gaussian lters of dierent scales (band-pass ltering), divided by the scale dierence. Rather than energy normalization, this variant favors the relation to Gaussian ltering and the simple formulae below, including the compatibility equation. The Fourier transform of 2 is 2 (, ) = 4 2 2 e4
2 2

(3)

For a signal u, the Mexican hat wavelet transform is written as u2 (, t) = u(t )2 (t t , ) dt = d2 (F u) dt2 (4)

Then the energy of the signal is given by | u |2 dt = 2 2 | u2 |2 d dt (5)

The inverse transform reconstructs the signal from its wavelet coecients by the relation u(t) = 4 u2 (, t ) 2 (t t , ) d dt . (6)

For the Mexican hat, a simpler exact alternative exists in the form

u(t) =

u2 (, t)d =
0 0

U (, t)

(7)

when we introduce the multiscale decomposition of u as U (, t) = u2 (, t) (8)

Because U can be interpreted as the u at scale and point x, and because of its relation to energy (see below), I prefer it for plotting purposes. A comparison with the orthodox presentation appears in the web pages. For plotting purposes, the conversion from to an equivalent frequency is needed. Taking the wavelet transform of cos(2t), two simple conversions can be adopted: the peak of the compensated energy spectrum corresponds to = 1/ 8, or the centroid of the spectrum coincides with = 1/2. The second alternative is adopted here, with the largest equal to the Nyquist frequency of our discrete signal. 2

amplitude (arb. units)

0

10

15

20

25

30

35

40

45

50

time(s)

10

freq(Hz)
10
1

10

15

20

25

30

35

40

45

50

time(s)

Figure 1: Mexican hat wavelet map U (bottom) and signal (top). The example signal and its Mexican hat wavelet map are shown on Fig. 1. The largest wavelet coecients (red for maxima, blue for minima) identify individual extrema of the signal. The near-periodicity of one contribution to the signal is visible in the energetic contributions along a horizontal strip of the map. The noise is intermittently distributed in the top third of the map, corresponding to larger frequencies. Additional features are discussed in connection with the Morlet transform. Parseval relation, spectra The Parseval/Plancherel formula takes the form 1 2 u dx = 2 2s u2 ds dx = 2 U2 ds dx. s (9)

Again, we see that U (squared) is proportional to the energy density (per unit time and per logarithmic frequency) of the signal. Cone of inuence and end eects For a wavelet located at a xed time, increasing its scale from some very small value gradually brings a larger part of the signal into the wavelets view, generating the cone of inuence of the wavelet. Signal values within the cone of inuence aect the wavelet coecient at that scale. If the wavelet is located near either end of a signal, it will see whatever information is located beyond the end-points, or the lack of information: the corresponding wavelet

coecients are of questionable value. Various techniques such as wraparound (implied periodicity of the signal), mirror symmetry or zero-padding do not entirely eliminate this problem, and results should be interpreted accordingly. Periodicity is used throughout the examples below. The cone of inuence is the envelope, for some suitably low threshold (2% of maximum in this case), of the wavelet coecients u of a Dirac function located at an end point. It is recommended that the cone of inuence of end-points be shown as a reminder that the wavelet coecients within the cone should be interpreted with caution.