The ICWT reconstructions have errors on the order of 10% or more relative to input waveforms. This seems to be a generic feature of Torrence and Compo (1998) style ICWTs.
The scale choosing conventions of T&C (sec 3f) are not followed in their own Fortran implementation. In a comment therein they suggest:
for accurate reconstruction ... set s0=dt (for Paul set s0=dt/4), and use a large "jtot"
This computes scales outside the usual Fourier range [1/Tobs, 1/(2*dt)].
- Is it possible to achieve FFT level accuracy?
- What choices of
dj, s0, M, etc maximize accuracy?
- How do these choices affect computational efficiency?
The ICWT reconstructions have errors on the order of 10% or more relative to input waveforms. This seems to be a generic feature of Torrence and Compo (1998) style ICWTs.
The scale choosing conventions of T&C (sec 3f) are not followed in their own Fortran implementation. In a comment therein they suggest:
This computes scales outside the usual Fourier range
[1/Tobs, 1/(2*dt)].dj,s0,M, etc maximize accuracy?