It is commonly known that the dyadic structure of wavelet expansions result
s in both time- and frequency-translation sensitivity of an input signal. W
e develop the first efficient method to reduce frequency-alignment sensitiv
ity by introducing a wavelet packet decomposition that is invariant to freq
uency shifts of a signal. A frequency-shifted wavelet packet library is pre
sented, and an efficient best-basis algorithm is developed to determine the
best signal representation among all frequency shifts of a signal. The alg
orithm computes all frequency-shifted coefficients in O(N-2) operations, fo
llowed by an efficient depth-first tree search of the same complexity.