The Eigenspaces of the fast Walsh Hadamard transform
For the normalized Hadamard transform (orthogonal, symmetric, leaves vector magnitude unchanged and self-inverse), you can say:
It has exactly two eigenspaces: one for eigenvalue and one for eigenvalue .
Given a vector x:
u =(I+H)x
v = (I-H)x
(I+H)u = u
(I-H)v = -v
x = 0.5(u+v)
Hx = 0.5(u-v)
That’s actually quite special, and it connects to several interesting areas in mathematics:
-
Two-point spectrum: rare for nontrivial transforms.
-
Perfect involution: exactly, no scaling.
-
Orthogonal split of the entire space into just two halves, making projection filtering trivial.
-
Group-theoretic structure: eigenspaces correspond to symmetry classes of the Boolean cube.

Comments
Post a Comment