Struct DiffFftPlan 

pub struct DiffFftPlan<T: Float> { /* private fields */ }

Differentiable FFT plan for automatic differentiation.

Wraps a standard FFT plan and provides differentiation capabilities.

Implementations

impl<T: Float> DiffFftPlan<T>

pub fn new(size: usize) -> Option<Self>

Create a new differentiable FFT plan.

pub fn forward(&self, input: &[Complex<T>], output: &mut [Complex<T>])

Execute forward FFT.

pub fn inverse(&self, input: &[Complex<T>], output: &mut [Complex<T>])

Execute inverse FFT (normalized).

pub fn forward_dual( &self, input: &[DualComplex<T>], ) -> (Vec<Complex<T>>, Vec<Complex<T>>)

Compute forward FFT with its Jacobian-vector product (forward mode AD).

Given input x and tangent dx, computes:

• y = FFT(x)
• dy = FFT(dx) (the directional derivative)

pub fn backward(&self, grad_output: &[Complex<T>]) -> Vec<Complex<T>>

Compute gradient of a scalar loss with respect to FFT input (backward mode AD).

Given the gradient of loss w.r.t. FFT output (grad_output), computes the gradient w.r.t. FFT input.

For FFT: y = F·x where F is the DFT matrix F[k,n] = exp(-2πi·k·n/N) The adjoint F^H satisfies: <v, F·x> = <F^H·v, x> F^H = conj(F^T) = IFFT_unnormalized = N · IFFT Gradient: ∂L/∂x = F^H · ∂L/∂y = conj(FFT(conj(∂L/∂y)))

pub fn backward_inverse(&self, grad_output: &[Complex<T>]) -> Vec<Complex<T>>

Compute gradient of a scalar loss with respect to IFFT input.

For normalized IFFT: y = (1/N)·Fᴴ·x The adjoint of (1/N)·Fᴴ is (1/N)·F Gradient: ∂L/∂x = (1/N)·F · ∂L/∂y = (1/N)·FFT(∂L/∂y)

pub fn size(&self) -> usize

Get the transform size.