Function construct_toeplitz

model/functional/toeplitz.py:17–33

Explicit construction of Krylov matrix [v A @ v A^2 @ v ... A^{n-1} @ v] where A = Z_f. This uses vectorized indexing and cumprod so it's much faster than using the Krylov function. Parameters: v: the starting vector of size n or (rank, n). f: real number Returns:

(v, f=0.0)

def construct_toeplitz(v, f=0.0):
    """Explicit construction of Krylov matrix [v  A @ v  A^2 @ v  ...  A^{n-1} @ v]
    where A = Z_f. This uses vectorized indexing and cumprod so it's much
    faster than using the Krylov function.
    Parameters:
        v: the starting vector of size n or (rank, n).
        f: real number
    Returns:
        K: Krylov matrix of size (n, n) or (rank, n, n).
    """
    n  = v.shape[-1]
    a = torch.arange(n, device=v.device)
    b = -a
    indices = a[:, None] + b[None]
    K = v[..., indices]
    K[..., indices < 0] *= f
    return K

def triangular_toeplitz_multiply_(u, v, sum=None):
    n = u.shape[-1]