Quantum Ml

Quantum Fisher Information and Natural Gradient

The quantum Fisher information defines the geometry of parameter space, enabling natural-gradient steps that often train quantum models faster than plain gradient descent.

Geometry of the parameter space

Plain gradient descent treats parameter space as flat Euclidean space, but the states a parameterized circuit produces do not change uniformly with the parameters: some directions move the state far, others barely at all. The quantum Fisher information matrix (also called the quantum geometric tensor) captures this by measuring how much the output state changes per unit change in each parameter and each pair of parameters.

Natural gradient

Kronos motion — parameter scan

The quantum natural gradient rescales the ordinary gradient by the inverse of the Fisher information matrix, so each step accounts for the local geometry. The update is theta_new = theta - eta * F^{-1} * grad, where F is the Fisher matrix. Steps then reflect distance in state space rather than parameter space, which typically means faster and more stable convergence, especially in landscapes with sharp curvature differences.

python

# Quantum natural gradient step (schematic)
import numpy as np
def qng_step(theta, grad, fisher, eta=0.1, reg=1e-3):
    F = fisher + reg * np.eye(len(theta))   # regularize for invertibility
    return theta - eta * np.linalg.solve(F, grad)

Estimating the matrix

Full Fisher information requires overlap circuits between shifted states and grows with the square of the parameter count.
A block-diagonal approximation, computed layer by layer, is far cheaper and often sufficient.
The diagonal alone gives a per-parameter learning-rate rescaling at minimal extra cost.

Why it helps with hard landscapes

Natural gradient can accelerate escape from flat or ill-conditioned regions because it stretches small-effect directions and compresses large-effect ones. It does not cure a true barren plateau, where gradients vanish exponentially and the Fisher information also becomes uninformative, but it noticeably improves training in landscapes that are merely ill-conditioned rather than flat.

Practical trade-off

The benefit is fewer optimization steps; the cost is more circuit evaluations per step to estimate the Fisher matrix. On NISQ hardware this trade-off is worth making when each step is expensive and the landscape is curved, particularly with the block-diagonal approximation. It is a standard tool in the hybrid training toolkit alongside the parameter-shift rule.