Expressivity of Quantum Circuits

Expressivity measures how large a family of states or functions a parameterized circuit can reach, and it trades off directly against trainability.

What expressivity means

The expressivity of a parameterized circuit is the size and richness of the set of states it can prepare, or equivalently the set of functions it can represent, as its parameters vary. A highly expressive ansatz can approximate almost any unitary; a low-expressivity ansatz explores only a small corner of the state space. Expressivity is a capacity concept, analogous to the size of a hypothesis class in classical learning.

How it is quantified

One common measure compares the distribution of states the circuit produces under random parameters to the uniform (Haar) distribution over all states. If the circuit's ensemble matches the Haar distribution up to the second moment, it is a 2-design and maximally expressive in that sense. The distance from Haar, often computed via frame potentials, gives a scalar expressivity score.

The central trade-off

• More expressive circuits reach more solutions but flatten the loss into a barren plateau with exponentially small gradients.
• Less expressive circuits train easily but may not contain any good solution for the task.
• The goal is the least expressive ansatz whose reachable set still includes a solution.

This tension is not incidental; it is a theorem in spirit. The same property that lets a circuit imitate random unitaries, uniform coverage of state space, is what causes gradient concentration. You cannot have unconstrained expressivity and reliable trainability at once on large registers.

Inductive bias over raw capacity

Because raw expressivity is a liability, good design injects structure. Problem-inspired ansatze restrict the reachable set to physically relevant states; symmetry-equivariant circuits reach only states consistent with a known invariance. These reduce expressivity in a targeted way, keeping the useful part and shedding the part that causes plateaus. This is the quantum analogue of choosing an architecture with the right inductive bias.

# Sketch of an expressivity estimate via state overlaps
import numpy as np
ovs = [abs(inner(state(random_theta()), state(random_theta())))**2
       for _ in range(5000)]
# Compare the histogram of `ovs` to the Haar prediction (N-1)*(1-F)^(N-2)
# Closer match -> more expressive (and likely harder to train).

Reading expressivity claims

A paper reporting high expressivity is not reporting good news by itself. The right pairing is expressivity with a trainability analysis and a task-relevant solution guarantee. Expressivity that a task cannot use, or that cannot be trained, is capacity wasted at best and a barren plateau at worst.