Ansatz Design for VQE

Overview

In VQE, the ansatz is the parameterized circuit that prepares the candidate state $|\psi(\theta)\rangle$.

Ansatz choice is crucial because it sets the tradeoffs:

• which states are reachable (expressivity)
• how easy it is to optimize (trainability)
• how deep the circuit is (noise sensitivity)

A “better” ansatz is not always deeper. In NISQ, deeper often means noisier, and noisier can mean less useful.

Intuition

What an ansatz really is

An ansatz is a structured guess. It is a family of states controlled by parameters.

You are not searching all possible quantum states. You are searching a subset that your circuit can produce.

Hardware-efficient vs problem-inspired

Two broad families:

1. Hardware-efficient ansatz

   • built from native gates and repeating layers
   • aims to be easy to run on the device
   • risk: may be hard to train or may not match the problem structure well

2. Problem-inspired ansatz

   • built using knowledge of the Hamiltonian’s structure
   • aims to reach relevant low-energy states with fewer parameters
   • risk: may require gates that are harder to implement or deeper circuits

Expressivity vs trainability

• If the ansatz is too simple, it might never reach the true ground state.
• If the ansatz is too expressive, the optimization landscape can become flat or chaotic.

In NISQ, you often want a “just expressive enough” ansatz.

Why deeper circuits are not always better

Depth increases:

• accumulated gate error
• decoherence exposure
• measurement cost (more shots needed for stable estimates)

So there is a practical sweet spot:

• enough depth to represent useful states
• not so much depth that noise dominates

Formal Description

We describe ansatz design as choosing a parameterized circuit class.

Ansatz as a mapping from parameters to states

An ansatz defines:

• parameters $\theta$ (a vector of angles and possibly other control values)
• a circuit $U(\theta)$
• the prepared state $|\psi(\theta)\rangle = U(\theta)|0\cdots 0\rangle$

VQE then minimizes the energy over $\theta$.

Practical ansatz design criteria

A good ansatz tends to have:

• compatibility with hardware: uses gates you can execute reliably
• reasonable parameter count: not too few (underfit) and not too many (hard to train)
• structured entanglement: enough entangling capability to represent correlated states
• manageable depth: to survive noise

Interpreting “expressivity” carefully

High expressivity means the ansatz can represent a wide variety of states. But in VQE, expressivity is only valuable if you can actually find the right parameters.

So a practical criterion is:

• “expressive enough, trainable enough”

Worked Example

Consider a 2-qubit Hamiltonian that couples qubits (so product states may be insufficient).

Option A: shallow ansatz

• single-qubit rotations on each qubit
• no entangling gate

This ansatz may fail because it cannot represent entangled ground states.

Option B: slightly deeper ansatz

• single-qubit rotations
• one entangling gate
• another layer of rotations

Now the ansatz can represent some entangled states while still being shallow.

This example shows the design goal:

• add the minimum structure needed to represent the physics, without adding unnecessary depth

Turtle Tip

Common Pitfalls

Quick Check

What’s Next

Now that we have an ansatz, we must evaluate its energy. Next we focus on the biggest practical cost in VQE: Hamiltonian measurement and the shot cost of estimating expectation values.