VQE Ansatz Design

Choosing the parameterized circuit for a variational eigensolver, balancing expressiveness, trainability, and hardware constraints.

What an ansatz must do

The ansatz is the family of trial states |psi(theta)> the optimizer searches over. A good ansatz must contain a state close to the true ground state (expressiveness), be trainable without vanishing gradients, and be shallow enough to run under hardware noise. These goals conflict, so ansatz design is a balancing act.

Chemistry-inspired: UCC

The unitary coupled cluster ansatz builds |psi(theta)> = e^(T(theta) - T(theta)^dagger)|HF>, where |HF> is the Hartree-Fock reference and T contains parameterized excitation operators. UCCSD (singles and doubles) is physically motivated and systematically improvable, but its circuits are deep after fermionic encoding.

Hardware-efficient ansatze

These use whatever native gates the device offers, layers of single-qubit rotations and entangling gates, without reference to the problem. They are shallow and easy to run, but they lack physical structure, are prone to barren plateaus, and may need many parameters to reach the ground state.

• UCCSD: physically grounded, systematically improvable, deep circuits.
• Hardware-efficient: shallow, native gates, weak physical guarantees.
• ADAPT-VQE: grows the ansatz operator by operator, keeping it compact.
• Hamiltonian variational ansatz: layers alternate the Hamiltonian's own terms.

ADAPT-VQE

Rather than fixing the circuit up front, ADAPT-VQE builds it greedily: at each step it adds the operator from a pool whose gradient most lowers the energy. This yields compact, problem-tailored circuits that often reach chemical accuracy with fewer parameters, at the cost of extra measurements to rank operators.

Symmetry and initialization

Enforcing symmetries (particle number, spin) restricts the search to the physical subspace, improving convergence and rejecting unphysical states. Starting near the Hartree-Fock state, rather than random parameters, avoids the worst barren-plateau regions. The right ansatz depends jointly on the molecule, the encoding, and the hardware, so design is problem-specific rather than one-size-fits-all.