Ansatz state for finding the lowest eigenvalue of a $2^n\times 2^n$ real matrix using VQE

I would like to find the lowest eigenvalue of a $$2^n\times 2^n$$ real matrix $$H$$ using the VQE procedure. The measurement part is simple — I just expand $$H$$ in a sum of all possible $$n$$-qubit Pauli terms.

But I have not yet come up with a nice way of preparing an arbitrary $$n$$-qubit state with real amplitudes. What would be a good ansatz operator $$U(\theta_1,\ldots\theta_{2^n-1})$$ to do the job?

(If I were solving the problem in the second-quantized formalism using $$2^n$$ qubits, I would simply use the Unitary Coupled Cluster with $$(2^n - 1)$$ single amplitudes — does it have an analogue in such a 'binary' encoding??)

For preparing an arbitrary $$n$$-qubit state with real amplitudes, you can simply:

1. Take a circuit for preparing arbitrary states,
2. Replace all single-qubits rotations with $$R_y$$ rotations.

Here is how I do it in Qiskit:

def arbitrary_real_superposition_qiskit( amps: List[ float ], qc_arb = None ):
_amps = np.array( amps )
if not misc.is_power_of_two( _amps.shape[ 0 ] + 1 ):
raise Exception( 'The number of parameters for arbitrary_real_superposition_qiskit() has to be 2^n-1.' )

nqubit = int( np.log2( _amps.shape[ 0 ] + 1 ) )

qc = qiskit.circuit.quantumcircuit.QuantumCircuit( nqubit )

if type( qc_arb ) != qiskit.circuit.quantumcircuit.QuantumCircuit:
qc_arb = init_state_qiskit( [ random.random() for i in range( _amps.shape[ 0 ] ) ] )

i = 0
for inst in qc_arb.data:
if type( inst[ 0 ] ) == qiskit.circuit.library.standard_gates.u3.U3Gate \
or type( inst[ 0 ] ) == qiskit.circuit.library.standard_gates.u2.U2Gate \
or type( inst[ 0 ] ) == qiskit.circuit.library.standard_gates.u1.U1Gate:
qc.ry( _amps[ i ], inst[ 1 ] )
i = i + 1
elif type( inst[ 0 ] ) == qiskit.circuit.library.standard_gates.x.CXGate:
qc.cx( inst[ 1 ][ 0 ], inst[ 1 ][ 1 ] )
else:
raise Exception( 'A gate other than u1/u2/u3 and cx was encountered.' )
return qc


where

def is_power_of_two( N ):
"""Returns True if n is a power of 2."""
n = int(N)
if n <= 0:
return False
else:
return n & (n - 1) == 0 # How the hell is this even working??? :D


and

def init_state_qiskit( amps: List[ complex ] ):

qubitn = int( np.log2( amps.shape[ 0 ] ) )

qc = qiskit.circuit.quantumcircuit.QuantumCircuit( qubitn )
qc.isometry( amps, list( range( qubitn ) ), None )

while True:
qc = qc.decompose()
if qc == qc.decompose():
break

qc = qiskittranspile( qc, optimization_level = 3, basis_gates = [ 'u1', 'u2', 'u3', 'cx' ] )

return qc


What happens here is that I initialize an $$n$$-qubit state from an arbitrary $$2^n$$ vector (to make sure that all the $$2^n-1$$ single-qubit rotations are present in the circuit), and replace those single-qubit rotations with $$R_y$$'s.

If you don't have any prior on your matrix, you can use the Strongly Entangling Circuit as a very generic ansatz for qubits.

Giving a general $$2^n \times 2^n$$ Hermitian matrix $$H$$, it's not guarantee that you can decompose/expand $$H$$ in polynomial strings of $$n$$-qubit Pauli matrices. So I don't know if VQE is a good method to find lowest eigenvalues of a matrix... unless you know that you can sum $$H$$ in polynomial terms of Pauli matrices. Even then, in general, VQE only gonna give you a good answer if you have some intuition about the problem. If you pick a variational form that is hardware efficient, using polynomial number of gates, then you can't expect it explore the entire Hilbert space...