bp_validation.py

"""
Compute the circuit gradient with single qubit random Haar averaged to
demonstrate the gradient vanishing, aka, barren plateau.
"""
import numpy as np
from tqdm import tqdm
import tensorcircuit as tc

xx = tc.gates._xx_matrix
yy = tc.gates._yy_matrix
zz = tc.gates._zz_matrix

nqubits, nlayers = 8, 6
# number of qubits and number of even-odd brick layers in the circuit
K = tc.set_backend("tensorflow")


def ansatz(param, gate_param):
    # gate_param here is a 2D-vector for theta and phi in the XXZ gate
    c = tc.Circuit(nqubits)
    # customize the circuit structure as you like
    for j in range(nlayers):
        for i in range(nqubits):
            c.ry(i, theta=param[j, 0, i, 0])
            c.rz(i, theta=param[j, 0, i, 1])
            c.ry(i, theta=param[j, 0, i, 2])
        for i in range(0, nqubits - 1, 2):  # even brick
            c.exp(
                i,
                i + 1,
                theta=1.0,
                unitary=gate_param[0] * (xx + yy) + gate_param[1] * zz,
            )
        for i in range(nqubits):
            c.ry(i, theta=param[j, 1, i, 0])
            c.rz(i, theta=param[j, 1, i, 1])
            c.ry(i, theta=param[j, 1, i, 2])
        for i in range(1, nqubits - 1, 2):  # odd brick with OBC
            c.exp(
                i,
                i + 1,
                theta=1.0,
                unitary=gate_param[0] * (xx + yy) + gate_param[1] * zz,
            )
    return c


def measure_z(param, gate_param, i):
    c = ansatz(param, gate_param)
    # K.real(c.expectation_ps(x=[i, i+1])) for measure on <X_iX_i+1>
    return K.real(c.expectation_ps(z=[i]))


gradf = K.jit(K.vvag(measure_z, argnums=0, vectorized_argnums=0), static_argnums=2)
# vectorized parallel the computation on circuit gradients for different circuit parameters

if __name__ == "__main__":
    batch = 100
    # tune batch as large as possible if the memory allows
    reps = 20
    measure_on = nqubits // 2
    # which qubit the sigma_z observable is on

    # we will average `reps*batch` different unitaries
    rlist = []
    gate_param = np.array([0, np.pi / 4])
    gate_param = tc.array_to_tensor(gate_param)
    for _ in tqdm(range(reps)):
        param = np.random.uniform(0, 2 * np.pi, size=[batch, nlayers, 2, nqubits, 3])
        param = tc.array_to_tensor(param, dtype="float32")
        _, gs = gradf(param, gate_param, measure_on)
        # gs.shape = [batch, nlayers, 2, nqubits, 3]
        gs = K.abs(gs) ** 2
        rlist.append(gs)
    gs2 = K.stack(rlist)
    # gs2.shape = [reps, batch, nlayers, 2, nqubits, 3]
    gs2 = K.reshape(gs2, [-1, nlayers, 2, nqubits, 3])
    gs2_mean = K.numpy(
        K.mean(gs2, axis=0)
    )  # numpy array for the averaged abs square for gradient of each element
    print(gs2_mean)