Joint Measurements: When Noisy X and Z Become Compatible

Joint Measurements: When Noisy X and Z Become Compatible

Sharp X and Z cannot share an exact joint measurement, but noisy versions can. Build a four-outcome POVM in NumPy and see why the threshold is 1/sqrt(2).

You cannot build one perfect qubit measurement that reproduces both a sharp X measurement and a sharp Z measurement as its marginals. But deliberately soften both measurements, and a single device can reproduce their noisy statistics exactly.

That is not a loophole in quantum mechanics. It is a useful distinction between measuring two sharp observables perfectly and jointly measuring two less informative observables. For equally noisy X and Z measurements, the dividing line is especially clean: a sharpness parameter of 1/sqrt(2).

What counts as a joint measurement?

Let each run of a device return a pair of classical signs, a and b, each either +1 or -1. A four-outcome positive operator-valued measure, or POVM, assigns an effect G[a,b] to each possible pair. Every effect must be positive semidefinite, and all four must add to the identity I.

For any input density matrix rho, the pair probability is

p(a,b) = Tr(rho G[a,b])

To recover an X-like measurement, ignore b and keep a. To recover a Z-like measurement, ignore a and keep b. The corresponding effects are the marginals:

sum_b G[a,b] = E_X[a]
sum_a G[a,b] = E_Z[b]

These must be operator identities, so the device works for every input state, not just a test state with convenient probabilities. This definition is about outcome statistics. It does not specify the postmeasurement state or promise that either measurement is nondisturbing.

For sharp projective observables, joint measurability is equivalent to commutativity. For general POVMs, it is not. Heinosaari, Reitzner, and Stano's notes explain that distinction and give the qubit compatibility criterion used below.

Make the desired marginals slightly blurry

Define the two binary measurements by

E_X[a] = (I + a eta X)/2
E_Z[b] = (I + b eta Z)/2
0 <= eta <= 1

Here X and Z are Pauli matrices. At eta = 1, these are the usual sharp measurements. At eta = 0, each output is a fair coin independent of the input. Intermediate values interpolate between them.

One operational description is to make the ideal measurement with probability eta and otherwise output a fair random sign. Equivalently, take its ideal sign and flip it with probability (1-eta)/2. The parameter eta describes measurement sharpness, not the purity of the input qubit.

A natural candidate joint measurement is

G[a,b] = (I + eta(a X + b Z))/4

Summing over b cancels the Z term and gives E_X[a]. Summing over a gives E_Z[b]. Adding everything gives I. The remaining question is positivity.

Because X and Z anticommute and each squares to I,

(a X + b Z)^2 = 2 I

Thus every G[a,b] has eigenvalues

(1 + eta sqrt(2))/4
(1 - eta sqrt(2))/4

All four effects are positive exactly when eta <= 1/sqrt(2). Below that threshold, the candidate really is a valid joint POVM.

A failed candidate alone would not prove that no better construction exists. Here we also have a necessity theorem: two unbiased binary qubit observables with Bloch vectors u and v are compatible exactly when

|u + v| + |u - v| <= 2

For u = eta x_hat and v = eta z_hat, this becomes 2 sqrt(2) eta <= 2. So the threshold is optimal for these equally sharp, orthogonal, unbiased observables. It is not a universal noise threshold for arbitrary measurements. The same joint-POVM construction appears as the two-generator case in Kunjwal, Heunen, and Fritz's paper.

A physical recipe at the threshold

At eta = 1/sqrt(2), flip a fair classical coin to choose between two sharp measurement axes:

N_plus  = (X + Z)/sqrt(2)
N_minus = (X - Z)/sqrt(2)

If you chose N_plus and obtained sign t, report (a,b) = (t,t). If you chose N_minus, report (a,b) = (t,-t).

For example, the (+1,+1) event comes from choosing N_plus and getting +1. Its effect is one half times that measurement's positive projector, exactly G[+1,+1]. The other three events work the same way. Suitable single-qubit basis rotations followed by ordinary Z readout can implement either axis measurement.

For a smaller nonnegative eta, run this threshold device with probability sqrt(2) eta. Otherwise output an independent, uniformly random sign pair. This adds enough classical noise to obtain the desired G effects without changing their form.

Verify the matrices, not just one distribution

This NumPy example checks normalization and both operator-valued marginals at three sharpness values. It prints the smallest effect eigenvalue, including one deliberately invalid candidate above the compatibility threshold. Only the valid eta = 0.6 device is used to calculate physical joint probabilities.

import numpy as np

I = np.eye(2)
X = np.array([[0.0, 1.0], [1.0, 0.0]])
Z = np.diag([1.0, -1.0])
signs = [1, -1]


def joint_effects(eta):
    return {(a, b): (I + eta * (a * X + b * Z)) / 4
            for a in signs for b in signs}


for eta in [0.5, 1 / np.sqrt(2), 0.8]:
    G = joint_effects(eta)
    assert np.allclose(sum(G.values()), I)
    for a in signs:
        assert np.allclose(sum(G[a, b] for b in signs),
                           (I + a * eta * X) / 2)
    for b in signs:
        assert np.allclose(sum(G[a, b] for a in signs),
                           (I + b * eta * Z) / 2)
    minimum = min(np.linalg.eigvalsh(g).min() for g in G.values())
    if abs(minimum) < 1e-12:
        minimum = 0.0
    print(f"eta={eta:.6f}: smallest eigenvalue={minimum:.6f}")

eta = 0.6
G = joint_effects(eta)
rho = (I + 0.3 * X + 0.4 * Z) / 2
p = {key: np.trace(rho @ g).real for key, g in G.items()}
assert min(p.values()) >= 0 and np.isclose(sum(p.values()), 1)
order = [(1, 1), (1, -1), (-1, 1), (-1, -1)]
print("joint probabilities:", " ".join(f"{p[k]:.6f}" for k in order))
px = sum(value for (a, b), value in p.items() if a == 1)
pz = sum(value for (a, b), value in p.items() if b == 1)
print(f"P(a=+1)={px:.6f}")
print(f"P(b=+1)={pz:.6f}")

Output:

eta=0.500000: smallest eigenvalue=0.073223
eta=0.707107: smallest eigenvalue=0.000000
eta=0.800000: smallest eigenvalue=-0.032843
joint probabilities: 0.355000 0.235000 0.265000 0.145000
P(a=+1)=0.590000
P(b=+1)=0.620000

Notice that the eta = 0.8 matrices still add to I and have the algebraically correct marginals. That is not enough. Their negative eigenvalues make them invalid effects: some input state would receive a negative supposed probability. Positivity is an essential physical constraint, not a numerical nicety.

What the two reported signs do not mean

The output pair is not a disclosure of two pre-existing sharp X and Z values. Its marginals describe the noisy POVMs we asked for, not the ideal projective measurements. Nor have we cloned the qubit or measured sharp X followed by sharp Z without disturbance.

You can estimate the ideal expectation values from repeated runs by dividing the mean reported signs by eta, provided eta is known and nonzero. But this rescales statistical fluctuations too. It does not recover two perfect sharp outcomes on an individual qubit.

The useful lesson is more precise than saying incompatible measurements can never share a device. Sharp X and Z cannot share one exact joint measurement. Their sufficiently noisy versions can, because the information surrendered by each marginal is exactly what makes a common positive measurement possible.