Channel Discrimination: A Bell Pair Can Reveal What One Qubit Cannot
Identify an identity channel or a random-Pauli channel with one use. A single probe tops out at 5/6 success, while a retained Bell reference enables perfect ideal discrimination.
Suppose someone hands you a sealed quantum device. It implements one of two known processes, chosen with equal probability, and you may use it exactly once. Your job is to identify which process it contains.
One process does nothing. The other randomly applies X, Y, or Z, each with probability one third. Surely a carefully chosen test qubit should settle the question?
Not perfectly. Even the best single-qubit probe leaves some ambiguity. But if you keep a reference qubit entangled with the probe, an ideal joint measurement identifies the device without error. The extra qubit never enters the device. What matters is the correlation it preserves.
Define the two devices before comparing strategies
Call the identity channel A and the random-Pauli channel B. On a one-qubit density matrix rho, they act as
A(rho) = rho
and
B(rho) = (X rho X + Y rho Y + Z rho Z) / 3.
B is a classical mixture of three unitary operations, not a coherent superposition of X, Y, and Z. Its Kraus operators are X/sqrt(3), Y/sqrt(3), and Z/sqrt(3), so it is a valid trace-preserving quantum channel.
The rules are important: one invocation, equal prior probabilities, no access to the device's internal random choice, and a final guess on every trial. First we forbid an entangled reference; then we allow one. State preparation, storage, and readout are ideal in this calculation.
This is an instance of the Pauli-channel discrimination problem studied by Massimiliano Sacchi. In particular, his Eq. 38 describes a Pauli unitary versus a mixture of the other three, for which entangled probes can enable perfect discrimination.
Why one qubit cannot settle it
Start with |0>. Device A returns |0>. Device B returns |1> after X or Y, but returns |0> after Z. Measuring in the computational basis gives a sensible rule: guess B after outcome 1, and A after outcome 0.
This succeeds with probability
(1/2) * 1 + (1/2) * (2/3) = 5/6.
That is not merely the performance of a convenient basis choice. It is optimal without entanglement.
For any normalized qubit state, the Pauli identity gives
B(rho) = (2I - rho) / 3.
For a pure input rho, its eigenvalues in the basis containing that input are 1 and 0. B's output therefore has eigenvalues 1/3 along the input and 2/3 along its orthogonal partner. The outputs always overlap, whatever direction we choose on the Bloch sphere.
More formally, write rho = (I + r dot sigma)/2. The Hermitian difference A(rho) - B(rho) has eigenvalues +2|r|/3 and -2|r|/3. Its trace norm, the sum of its absolute eigenvalues, is 4|r|/3.
For two equally likely states sigma and tau, the optimal guessing probability is the Holevo-Helstrom expression
P_success = 1/2 + ||sigma - tau||_1 / 4.
Since |r| <= 1, no unassisted input beats 5/6. Pure states reach it; mixing the input cannot improve it. Keeping an ordinary classical record of which probe you prepared cannot help either, since each recorded choice obeys the same bound.
The trace norm here is not the Frobenius norm and not the sum of absolute matrix entries. That distinction matters when implementing the test.
Keep half a Bell pair outside the device
Prepare the probe S and reference R in
|Phi+> = (|00> + |11>) / sqrt(2).
Use tensor order S then R, and send only S through the unknown channel. A returns |Phi+> unchanged. Applying X, Y, or Z to S instead produces, up to an irrelevant global phase,
- X:
|Psi+> = (|01> + |10>) / sqrt(2). - Y:
|Psi-> = (|01> - |10>) / sqrt(2). - Z:
|Phi-> = (|00> - |11>) / sqrt(2).
These four Bell states are mutually orthogonal. B's output is an equal classical mixture of the three states other than |Phi+>. Its support is orthogonal to A's output.
We do not need to determine which Pauli B sampled. A two-outcome measurement with effects P = |Phi+><Phi+| and I - P is enough. The P outcome means A; the other outcome means B. Neither channel can produce the wrong outcome in this ideal setup.
A full Bell-basis measurement also works. One implementation applies CNOT from S to R, then H on S, then measures both qubits in the computational basis. The |Phi+> state decodes to 00, while the other Bell states decode to the other three strings.
Check both strategies in NumPy
The code below computes the optimal output-state guessing probability using Hermitian eigenvalues. It also verifies that B's two-qubit output has zero overlap with the |Phi+> projector.
import numpy as np
I = np.eye(2, dtype=complex)
X = np.array([[0, 1], [1, 0]], dtype=complex)
Y = np.array([[0, -1j], [1j, 0]], dtype=complex)
Z = np.diag([1, -1]).astype(complex)
paulis = [X, Y, Z]
def random_pauli(rho, reference=False):
operators = [np.kron(P, I) for P in paulis] if reference else paulis
return sum(K @ rho @ K.conj().T for K in operators) / 3
def success(sigma, tau):
trace_norm = np.abs(np.linalg.eigvalsh(sigma - tau)).sum()
return 0.5 + 0.25 * trace_norm
zero = np.array([1, 0], dtype=complex)
rho = np.outer(zero, zero.conj())
phi = np.array([1, 0, 0, 1], dtype=complex) / np.sqrt(2)
P = np.outer(phi, phi.conj())
bell_output = random_pauli(P, reference=True)
print(f"One-qubit success: {success(rho, random_pauli(rho)):.6f}")
print(f"Bell-assisted success: {success(P, bell_output):.6f}")
print(f"Phi+ probability after B: {np.trace(P @ bell_output).real:.6f}")
print("B output eigenvalues:", np.round(np.linalg.eigvalsh(bell_output), 6))
assert np.allclose(bell_output, (np.eye(4) - P) / 3)
# Discard R and compare the remaining S states.
def discard_reference(joint):
return np.trace(joint.reshape(2, 2, 2, 2), axis1=1, axis2=3)
print("Probe-only outputs equal:",
np.allclose(discard_reference(P), discard_reference(bell_output)))
The output is
One-qubit success: 0.833333
Bell-assisted success: 1.000000
Phi+ probability after B: 0.000000
B output eigenvalues: [0. 0.333333 0.333333 0.333333]
Probe-only outputs equal: True
This is an exact density-matrix calculation up to floating-point precision, not a noisy hardware experiment or an estimate from finite measurement shots.
The information is in the joint state
There is a useful twist in the last line. Each half of |Phi+> is maximally mixed. Both A and B leave that local input I/2 unchanged. If you throw away R and examine only the probe, this particular experiment gives no information at all about the channel.
The Bell strategy wins because the two-qubit outputs have different correlations, not because the probe alone becomes easier to read. The retained reference and joint readout are real resources, not free details hidden in the word "input."
Nor does this mean that entanglement makes every pair of channels perfectly distinguishable. Our result uses a specific pair with disjoint Pauli alternatives and a perfect Bell probe. For other devices, the output supports can overlap even with a reference. Sacchi's Pauli-channel analysis gives broader conditions rather than a universal promise.
When testing a quantum process, ask two separate questions: which probe should enter the device, and what reference system should remain available at readout? Optimizing only the first can miss information that exists solely in correlations.