Does Entanglement Matching Survive Real Noise? A Hardware Test of Probabilistic Teleportation

Teleportation moves a quantum state from Alice to Bob using a shared entangled pair and a measurement on Alice's side. When the shared pair is only partially entangled, perfect transfer is still possible, but only some of the time. A theoretical rule called entanglement matching, from Probabilistic teleportation and entanglement matching, says Alice should tune how entangled her measurement is to match the entanglement of the shared pair. Any extra entanglement in the measurement is wasted. Any shortfall leaves the channel underused. The rule was derived for perfectly pure states, gates without error, and a channel whose entanglement Alice knows exactly. None of those hold on real quantum processors, and no paper in our survey has run the protocol on hardware or a realistic noise simulator.

This note proposes the missing test, along with the theory needed to make it a sharp test rather than a demonstration. Read carefully, the pure-state rule makes an uncomfortable prediction. Success probability is $P = 2\min(|b|^2,|b'|^2)$, where $|b|^2 \le 1/2$ measures the channel's entanglement and $|b'|^2 \le 1/2$ measures the measurement's. A standard, maximally entangled Bell measurement ($|b'|^2=1/2$) therefore already achieves the matched value $2|b|^2$. In the ideal model, matching buys no success probability. It only saves measurement entanglement. Whether matching earns anything operationally can only be settled where the ideal model breaks: under mixed noise, gate errors, and uncertainty about $|b|^2$. That is the regime we study.

Core Claim

Our hypothesis is that noise and channel uncertainty turn the measurement parameter $|b'|^2$ into a real control knob. With decohered channels and imperfect gates, Alice's choice of $|b'|^2$ and Bob's choice of recovery operation trade off two quantities: how often the protocol reports success, and how faithful the state is when it does. The best setting need not coincide with the matched point $|b'|^2=|b|^2$. We will compute that trade-off curve, which we call the success-fidelity frontier, for noise models calibrated to hardware. We will then measure it on small processors and report where theory and experiment agree or disagree.

The new elements are threefold. We extend matching from pure, known channels to mixed channels and to channels known only up to an interval. We make falsifiable predictions before running anything. And we execute the protocol at the 3–5 qubit scale, where full state reconstruction is still affordable. A null result is informative: if the frontier for a plain Bell measurement dominates the matched one on hardware, the matching principle is a statement about resource accounting rather than performance, and that is worth knowing.

Research Program

Mixed-state extension. We model the shared pair as the pure state $a|00\rangle+b|11\rangle$ passed through amplitude damping, dephasing, and depolarizing noise, with rates taken from device calibration data. For each noisy pair, we numerically optimize Alice's two-qubit measurement within the family parameterized by $(a',b')$. We also optimize Bob's recovery step, which couples his qubit to a helper qubit through a joint unitary and keeps the run only if the helper reads the right outcome. The output is the frontier of success probability $P$ against conditional fidelity $F$. We will report the optimal $|b'|^2$ as a function of noise strength and compare it to the matched value. We register two predictions in advance. First, without noise the matched and Bell frontiers coincide in $P$. Second, the gap between them under noise has a sign and size that we will compute and commit to before any hardware run.

Uncertain channels. We borrow the setting of compound channels: the channel is fixed but unknown, and is only guaranteed to lie in a known set. The Bjelaković–Boche–Nötzel line of work (Entanglement transmission capacity of compound channels, Universal Quantum Channel Coding) shows that for asymptotic capacity, the rate is set by the worst channel in the set. It also shows that giving the receiver knowledge of the channel does not raise capacity, while giving the sender that knowledge can.

We use these results as a design lens rather than a method, since they concern long codes and our protocol uses a single pair. Here $|b|^2$ is known only to lie in an interval. We compute two worst-case policies over that interval. In the first, Alice is uninformed and picks one $|b'|^2$ and one recovery for the whole interval. In the second, only Bob knows the true $|b|^2$. Comparing the two at a single-pair blocklength asks whether the asymptotic result that receiver knowledge does not help survives at the smallest possible scale. Because entanglement matching is a sender-side adaptation, sender knowledge is the natural place for any gain to appear, and we will check that too.

Hardware execution. We compile each protocol variant to OpenQASM. The circuit uses the qubit to be sent, the two halves of the shared pair, and Bob's helper qubit, so four qubits in all, plus one more when we build the purification needed to measure average fidelity. This places every circuit in the small-scale tier of QASMBench. That paper ran its suite on IBM-Q, IonQ, and Rigetti processors. It found that state reconstruction by measurement is affordable only at this 2–5 qubit scale, and that small refined devices can outperform the 127-qubit machine. Both findings tell us where to run.

We also borrow QASMBench's practice of reporting circuit depth, measurement density, and gate counts next to fidelity. That lets us attribute any loss to specific overhead, such as Bob's recovery unitary or the mid-circuit measurement of the helper qubit. The same circuits are run in a density-matrix simulator with the calibrated noise model. The gap between simulator and device is a reported result in its own right.

Baselines

The do-nothing reference is standard teleportation with a Bell measurement over the partially entangled pair and no recovery step on Bob's side. It always reports success, and its fidelity is below one. This anchors the frontier at $P=1$.

The strongest competitor is Bell-measurement probabilistic teleportation ($|b'|^2=1/2$) with Bob's recovery step tuned to the channel. In the pure-state theory of Probabilistic teleportation and entanglement matching it achieves exactly the matched success probability, and it uses the measurement basis that hardware supports natively. We found no reported hardware or noisy-simulation result on Lacuna for probabilistic teleportation with entanglement matching or with this Bell variant, so there is no published number to beat. The pure-state value $P=2|b|^2$ at $F=1$ serves as the reference.

The second strong competitor is to filter first, then teleport. One party applies a local filtering operation that converts the pair into a maximally entangled pair with probability $2|b|^2$, followed by ordinary teleportation. This reaches the same ideal success probability through a different circuit.

The ablation that isolates our key ingredient is the matched protocol with $|b'|^2$ set to the pure-state matched value, compared against our noise-optimized $|b'|^2$ using the same recovery circuit family.

Fairness. All variants use the same qubits on the same device, interleaved within one calibration window, with the same number of shots per point (we will fix 8,192). Each variant gets the same compilation optimization level. Every optimized parameter, whether ours or a baseline's, is tuned with the same budget against the same calibrated noise model before any device run.

What counts as beating a baseline. A method beats a baseline if its measured frontier lies strictly above the baseline's, by more than the bootstrap 95% confidence band, over a contiguous range of target fidelities on at least two device families. Equal frontiers count as a confirmed null result, not a loss.

Why not just…?

Why not just use the Bell measurement and forget matching? This is the real threat, and the pure-state formula already backs it, since $\min(|b|^2,1/2)=|b|^2$. Our experiment is designed to settle it. If the Bell frontier matches or dominates the matched and noise-optimized frontiers on hardware, we will conclude that matching has no operational value beyond saving measurement entanglement. The one place matching could still win is fidelity under noise. A less entangled measurement may be compiled with fewer or weaker two-qubit interactions, and we will check that directly in the depth and gate-count data.

Why not just distill the pair to full entanglement first? For a single pair, distillation is impossible: it consumes several noisy pairs to make one better pair. Scaling Limits of Quantum Repeater Networks found that at gate fidelity 0.99, extra distillation rounds eventually lower end-to-end fidelity, and that hardware errors cut the reachable network length to 115 km, against roughly 267 km with perfect hardware. The single-pair analogue of distillation is the filter-first baseline above, so this alternative is tested head to head rather than argued away.

Why not just estimate the channel, then match exactly? Estimation costs pairs and time, and the cost of state reconstruction grows quickly with system size, as QASMBench notes. Our uncertain-channel arm measures exactly what knowing $|b|^2$ is worth. The gap between the worst-case policy and the policy with full knowledge is an upper bound on what any estimation scheme could gain. If that gap is small next to hardware error bars, estimation is not worth doing.

Why not just simulate? Simulation alone reproduces the gap between theory and practice that motivates this note. The value lies in comparing calibrated simulation against the device. If they agree, future tests of matching can be run cheaply in simulation.

Proposed Experiments and Decisive Checks

We sweep $|b|^2$ over ten values in $(0,1/2]$ and, for each, $|b'|^2$ over a grid. On IBM-Q, IonQ, and Rigetti devices we measure success probability from the helper qubit's outcome and conditional fidelity by state reconstruction, averaged over six input states that probe the whole Bloch sphere.

The first decisive check is the noiseless limit in simulation. All three probabilistic variants must give $P=2|b|^2$ at $F=1$. This validates the implementation.

The second check compares the pre-registered noise-optimal $|b'|^2$ with the value measured on hardware. Agreement within error bars supports extending matching to mixed states. Systematic disagreement localizes the failure to specific overheads.

The third check is the uncertain-channel arm. We compare the worst-case frontier over an interval of $|b|^2$ against the frontier with only Bob informed and with only Alice informed. This tests whether the asymptotic asymmetry between sender and receiver knowledge appears at a blocklength of one.

The single figure we commit to publishing either way is a plot of the frontier, matched against Bell, on each device. That figure decides whether entanglement matching is a design rule or only an accounting identity.