Entanglement purification

Bell pairs delivered through a noisy quantum channel are imperfect — their fidelity to the ideal Bell state is below 1. Heralded Entanglement Purification (HEP) — also called distillation — is the family of protocols that take two (or more) low-fidelity Bell pairs and produce one pair of higher fidelity, at the cost of throughput. "Heralded" because the protocol announces success or failure on each round, and the network only consumes pairs that have been announced clean. The canonical protocol is BBPSSW, from Bennett, Brassard, Popescu, Schumacher, Smolin and Wootters (1996) ^{Bennett et al. 1996}.

Purification is the network's response to fidelity loss. Each entanglement- generation event has imperfections; each swap at a relay multiplies them. By sacrificing some pairs to clean others up, the network can keep delivered fidelity above whatever threshold downstream applications need (teleportation, QKD key extraction, distributed quantum computing).

Why purify

A perfect Bell pair has fidelity $F = 1$ . Real channels deliver mixed states with $F < 1$ — typical numbers in current memory-based repeater experiments are $F \sim 0.7 - 0.9$ on a single link. Two problems compound:

Multi-hop multiplication. Chained entanglement swapping multiplies per-link fidelities (see the swapping subject). Two hops at $F = 0.85$ each give an end-to-end pair at roughly $F \sim 0.72$ after the swap; ten hops at $F = 0.95$ each fall to $\sim 0.6$ . Without countermeasures, fidelity decays exponentially in the number of hops.
Application thresholds. Many downstream protocols simply do not work below a fidelity threshold. For teleportation of an arbitrary qubit through a Bell pair, the average output fidelity is $(2 F + 1) /3$ , where $F$ is the pair's singlet fraction (its overlap with the closest Bell state) ^{Horodecki et al. 1999}. Below $F = \frac{1}{2}$ the output drops to ⅔ — what a classical measure-and-prepare strategy alone can already achieve ^{Massar et al. 1995} — and the entanglement provides no quantum advantage. Practical applications need the Bell pair well above this threshold.

Purification flips the trade. Spend more pairs to get fewer pairs, but each one is cleaner. Since long-distance distribution is throughput-limited anyway (Bell-pair generation rates are slow), trading a factor of throughput for a factor of fidelity is often the right deal.

The BBPSSW protocol

The simplest purification primitive consumes two Bell pairs and probabilistically produces one cleaner pair. Alice and Bob each hold one half of each pair — $(a_{1}, b_{1})$ are the source pair (the one we hope to keep) and $(a_{2}, b_{2})$ are the target pair (consumed in the measurement). The protocol is local at each side, with one round of classical communication to reconcile outcomes.

Local CNOTs. Alice applies a CNOT with $a_{1}$ as control and $a_{2}$ as target. Bob applies the same operation locally: $b_{1}$ control, $b_{2}$ target.
Measure the target pair. Alice measures $a_{2}$ in the computational basis (the standard $∣0 ⟩$ / $∣1 ⟩$ basis); Bob measures $b_{2}$ . Each gets a 1-bit outcome.
Compare classically. Alice and Bob exchange their target-pair outcomes over a classical channel.
Decide. If the outcomes match (both 0 or both 1), the source pair $(a_{1}, b_{1})$ is kept — its fidelity has been increased. If they disagree, the source pair is discarded; the round is a failure and produces nothing.

1. Two noisy Bell pairs at $F = 0.70$
2. Alice and Bob each apply a local CNOT (source → target)
3. Measure the target pair locally
4. Compare outcomes over the classical channel
5a. Match → keep source pair, heralded at $F^{'} \approx 0.74$
5b. Mismatch → discard, try again

Each cycle alternates between a success (outcomes match — source pair is heralded as kept, its wavy bond thickens and its fidelity ticks from

F = 0.70

F^{'} \approx 0.74

per BBPSSW) and a failure (outcomes disagree — source pair fades and the round produces nothing). The target pair is consumed either way. Real protocols can reach much higher fidelity by recursing — feeding purified pairs back in as input to another round.

Throughput vs fidelity

BBPSSW spends two pairs to (probabilistically) produce one cleaner pair. With $ε \equiv (1 - F) /3$ , the success probability is $P_{succ} = F^{2} + 2 F ε + 5 ε^{2}$ , and on success the output fidelity is $F^{'} = (F^{2} + ε^{2}) / (F^{2} + 2 F ε + 5 ε^{2})$ . Plugging the figure's starting point: $F = 0.70$ gives $F^{'} \approx 0.74$ with $P_{succ} \approx 0.68$ . Two takeaways:

Each round at least halves the throughput. Two pairs in, zero or one out. Successive rounds compound this: purifying once costs ½ throughput, twice costs ¼, etc.
Each round increases fidelity provided the input was already above $F = 1/2$ . Below that threshold, BBPSSW's output is worse than the input — there's no useful entanglement to amplify, so purification can't bootstrap from purely classical noise. This is why generation must already deliver above-threshold pairs before purification can help.

Network role

Purification lives at the link or end-to-end layer. A network's entanglement-generation rate sets the supply of low-fidelity Bell pairs; purification is the post-processing that turns that supply into a smaller supply of high-fidelity pairs the application layer can actually use. RFC 9340 treats purification as part of the layered service contract — the link layer delivers Bell pairs, optionally with purification rounds applied, at a specified fidelity ^{RFC 9340}.

Two related subjects pick up the threads:

Repeaters integrate generation, swapping, and purification into the 1G / 2G / 3G architectural families. The role of purification differs across generations — central in 1G and 2G, replaced by quantum error correction in 3G. Covered in the repeaters subject.
Metrics grounds the fidelity numbers we've been throwing around — what counts as "low" or "high" fidelity, what application thresholds actually require, and how fidelity composes with loss along a real channel. Covered in the metrics subject.