Forward-Assisted Purifications

• Forward-assisted purifications are advanced noise-mitigation protocols that integrate pre-, post-, and memory-assisted processing to improve quantum channel fidelity.
• They utilize coherent control, interference of noisy channels, and auxiliary resources such as entangled qudits or carrier qubits to achieve superior error suppression.
• These protocols enhance repeater architectures and error-correction schemes by optimizing the purification process architecture beyond conventional post-processing.

Forward-assisted purifications are purification protocols in which the intervention is not confined to post-processing on already degraded outputs, but is instead distributed before, during, or after the noise process, or supplemented by auxiliary forward resources along the transmission path. In the most explicit formulation, purification is recast as a spatiotemporal task over superchannels that insert pre-processing, memory, and post-processing around noise, rather than as a static CPTP recovery map acting only on $N(\psi)^{\otimes n}$ (Meng et al., 2 Jun 2026). Related work uses the same or closely aligned idea at several levels: coherent postselected cooling beyond the fixed point of heat-bath algorithmic cooling (Goldberg et al., 2021), channel-centric purification by coherent interference of noisy channels in optics (Fei et al., 31 Oct 2025), optimistic purification on half-RGS primitives in all-photonic repeaters (Benchasattabuse et al., 25 Apr 2025), carrier-assisted entanglement purification by forward quantum communication (Kim et al., 9 Sep 2025), and auxiliary-entanglement-assisted syndrome extraction on finite ensembles (Sàbat et al., 2020).

1. Definition and formal setting

In the conventional setting, purification acts only after noise has taken effect. The basic object is the noisy output state $N(\psi)$, and one optimizes a CPTP map $E$ acting on $n$ copies, $N(\psi)^{\otimes n}$, to maximize fidelity with the ideal target $\psi$ (Meng et al., 2 Jun 2026). The spatiotemporal formulation replaces this channel-level viewpoint by a higher-order one: the noise is treated as a process, and purification is performed by a superchannel

$$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$

This makes pre-processing, memory, and post-processing native parts of the optimization problem (Meng et al., 2 Jun 2026).

The same framework isolates physically meaningful subclasses. The paper introducing the formalism studies PostP (post-processing only), PreP (pre-processing only), UA (unassisted, i.e. pre + post, but no memory), EA (entanglement-assisted), FCA (forward-classical-assisted memory), FHA (forward-Horodecki-assisted memory), PPT, NS, and PPT $\cap$ NS classes (Meng et al., 2 Jun 2026). Conventional purification is recovered as the special case with only post-processing. Efficient purification is defined by strict improvement over the no-purification benchmark,

$F(E,N,S) > F(N,S),$

where the average fidelity is the figure of merit (Meng et al., 2 Jun 2026).

This formulation makes “forward assistance” a structural notion rather than a single protocol family. It denotes assistance that is inserted along the temporal trajectory of the noisy process, or along the forward communication path, instead of being deferred to a terminal LOCC step.

2. Operational mechanisms of forward assistance

Across the cited literature, forward assistance is instantiated by several distinct resource patterns.

• Pre-/post-distributed interventions: the intervention is split around the noise, optionally with memory linking the stages, as in the superchannel formulation (Meng et al., 2 Jun 2026).
• Coherent control of operation order: a control qubit prepares a superposition of two unitary orders, and a single postselected measurement filters the purified branch (Goldberg et al., 2021).
• Coherent interference between noisy channels: two copies of a noise channel are embedded in a coherent Fredkin-gate circuit so that unwanted error components interfere destructively (Fei et al., 31 Oct 2025).
• Auxiliary entanglement as a measurement resource: maximally entangled qudit pairs implement entangling nonlocal measurements that determine the number and position of errors in a finite ensemble (Sàbat et al., 2020).
• Forward quantum communication by carriers: one stored noisy pair is purified using one or more traveling carrier qubits rather than multiple simultaneously stored noisy pairs (Kim et al., 9 Sep 2025).
• Purification along the repeater path: purification is performed directly on half-RGS primitives before they are assembled into full repeater graph states (Benchasattabuse et al., 25 Apr 2025).

This suggests that forward-assisted purifications are best understood as an operational family unified by where the extra resource is inserted: before noise, around noise, or along the transmission path. The common departure from standard purification is that the protocol acts on the process architecture, not only on the noisy outputs.

3. Postselected coherent-control purification of states

A canonical state-level example is the extension of heat-bath algorithmic cooling (HBAC) by postselection. Standard HBAC starts with $n+1$ qubits and repeatedly resets one qubit by replacing it with a thermalized qubit from a bath and then applies a compression/sorting unitary to move entropy toward the reset qubit. The reset qubit is thermal,

$N(\psi)$0

and the usual HBAC compression is summarized as

$N(\psi)$1

In the asymptotic limit HBAC converges to the “ultimate” distribution $N(\psi)$2, but that state is not fully pure: the most pure qubit is approximately $N(\psi)$3, and the least pure qubit is $N(\psi)$4 (Goldberg et al., 2021).

The forward-assisted ingredient is a single control qubit prepared in

$N(\psi)$5

together with a quantum switch that applies two unitaries $N(\psi)$6 and $N(\psi)$7 in a control-dependent order,

$N(\psi)$8

With the control prepared in $N(\psi)$9, measuring the control in the $E$0 basis produces interference terms that can cancel the mixed components left by standard HBAC (Goldberg et al., 2021).

In the main HBAC+PS protocol, one first runs ordinary HBAC until the system reaches $E$1, then replaces the standard compression step with a coherently controlled pair of unitaries, for example

$E$2

and finally measures only the control qubit in the $E$3 basis. For the successful branch $E$4, the effective transfer matrix becomes

$E$5

so that all intermediate mixed components are annihilated (Goldberg et al., 2021). After measuring one target qubit in the ground/excited basis, the remaining qubits collapse to a completely pure state, yielding $E$6 pure qubits with

$E$7

The purification is heralded. The success probability is

$E$8

for moderate-to-large $E$9, and is independent of $n$0 (Goldberg et al., 2021). Variants denoted HBAC+$n$1PS purify $n$2 qubits with

$n$3

The distinctive forward-assisted feature is that a single auxiliary pure control qubit, a coherent superposition of unitary orders, and one binary-outcome measurement certify the complete purification of many qubits.

4. Channel-centric forward-assisted purification

A channel-level realization appears in experimental quantum channel purification. Here the target is not a distributed noisy state but the effective transmission map itself. The theoretical protocol consumes two copies of a noise channel $n$4 and outputs a purified channel $n$5. The circuit uses three registers: a main register initialized in $n$6, an ancillary register initialized in the maximally mixed state $n$7, and a control register initialized in $n$8. The sequence is: apply a Fredkin gate, pass the registers through two noise channels, apply a second Fredkin gate, measure the control qubit in the Pauli-$n$9 basis, and keep the main register only when the measurement outcome is $N(\psi)^{\otimes n}$0 (Fei et al., 31 Oct 2025).

For a Pauli-twirled channel

$N(\psi)^{\otimes n}$1

the purified channel remains Pauli-diagonal, with coefficients

$N(\psi)^{\otimes n}$2

If $N(\psi)^{\otimes n}$3 is the largest coefficient, then $N(\psi)^{\otimes n}$4, so the purified channel is closer to the ideal channel whenever the intended unitary component is already dominant (Fei et al., 31 Oct 2025). The same work defines a “virtual” purified channel from discarded $N(\psi)^{\otimes n}$5 outcomes,

$N(\psi)^{\otimes n}$6

which is not a physical output channel but can be used for expectation-value evaluation and can suppress noise more strongly, with $N(\psi)^{\otimes n}$7.

The experiment implements the protocol entirely in linear optics. Four logical qubits are encoded across two photons and two degrees of freedom: polarization encodes the main and ancillary registers, while spatial mode encodes the control registers. The control is realized by preparing the two spatial qubits in

$N(\psi)^{\otimes n}$8

and a newly designed spatial beam splitter (SBS) realizes the Fredkin action optically by routing the two possible channel orderings coherently (Fei et al., 31 Oct 2025). Two Fredkin gates are required: the first places the registers into the two possible channel orderings, and the second recombines them so that the control measurement can select the purified branch.

The reported spatial-mode interference visibility reaches 0.936 when the Pauli rotations are identical (Fei et al., 31 Oct 2025). In tomography of two distinct noise channels, the Pauli transfer matrix element $N(\psi)^{\otimes n}$9, representing the noiseless component, increases from $\psi$0 and $\psi$1 for the two initial channels to $\psi$2 after physical purification and to $\psi$3 after virtual purification. For depolarizing channels, the reported average-fidelity improvement reaches a maximum from 0.744 to 0.913 under virtual purification. In entanglement distribution, the protocol preserves entanglement even in a regime where the original noisy channel destroys it entirely: at $\psi$4, the unpurified output is verified as separable using the PPT criterion, whereas after purification the output retains entanglement with fidelity $\psi$5 (Fei et al., 31 Oct 2025).

The significance is conceptual as well as practical. The protocol is explicitly distinguished from conventional entanglement purification: it targets the noisiest channel itself, uses coherent interference around the channel, and requires no conventional entanglement distillation machinery, quantum memory, or encoding/decoding overhead (Fei et al., 31 Oct 2025).

5. Auxiliary-resource and carrier-assisted entanglement purification

A different forward-assisted line uses auxiliary entanglement as a nonlocal measurement resource. In entanglement-assisted entanglement purification protocols (EIPs), one starts from an ensemble of noisy pairs,

$\psi$6

and uses auxiliary maximally entangled qudit pairs of dimension $\psi$7 to perform entangling nonlocal measurements that determine the number and position of errors without disturbing the good pairs (Sàbat et al., 2020). The central primitive is the bilateral controlled-$\psi$8 “counter gate,”

$\psi$9

For the rank-2 state

$$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$0

the protocol first determines the number of errors with an error number gate using $$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$1, obtaining

$$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$2

and then, when $$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$3, locates the error with an error position gate using $$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$4 (Sàbat et al., 2020). The conditioned yields are

$$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$5

Approximate EIPs extend the method to rank-3 and full-rank states. These protocols are described as especially effective in the few-error regime, for moderate ensemble sizes, and for decay noise; in the reported finite-size regime they can outperform both hashing and recurrence protocols (Sàbat et al., 2020).

Carrier-assisted entanglement purification replaces multiple stored noisy pairs by forward quantum communication. In the simplest CAEPP, Alice and Bob keep one shared noisy Bell-diagonal pair,

$$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$6

apply local pre-processing

$$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$7

encode a carrier qubit initialized in $$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$8 by a CNOT, send the carrier through a qubit channel $$\theta(N^{\otimes n})=\theta^{\mathrm{Post}} \circ \bigl(\theta^{\mathrm{Mem}}\otimes N^{\otimes n}\bigr)\circ \theta^{\mathrm{Pre}}.$$9, and let Bob decode and measure the carrier in the $\cap$0 basis. The round succeeds only when the measurement outcome is $\cap$1 (Kim et al., 9 Sep 2025). When the carrier transmission is noiseless, one successful round yields

$\cap$2

so all odd-parity components are removed. After a second successful round,

$\cap$3

and two successful rounds are therefore sufficient to distill a perfect ebit (Kim et al., 9 Sep 2025).

With noisy carriers, the single-carrier protocol can still improve fidelity, but full convergence to $\cap$4 occurs only for Pauli channels with no $\cap$5-error component,

$\cap$6

for which

$\cap$7

To remove this restriction, the multi-carrier extension mCAEPP encodes $\cap$8 carriers with a stabilizer code and accepts only when all syndrome bits are $\cap$9. The paper states a general condition: with sufficiently many carriers, mCAEPP can purify noisy entanglement to a perfect ebit provided the channel is not entanglement-breaking (Kim et al., 9 Sep 2025).

These two lines illustrate complementary forms of forward assistance. In EIPs the forward resource is auxiliary high-dimensional entanglement used for controlled syndrome extraction; in CAEPP it is one or more traveling carrier qubits used in place of multi-copy coherent storage.

6. Forward assistance in repeaters and network architectures

Forward assistance has also been integrated into repeater architectures. In the purification-enhanced all-photonic repeater based on repeater graph states (RGS), the key enabling primitive is the half-RGS: a group of $F(E,N,S) > F(N,S),$0 inner qubits connected to an anchor qubit, with each inner qubit also connected to an outer qubit. Rather than first assembling full RGSs, the source stations generate multiple half-RGS copies per side, use their anchor qubits as the purification inputs, and only after purification pair the surviving half-RGSs and join them into full RGSs by a CZ gate between anchors followed by $F(E,N,S) > F(N,S),$1-basis measurements on the anchors (Benchasattabuse et al., 25 Apr 2025). The protocol used is optimistic entanglement purification: purification rounds are executed back-to-back without waiting for intermediate heralding, and all classical outcomes are communicated once at the end.

The error model is tracked with vectors

$F(E,N,S) > F(N,S),$2

and the paper gives explicit success probabilities for three purification circuits,

$F(E,N,S) > F(N,S),$3

together with the corresponding post-purification maps and the BSM update rule $F(E,N,S) > F(N,S),$4 (Benchasattabuse et al., 25 Apr 2025). A central latency claim is that optimistic purification reduces the communication latency in the memory requirement from $F(E,N,S) > F(N,S),$5 to $F(E,N,S) > F(N,S),$6, because the protocol avoids waiting after each purification stage. The overhead is described as modest: each purification round requires one additional quantum emitter per half-RGS source, and the RGS generation slows down proportionally with the number of purification rounds (Benchasattabuse et al., 25 Apr 2025).

In a simulated 10-hop chain with 9 intermediate RGSS nodes, node spacing of 2 km, channel loss 0.2 dB/km, 18-arm half-RGSs, tree encoding with branching parameters $F(E,N,S) > F(N,S),$7, depolarizing noise on photons, noiseless gates and measurements, ideal memories, and half-RGS generation tuned to near-deterministic success rates $F(E,N,S) > F(N,S),$8, both purified schemes reach fidelities above 0.9 over the noise range considered, but the purification-enhanced scheme achieves higher fidelity than baseline end-node purification. Its distribution rate exceeds the baseline by a factor of roughly 45 to 65, with both purified schemes using the same number of half-RGSs per hop—five per side per RGSS and end-node side (Benchasattabuse et al., 25 Apr 2025). The result is that all-photonic repeaters need not choose between a memory-free architecture and entanglement purification.

A structurally related, earlier repeater example appears in cavity QED. There, one additional entanglement protocol is inserted inside each repeater node before the main purification step, converting two temporary Bell-diagonal pairs into a correlated four-qubit state $F(E,N,S) > F(N,S),$9. When this resource is fed into the cavity-mediated $n+1$0 purification dynamics, the one-round output fidelity becomes

$n+1$1

compared with the original

$n+1$2

The paper states that the fidelity growth is “almost twice as large” as in the original scheme and that for $n+1$3 the modified protocol yields almost unit output fidelity after three successful rounds, without any controlled-NOT gates (Gonta et al., 2012). This suggests a node-local precursor of later forward-assisted language: an extra entanglement-generation step is moved forward into the repeater node so that the subsequent purification stage acts on a more favorable correlated resource.

7. Hierarchies, no-go circumvention, and full-stack optimization

The formal spatiotemporal framework establishes strict hierarchies of achievable fidelity. The reported bounds include

$n+1$4

and

$n+1$5

Conventional post-processing is therefore always at the weakest end of the hierarchy (Meng et al., 2 Jun 2026). The same work reports a headline separation: a single-copy UA protocol can beat 5-copy conventional post-processing for certain ensembles, and in some regimes it beats up to 50-copy conventional purification; extrapolated crossover points are reported around $n+1$6, $n+1$7, $n+1$8, or $n+1$9, depending on the ensemble (Meng et al., 2 Jun 2026). For qubits, symmetry reduction by Schur-Weyl methods and Clebsch-Gordan recursion yields constructive scaling

$N(\psi)$00

which makes benchmarking to $N(\psi)$01 copies feasible (Meng et al., 2 Jun 2026).

A major conceptual claim is that forward assistance can circumvent no-purification theorems that apply to the conventional static setting. For Bell states under local depolarizing noise, the paper recalls that conventional PPT post-processing alone gives no improvement, but adding pre-processing $N(\psi)$02 before the entangling/noisy stage can yield

$N(\psi)$03

leading to the theorem that Bell states under local depolarizing noise can be efficiently purified in the 2-to-1 setting using UA/FA strategies (Meng et al., 2 Jun 2026). The mechanism is described as symmetry breaking by pre-processing: the effective purification problem is reshaped before the noise acts.

A related but indirect development is full-stack optimization of purification circuits for downstream tasks. A faster-than-Clifford simulator for Bell-diagonal purification circuits achieves $N(\psi)$04 gate-simulation complexity per step and enables practical optimization of $N(\psi)$05-to-$N(\psi)$06 circuits by a genetic algorithm (Addala et al., 2023). The central insight is that optimizing only for output Bell-pair fidelity can be misleading when the purified pairs are later consumed by error correction. The paper therefore distinguishes

$N(\psi)$07

where $N(\psi)$08 is the logical fidelity after teleportation and error correction, and shows that optimizing for $N(\psi)$09 or even $N(\psi)$10 can reduce $N(\psi)$11 because correlated error patterns matter to the downstream code (Addala et al., 2023). This does not introduce a forward-assisted protocol by name, but it reinforces the same design principle: purification should be optimized in anticipation of the process that follows it.

Taken together, these results define forward-assisted purifications as a broad shift in purification theory and practice. The shift is from static recovery on noisy outputs to process-aware architectures in which auxiliary systems, coherent control, pre-processing, path-level scheduling, or forward communication are used to reshape the noise problem before final recovery is attempted.