Quantum Decoder for Turbo Codes

• Quantum decoder for turbo codes is an iterative decoding framework that employs turbo-style concatenation, trellis or factor-graph representations, and soft information exchange.
• It leverages classical BCJR-like algorithms and entanglement assistance to achieve non-catastrophic, recursive behavior, ensuring improved error correction near quantum capacity limits.
• Variants such as coherent BPQM, quantum-inspired iterative decoders, and variational quantum circuit receivers extend the architecture to diverse channel models and performance trade-offs.

Searching arXiv for the cited papers to ground the article in the latest indexed metadata. Using arXiv search for the core references: (Wilde et al., 2010, Piveteau et al., 2021, Franck, 2019, Liu et al., 2022), and (Chandra et al., 2019). A quantum decoder for turbo codes is an iterative decoding architecture built around turbo-style concatenation, trellis or factor-graph representations, and repeated exchange of soft or quantum information between constituent decoders. In the most direct quantum error-correction sense, it denotes the soft-input soft-output decoder for quantum serial turbo codes, especially the entanglement-assisted construction in which two quantum convolutional encoders are connected by a quantum interleaver and decoded by BCJR-like constituent decoders operating on Pauli error probabilities and measured syndromes. Related literature uses the same phrase for coherent message-passing decoders on pure-state classical-quantum channels, for quantum-inspired iterative decoders on trellis-constrained codes, and for variational quantum circuits embedded in turbo equalization loops. Taken together, these works suggest that the topic is not a single algorithm but a family of decoder models unified by iterative inference on concatenated graphical structures (Wilde et al., 2010, Piveteau et al., 2021, Franck, 2019, Liu et al., 2022, Chandra et al., 2019).

1. Architectural scope and terminological boundaries

The canonical quantum turbo-code architecture is a serial concatenation of two quantum convolutional encoders joined by an interleaver. In the entanglement-assisted setting, the encoders may act on information qubits, ancilla qubits, and shared ebits. The inner encoder is closer to the physical channel, while the outer encoder is farther from the channel. The constituent decoders are soft-input soft-output decoders that exchange soft information iteratively, in direct analogy with classical turbo decoding, but the latent variables are Pauli errors and syndrome constraints rather than binary symbol errors alone (Wilde et al., 2010).

Other papers broaden the scope. Trellis-Constrained Codes are defined as the intersection of two trellis-defined codes with an interleaver constraint, so Turbo codes appear as a subclass of this more general family. In a different direction, coherent quantum decoding on pure-state classical-quantum channels models turbo-like decoding through cyclic factor graphs and computation trees. A further shift occurs in variational-quantum work, where “turbo” refers to turbo equalization with iterative exchange between a detector and a decoder rather than to two recursive systematic convolutional constituents (Franck, 2019, Piveteau et al., 2021, Liu et al., 2022).

Paradigm	Code or graph structure	Decoder carrier
Entanglement-assisted quantum turbo code	Serially concatenated quantum convolutional encoders with quantum interleaver	Classical APP and extrinsic probabilities over Pauli errors
QSBC-QURC short-block quantum turbo code	Outer QSBC, random interleaver, inner QURC	Classical SISO exchange with BCJR and parity-node updates
BPQM turbo-like decoding	Unrolled computation tree of a cyclic factor graph	Quantum messages consisting of data qubits and angle registers
TCC amplitude-amplification decoder	Intersection of two trellis-defined codes with interleaver constraint	Classical weights and partition functions
VQC turbo detection	EMA-OAMP receiver with coded constraints	Soft outputs from a QAOA-based variational quantum circuit

A common misconception is that every “quantum turbo decoder” is a coherent quantum circuit acting on a quantum code block. The literature distinguishes at least three meanings: a classical post-processing decoder for a quantum error-correcting turbo code, a coherent quantum message-passing decoder for a classical-quantum channel, and a hybrid or quantum-inspired turbo receiver. The distinction is substantive because the channel model, inferential object, and implementation constraints differ across these settings (Wilde et al., 2010, Piveteau et al., 2021, Liu et al., 2022).

2. Soft-input soft-output decoding for entanglement-assisted quantum turbo codes

In entanglement-assisted quantum turbo codes, each constituent encoder is specified by a seed transformation $U$ acting on memory qubits, information qubits, ancillas, and ebits, producing new memory and physical channel qubits. In the binary Pauli representation, a single frame transformation is

$(M':P) = (M : L : S : E) U.$

The associated state diagram is a directed multigraph with $4^m$ vertices, where $m$ is the number of memory qubits. An edge $M \to M'$ with label $(L,P)$ exists if there is $S^z \in \{I,Z\}^a$ such that

$(M':P) = (M : L : S^z : I_c) U.$

This state-diagram description governs both minimum-distance analysis and the BCJR-like forward-backward decoding (Wilde et al., 2010).

Each constituent decoder accepts a priori probabilities for logical errors and physical errors, together with measured syndrome bits, and returns a posteriori probabilities for logical and physical errors. The modified decoding algorithm exchanges only extrinsic information, rather than the a posteriori probabilities exchanged in the decoder of Poulin et al. The probability-domain update is

$$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$

and the log-domain form is

$$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$

The stated motivation is to avoid detrimental positive feedback caused by reusing the same a priori information across iterations (Wilde et al., 2010).

The inner decoder uses the physical noise model, the measured syndrome for the inner code, and an a priori belief about logical errors. It outputs extrinsic information for the interleaved bits, which becomes the a priori input to the outer decoder. The outer decoder similarly generates extrinsic information that is deinterleaved and fed back to the inner decoder. Iterations continue until a hard decision stabilizes or a maximum number of iterations is reached; the reported empirical bound is $(M':P) = (M : L : S : E) U.$0 iterations (Wilde et al., 2010).

The APP computation is BCJR-like over the encoder’s state diagram. For each trellis time $(M':P) = (M : L : S : E) U.$1 and edge $(M':P) = (M : L : S : E) U.$2 with label $(M':P) = (M : L : S : E) U.$3, the branch metric satisfies

$(M':P) = (M : L : S : E) U.$4

where the memoryless depolarizing channel is

$(M':P) = (M : L : S : E) U.$5

For a single-qubit Pauli symbol $(M':P) = (M : L : S : E) U.$6,

$(M':P) = (M : L : S : E) U.$7

The multiqubit probability is the product over qubits, because the channel is memoryless (Wilde et al., 2010).

Measured syndromes enter through the decomposition

$(M':P) = (M : L : S : E) U.$8

with $(M':P) = (M : L : S : E) U.$9 and $4^m$0. Ancilla $4^m$1 bits are obtained from $4^m$2-basis measurements, while ebit Bell outcomes $4^m$3 identify the ebit Pauli error with no degeneracy on ebits. This is the technical point that allows the decoder to enforce branch consistency on auxiliary resources more strongly than in the unassisted case (Wilde et al., 2010).

The computational profile is explicit: the trellis has $4^m$4 states, per-iteration complexity scales linearly with blocklength $4^m$5 for fixed $4^m$6, and overall complexity grows exponentially in $4^m$7. The decoder is entirely classical in the sense that it processes syndromes and computes probabilities over the state diagram; the quantum processing lies in encoding, transmission, and syndrome acquisition rather than in the iterative decoder itself (Wilde et al., 2010).

3. Encoder properties, entanglement assistance, and reported performance

Two structural properties dominate the theory of quantum turbo decoding: non-catastrophicity and recursiveness. An encoder is non-catastrophic if every zero physical-weight cycle in the state diagram has zero logical weight. An encoder is quasi-recursive if every weight-one logical Pauli input maps to an infinite-weight physical output, and recursive if every element in the logical cosets $4^m$8, $4^m$9, and $m$0 has infinite physical-weight response. These properties are linked to minimum-distance growth and iterative-decoder convergence, respectively (Wilde et al., 2010).

The paper proves that encoders acting only on information qubits, classical bits, gauge qubits, and ancillas cannot be simultaneously recursive and non-catastrophic. Its corollary states that if such an encoder is recursive, it is catastrophic. Entanglement assistance is identified as the enabling resource: Bell-measurement syndromes on ebits remove degeneracy on the auxiliary resource, permitting both properties to coexist in a single convolutional encoder (Wilde et al., 2010).

This structural result has direct decoding consequences. In unassisted codes, ancilla-$m$1 degeneracy can make some cycles invisible to the syndrome, which undermines iterative correction. In the entanglement-assisted setting, Bell measurements distinguish all four Pauli errors on an ebit half, so the decoder’s branch-metric consistency test excludes precisely the problematic hidden cycles. This suggests that entanglement assistance changes not only achievable rates but the inferential geometry of the trellis itself (Wilde et al., 2010).

Concrete code families illustrate the effect. PTO1R is an unassisted encoder with $m$2 memory qubits, $m$3 information qubit, $m$4 ancillas, and $m$5 physical outputs; serial concatenation yields a rate $m$6 turbo code. PTO1REA is its entanglement-assisted version, obtained by replacing ancillas with ebits, and becomes recursive with improved distance spectrum. PTO3R and PTO3REA form the analogous rate-$m$7 family, with $m$8, $m$9, $M \to M'$0, and $M \to M'$1 physical outputs (Wilde et al., 2010).

The reported benchmarks are specific. For the rate-$M \to M'$2 unassisted PTO1R/PTO1R code, the pseudothreshold is approximately $M \to M'$3, within $M \to M'$4 dB of the unassisted hashing bound noise limit for rate $M \to M'$5 at approximately $M \to M'$6. For the fully assisted PTO1REA/PTO1REA “father protocol,” with ebit rate $M \to M'$7, the threshold is approximately $M \to M'$8, which is $M \to M'$9 dB beyond the unassisted turbo-code pseudothreshold and within $(L,P)$0 dB of the entanglement-assisted hashing limit at approximately $(L,P)$1. An inner-assisted rate-$(L,P)$2 configuration with PTO1REA inner and PTO1R outer has threshold approximately $(L,P)$3 and lies within $(L,P)$4 dB of the hashing-region boundary at approximately $(L,P)$5 for $(L,P)$6 and $(L,P)$7. An outer-assisted configuration has pseudothreshold approximately $(L,P)$8, at $(L,P)$9 dB from the boundary at approximately $S^z \in \{I,Z\}^a$0 for $S^z \in \{I,Z\}^a$1 and $S^z \in \{I,Z\}^a$2 (Wilde et al., 2010).

For the rate-$S^z \in \{I,Z\}^a$3 PTO3 family, the reported thresholds are approximately $S^z \in \{I,Z\}^a$4 for the unassisted code, $S^z \in \{I,Z\}^a$5 for the fully assisted code with ebit rate $S^z \in \{I,Z\}^a$6, $S^z \in \{I,Z\}^a$7 for the inner-assisted case with ebit rate $S^z \in \{I,Z\}^a$8, and $S^z \in \{I,Z\}^a$9 for the outer-assisted case with ebit rate $(M':P) = (M : L : S^z : I_c) U.$0. Their distances to hashing limits are approximately $(M':P) = (M : L : S^z : I_c) U.$1 dB, $(M':P) = (M : L : S^z : I_c) U.$2 dB, $(M':P) = (M : L : S^z : I_c) U.$3 dB, and $(M':P) = (M : L : S^z : I_c) U.$4 dB, respectively. Experiments with ebit noise rates $(M':P) = (M : L : S^z : I_c) U.$5, $(M':P) = (M : L : S^z : I_c) U.$6, and $(M':P) = (M : L : S^z : I_c) U.$7 further show that placing ebits in the inner encoder maintains better performance under ebit noise than placing them in the outer encoder, even though the inner-assisted construction consumes more ebits (Wilde et al., 2010).

The paper also contrasts its modified extrinsic-information decoder with the original decoder of Poulin et al. The stated impact is “significant performance gains in unassisted turbo codes compared to PTO09,” including lower WER and a steeper waterfall, while in entanglement-assisted codes thresholds approach hashing limits, often within approximately $(M':P) = (M : L : S^z : I_c) U.$8 dB. EXIT-chart analysis is cited as supporting the claim that extrinsic-information exchange opens the decoding tunnel and yields true thresholds in the rate-$(M':P) = (M : L : S^z : I_c) U.$9 father-code case (Wilde et al., 2010).

4. Short-block quantum turbo decoding: the QSBC-QURC construction

A distinct quantum turbo-decoder line is the QSBC-QURC architecture, where the outer code is a quantum single-parity-check code and the inner code is a quantum unity-rate code. The outer QSBC is a dual-containing CSS code built from classical SPC codes, with

$$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$0

so that the block stabilizers are

$$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$1

Its rates are $$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$2 with minimum distance $$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$3, and examples include $$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$4, $$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$5, and $$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$6. The inner QURC is a non-recursive, non-catastrophic quantum convolutional unity-rate code with seed transformation

$$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$7

Serial concatenation with a random interleaver produces the QSBC-QURC quantum turbo code (Chandra et al., 2019).

Decoding begins by applying the inner inverse encoder $$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$8 to the corrupted block. Because the inner code is unity rate, there are no measured syndromes at the inner stage:

$$P^e(L^j_i) = N_{L^j}\cdot\left[\frac{P^o(L^j_i)}{P^a(L^j_i)}\right], \qquad P^e(P^j_i) = N_{P^j}\cdot\left[\frac{P^o(P^j_i)}{P^a(P^j_i)}\right],$$9

After deinterleaving, the outer inverse encoder $$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$0 is applied and the QSBC stabilizer ancilla is measured to obtain $$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$1:

$$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$2

The first inner iteration uses uniform a priori information $$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$3, while the per-qubit channel priors are those of the depolarizing channel, $$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$4 and $$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$5 (Chandra et al., 2019).

The inner SISO decoder is a BCJR decoder on the QURC trellis. Its forward and backward recursions are

$$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$6

with branch metric

$$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$7

Since the inner stage has no measured syndrome, the metric depends only on channel priors and a priori logical information. The outer QSBC SISO instead enforces stabilizer parity constraints. For the dual-containing QSBC,

$$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$8

and the Bayesian update is

$$\ln[P^e(L^j_i)] = \ln[P^o(L^j_i)] - \ln[P^a(L^j_i)], \qquad \ln[P^e(P^j_i)] = \ln[P^o(P^j_i)] - \ln[P^a(P^j_i)].$$9

where $(M':P) = (M : L : S : E) U.$00 is $(M':P) = (M : L : S : E) U.$01 for parity-consistent errors and $(M':P) = (M : L : S : E) U.$02 otherwise (Chandra et al., 2019).

The decoder exchanges extrinsic information between the inner QURC SISO and outer QSBC SISO through the interleaver and deinterleaver. The paper emphasizes normalization of $(M':P) = (M : L : S : E) U.$03 and $(M':P) = (M : L : S : E) U.$04, optional extrinsic LLR damping $(M':P) = (M : L : S : E) U.$05, and a practical stopping regime of $(M':P) = (M : L : S : E) U.$06 iterations, with EXIT trajectories typically saturating after about $(M':P) = (M : L : S : E) U.$07 iterations. Because the inner QURC is non-recursive, the inner EXIT curve terminates at $(M':P) = (M : L : S : E) U.$08 with $(M':P) = (M : L : S : E) U.$09, which yields a narrow tunnel near $(M':P) = (M : L : S : E) U.$10 and a residual error floor (Chandra et al., 2019).

Performance is reported relative to the quantum hashing bound

$(M':P) = (M : L : S : E) U.$11

For the half-rate case, the bound is approximately $(M':P) = (M : L : S : E) U.$12. The distance to the hashing bound is defined as

$(M':P) = (M : L : S : E) U.$13

where $(M':P) = (M : L : S : E) U.$14 is the operational depolarizing probability at which QBER $(M':P) = (M : L : S : E) U.$15 is achieved. For $(M':P) = (M : L : S : E) U.$16 logical qubits and $(M':P) = (M : L : S : E) U.$17 iterations, QSBC-QURC reaches QBER $(M':P) = (M : L : S : E) U.$18 at approximately $(M':P) = (M : L : S : E) U.$19, giving $(M':P) = (M : L : S : E) U.$20, while the QIrCC-QURC baseline operates at approximately $(M':P) = (M : L : S : E) U.$21, giving $(M':P) = (M : L : S : E) U.$22. For the multi-rate construction, the reported values are $(M':P) = (M : L : S : E) U.$23 at rate $(M':P) = (M : L : S : E) U.$24 and $(M':P) = (M : L : S : E) U.$25 at rate $(M':P) = (M : L : S : E) U.$26 (Chandra et al., 2019).

The multiple-rate feature is central. The same scalable QSBC structure supports rate switching among $(M':P) = (M : L : S : E) U.$27 by changing the blocklength and corresponding all-ones stabilizer. The paper gives explicit switching ranges for $(M':P) = (M : L : S : E) U.$28 and $(M':P) = (M : L : S : E) U.$29 iterations: for target QBER $(M':P) = (M : L : S : E) U.$30, use rate $(M':P) = (M : L : S : E) U.$31 for $(M':P) = (M : L : S : E) U.$32, rate $(M':P) = (M : L : S : E) U.$33 for $(M':P) = (M : L : S : E) U.$34, rate $(M':P) = (M : L : S : E) U.$35 for $(M':P) = (M : L : S : E) U.$36, and uncoded transmission below $(M':P) = (M : L : S : E) U.$37. This is presented as the first instantiation of QTCs capable of adjusting the quantum encoders according to the required quantum coding rate (Chandra et al., 2019).

5. Coherent quantum message passing on turbo-like factor graphs

A fully coherent notion of a quantum decoder for turbo codes arises in belief propagation with quantum messages. In this framework, the channel is a pure-state classical-quantum channel with binary input $(M':P) = (M : L : S : E) U.$38 and qubit output

$(M':P) = (M : L : S : E) U.$39

For a binary linear codeword $(M':P) = (M : L : S : E) U.$40 and possibly non-uniform channel parameters $(M':P) = (M : L : S : E) U.$41, the receiver obtains the product state $(M':P) = (M : L : S : E) U.$42. On tree Tanner graphs, BPQM is proved block-optimal: it realizes the optimal measurement, identified as the pretty-good measurement, for discriminating the codeword outputs (Piveteau et al., 2021).

The local operations correspond to factor-graph primitives. Equality nodes apply a two-qubit unitary $(M':P) = (M : L : S : E) U.$43 satisfying

$(M':P) = (M : L : S : E) U.$44

with composite angle

$(M':P) = (M : L : S : E) U.$45

Check nodes apply a CNOT and induce the conditioned angle update

$(M':P) = (M : L : S : E) U.$46

with branch probability

$(M':P) = (M : L : S : E) U.$47

The final decision for a target bit is a Helstrom measurement in the $(M':P) = (M : L : S : E) U.$48 basis (Piveteau et al., 2021).

The original BPQM description contained a global-control flaw: equality-node operations depend on ancillas produced by all preceding check nodes, so naive circuit realizations are exponentially large in the code dimension. The corrected design makes angle information local by attaching a finite-precision angle register to each message and updating it on the fly. With $(M':P) = (M : L : S : E) U.$49-qubit angle registers, choosing

$(M':P) = (M : L : S : E) U.$50

ensures success probability within $(M':P) = (M : L : S : E) U.$51 of ideal BPQM, and the decoder’s circuit complexity becomes

$(M':P) = (M : L : S : E) U.$52

A detailed bound given in the paper is depth $(M':P) = (M : L : S : E) U.$53 and width $(M':P) = (M : L : S : E) U.$54 (Piveteau et al., 2021).

Turbo-like graphs are cyclic, so exact tree-based BPQM does not apply directly. The proposed extension unrolls the cyclic graph around a target bit for $(M':P) = (M : L : S : E) U.$55 rounds into a computation tree and supplies duplicated leaves by approximate cloning. Two cloning options are described. The ENU cloner uses $(M':P) = (M : L : S : E) U.$56 with

$(M':P) = (M : L : S : E) U.$57

for the $(M':P) = (M : L : S : E) U.$58 case. An “optimal $(M':P) = (M : L : S : E) U.$59 cloner for the two-state ensemble” is also described, and the effective $(M':P) = (M : L : S : E) U.$60 used by BPQM may be tuned numerically. The resulting turbo-decoding outline is explicit: construct the Turbo factor graph from two constituent convolutional trellises and an interleaver, unroll for $(M':P) = (M : L : S : E) U.$61 rounds, clone repeated leaves, run BPQM node updates, and measure the root qubit (Piveteau et al., 2021).

The paper stresses that increasing $(M':P) = (M : L : S : E) U.$62 does not necessarily improve performance, because more unrolling requires more approximate cloning. Empirically, a small $(M':P) = (M : L : S : E) U.$63, such as $(M':P) = (M : L : S : E) U.$64 or $(M':P) = (M : L : S : E) U.$65, balances additional context against clone-induced degradation. There are no general convergence guarantees on cyclic graphs. Numerical results on an $(M':P) = (M : L : S : E) U.$66-bit cyclic code nevertheless show that BPQM with approximate cloning can significantly outperform the best classical decoder that first measures each qubit optimally and then runs MAP or maximum-likelihood decoding, and can approach the optimal global quantum decoder for moderate channel overlap (Piveteau et al., 2021).

This coherent BPQM line differs fundamentally from Pauli-syndrome turbo decoding. It does not decode a stabilizer syndrome over a depolarizing channel; it performs collective quantum discrimination of codeword states on a pure-state classical-quantum channel. A plausible implication is that “quantum decoder for turbo codes” can refer either to a decoder for quantum codes or to a quantum-mechanical decoder for classical-coded communication, depending on whether the quantum structure lies in the code or in the receiver (Piveteau et al., 2021).

6. Quantum-inspired and variational turbo receivers

The amplitude-amplification-inspired decoder for Trellis-Constrained Codes is explicitly “quantum-inspired, not a quantum circuit.” A TCC is defined as

$(M':P) = (M : L : S : E) U.$67

where both constituent codes admit low-complexity trellis representations and $(M':P) = (M : L : S : E) U.$68 is the interleaved version of $(M':P) = (M : L : S : E) U.$69. The decoder maintains two weight vectors, initialized as

$(M':P) = (M : L : S : E) U.$70

and alternates an even-iteration $(M':P) = (M : L : S : E) U.$71-step

$(M':P) = (M : L : S : E) U.$72

with an odd-iteration $(M':P) = (M : L : S : E) U.$73-step

$(M':P) = (M : L : S : E) U.$74

Its key guarantee is monotonic improvement of the relative likelihood of the best codeword,

$(M':P) = (M : L : S : E) U.$75

with bitwise updates derived from BCJR-computed $(M':P) = (M : L : S : E) U.$76-sums on the constituent trellises. The paper states that per-iteration complexity is comparable to one Turbo decoding iteration with BCJR or Log-MAP, but its experiments section is “TBD,” so no BER or FER curves are reported (Franck, 2019).

A different strand is the variational quantum circuit turbo receiver built around EMA-OAMP. Here the channel model is $(M':P) = (M : L : S : E) U.$77 with BPSK symbols $(M':P) = (M : L : S : E) U.$78, and the VQC acts as a channel decoder within the turbo loop rather than as a decoder for a convolutional turbo code. The cost Hamiltonian is

$(M':P) = (M : L : S : E) U.$79

with mixer

$(M':P) = (M : L : S : E) U.$80

The QAOA ansatz uses one qubit per information bit and produces Pauli-$(M':P) = (M : L : S : E) U.$81 expectations $(M':P) = (M : L : S : E) U.$82, which are mapped to soft decisions by

$(M':P) = (M : L : S : E) U.$83

and

$(M':P) = (M : L : S : E) U.$84

These are converted to coded-bit LLRs and fed back to the OAMP detector as extrinsic information (Liu et al., 2022).

The proposed learning-to-learn framework uses an LSTM meta-optimizer to predict QAOA parameters from previous parameters, a cost estimate, weighted syndromes, and magnitudes of the detector outputs. The training objective is an exponentially decayed binary cross-entropy with decay factor $(M':P) = (M : L : S : E) U.$85 in experiments, optimized by Adamax with learning rate $(M':P) = (M : L : S : E) U.$86, LSTM depth $(M':P) = (M : L : S : E) U.$87, time steps $(M':P) = (M : L : S : E) U.$88 for the LDPC code and $(M':P) = (M : L : S : E) U.$89 for the BKLC, and QAOA depth $(M':P) = (M : L : S : E) U.$90 and $(M':P) = (M : L : S : E) U.$91, respectively. Reported performance is within approximately $(M':P) = (M : L : S : E) U.$92 dB of exact joint ML detection-and-decoding after $(M':P) = (M : L : S : E) U.$93 EMA-OAMP iterations, for short codes with $(M':P) = (M : L : S : E) U.$94, $(M':P) = (M : L : S : E) U.$95 (Liu et al., 2022).

The paper is explicit that no convolutional constituent RSC encoders or interleaver typical of classical turbo codes are used; “turbo” refers to turbo equalization with iterative exchange between detector and decoder. This is important for classification. The VQC turbo detector belongs to the broader family of quantum-assisted iterative receivers, but not to the narrower category of quantum serial turbo-code decoders in the sense of entanglement-assisted quantum convolutional coding (Liu et al., 2022).

7. Limitations, design principles, and open problems

Several limitations recur across the literature. In entanglement-assisted quantum turbo decoding, complexity is linear in blocklength only for fixed memory size and still exponential in the number of memory qubits because of the $(M':P) = (M : L : S : E) U.$96 trellis state space. The paper therefore recommends constraining $(M':P) = (M : L : S : E) U.$97, and the reported experiments use $(M':P) = (M : L : S : E) U.$98. Robustness to mismatch in the depolarizing parameter $(M':P) = (M : L : S : E) U.$99 is not explicitly studied. Interleaver design remains open, and random interleavers are used in the reported simulations (Wilde et al., 2010).

Design guidance is correspondingly specific. For strong waterfall performance and minimum-distance growth, the recommendation is to choose a recursive, non-catastrophic inner encoder made possible by entanglement assistance, together with a non-catastrophic outer encoder with high free distance. Inner-assisted designs are also reported to be more robust than outer-assisted designs under moderate ebit-noise levels $4^m$00. The trade-off is that ebits are a precious resource, so entanglement consumption rate $4^m$01 must be treated as a design parameter rather than a negligible overhead (Wilde et al., 2010).

In the QSBC-QURC line, the central limitation is the residual error floor caused by the outer QSBC minimum distance $4^m$02 and by the inner non-recursive design, whose EXIT curve terminates at $4^m$03 with $4^m$04. The benefit is rate compatibility and simple stabilizer structure, but deep low-$4^m$05 behavior may still favor alternatives such as QIrCC-QURC. This suggests a structural trade-off between scalability and low-floor performance in short-block quantum turbo design (Chandra et al., 2019).

For coherent BPQM-based turbo decoding, the main restrictions are the pure-state classical-quantum channel model, the absence of guarantees on cyclic graphs, and the cost of approximate cloning. The paper states that more unrolling can degrade performance because it increases clone usage. Hardware requirements include coherent angle-register storage, controlled rotations, reversible arithmetic, and sufficiently low noise to preserve the gain over classical decoders. Extending BPQM to mixed-state outputs, realistic optical systems, and larger cyclic graphs is left open (Piveteau et al., 2021).

The amplitude-amplification-inspired decoder and the QAOA-based turbo detector each expose a different uncertainty. In the former, formal convergence-rate analysis and numerical performance evidence remain open because the preliminary version leaves experiments and conclusions as “TBD.” In the latter, larger block lengths require more qubits and higher-order cost terms, while decoherence and readout bias directly perturb the soft outputs used in the iterative loop. Both cases indicate that “quantum” advantages in turbo-like decoding remain strongly dependent on the exact channel model, hardware assumptions, and graphical structure under consideration (Franck, 2019, Liu et al., 2022).

Across all of these lines, the unifying research question is how to preserve the iterative, extrinsic-information logic of turbo decoding while exploiting genuinely quantum resources. In entanglement-assisted quantum turbo codes, the crucial resource is shared entanglement, which removes degeneracy on the auxiliary resource and permits simultaneous recursiveness and non-catastrophicity. In BPQM, the crucial resource is joint quantum measurement on channel outputs. In quantum-inspired and variational schemes, the quantum contribution is instead algorithmic analogy or a learned variational measurement. The literature therefore supports no single canonical definition of a quantum decoder for turbo codes; it supports a structured family of decoder paradigms whose common elements are trellises, interleavers, iterative inference, and the search for near-capacity operation under quantum constraints (Wilde et al., 2010, Piveteau et al., 2021, Chandra et al., 2019).