Quantum turbo codes reveal a trade-off between decoding rounds and computing cost

Building protection from smaller blocks

How can quantum error correction protect several logical qubits without devoting an overwhelming number of physical qubits to the task? Khaled Jebari and colleagues investigate one approach in High Rate Quantum Block Turbo Codes, an arXiv preprint posted on November 9, 2026. The work combines decoding-complexity analysis with simulations, rather than demonstrating a working quantum device. The supplied source does not establish peer-reviewed publication.

A logical qubit is quantum information encoded across physical qubits for protection. The code rate measures how many logical qubits a code carries per physical qubit. A higher rate means less physical-qubit overhead by that measure, but does not by itself establish stronger protection or easier decoding.

The researchers study quantum block turbo codes: constructions made by combining two smaller, constituent error-correcting codes through successive row and column encoding operations. Their decoder repeatedly exchanges probability estimates between two components, allowing each to refine its interpretation of the errors.

The paper lists combinations with rates around one-fifth. One example encodes 25 logical qubits into 121 physical qubits; another encodes 49 into 225. These are code constructions, not demonstrated hardware capacities.

What makes decoding “turbo”?

The two decoding components use error-check measurements, called syndromes, together with probabilities for possible errors. Rather than passing only a single best guess, they exchange probabilistic information. Each additional round can improve the estimate of which logical error occurred.

The authors assess performance over a quantum depolarizing channel, their chosen noise model. They track word error rate, a measure of failures at the encoded-block level, as the noise rate and number of decoding rounds change.

Across the tested constructions, the largest improvement generally occurs between the first and second rounds. Constituents carrying more logical qubits show clearer additional gains between the second and third rounds, particularly the combinations encoding 25 and 49 logical qubits overall.

More rounds are not automatically better. For the smallest example, built from two four-physical-qubit constituents, improvement largely saturates after the second round. The third can even perform worse. The authors suggest this behavior is related to the constituents’ small code distance, a parameter describing their error-protection capability; that explanation remains an interpretation.

Distance also matters separately from iterative improvement. Two constructions carrying nine logical qubits have broadly similar performance-curve shapes, but the higher-distance construction achieves lower word error rates in the reported comparison.

Order changes the computational bill

The constituent closest to the noisy channel is called the inner code. Swapping the inner and outer codes can change both convergence and decoding cost.

In one comparison, putting a constituent carrying just one logical qubit in the inner position produces little improvement from repeated decoding. Reversing the arrangement restores an iterative gain, with the two arrangements eventually reaching similar reported performance.

Across other tested pairs, a smaller inner code generally reaches its performance sooner in terms of decoding rounds. A larger inner code tends to need more rounds but costs less computation per round.

For constituents encoding three logical qubits into eight physical qubits and five into eleven, the authors’ operation-count estimates differ by roughly a factor of 3.5 per round between orderings. That is an algorithmic estimate—not a measured hardware speedup. Fewer rounds therefore cannot automatically be equated with a faster decoder overall.

Promising constructions, bounded evidence

The complexity analysis identifies an important limitation: exact decoding remains exponential in the sizes of the constituent codes, though not directly in the total product-code size. Small constituents can therefore keep decoding manageable even when the combined block is considerably larger.

The paper also proposes reducing work by discarding sufficiently unlikely error candidates. Such truncation introduces an approximation; it is not a free reduction in the cost of exact decoding.

The findings concern the tested codes and depolarizing-noise simulations. They do not establish performance on a particular quantum processor. The authors also conjecture that different orderings will eventually converge to the same performance after sufficiently many rounds; they do not prove that generally.

The practical lesson is a design trade-off: code rate, error protection, decoding rounds and work per round must be considered together. Choosing the constituent codes is only part of the problem. Choosing their order matters too.