Quantum Computers Edge Closer to Reality with Boosted Error Correction

Syndrome resampling enhances error correction thresholds for any decoder and diminishes logical errors without requiring new hardware or code-specific alterations. Luis Colmenarez of the Institute for Theoretical Nanoelectronics (PGI-2) and RWTH Aachen University, and colleagues, use the correlation between improbable syndromes and potential logical failures to effectively bias syndrome measurements, improving logical fidelities. Simulations utilising surface codes reveal sharp increases in thresholds and reductions in logical error rates of up to four orders of magnitude, providing potential for practical improvements in near-term quantum error correction experiments.

Syndrome resampling boosts quantum error correction performance via R enyi coherent information

Low-probability syndromes frequently lead to logical failure, so biasing syndrome averages towards the most likely syndromes effectively increases logical fidelities. Connecting the R enyi coherent information (RCI) with powers of the syndrome probability distribution reveals that resampling syndromes according to these powers, combined with maximum likelihood decoding (MLD), realises a family of optimal thresholds associated with phase transitions in the RCI. Numerical simulations of surface codes demonstrate that syndrome resampling substantially increases thresholds for both optimal and suboptimal decoders, and reduces logical error rates by up to four orders of magnitude in experimentally relevant regimes. This combination with decoding-based post-selection achieves additional gains from finite data.

Quantum error correction (QEC) is central to fault-tolerant (FT) quantum computation, introducing redundant degrees of freedom to detect and correct errors by encoding one or more logical qubits into many noisy physical qubits. Recent experiments utilising superconducting qubits, trapped ions, and neutral atoms have demonstrated key QEC protocols, marking the onset of the early FT era. Applying this method to existing experimental QEC data yields up to two orders of magnitude reduction in logical error rates without requiring additional measurements.

Scaling QEC to large FT algorithms requires device noise to lie below certain thresholds. While several experiments report operation below these thresholds, further progress demands both more physical qubits and lower noise levels. Suppressing logical errors using methods that do not require substantial additional hardware resources is therefore desirable. The Maximum Likelihood Decoder (MLD) achieves optimal performance, but its computational intractability reflects fundamental limits in the coherent information of the noisy quantum state associated with the QEC code.

An alternative approach is post-selection (PS), where runs exhibiting detected errors are discarded. However, the probability of obtaining error-free runs decays exponentially with the number of faulty operations. Recently, intermediate strategies between decoding and PS have been proposed, including confidence-based decoding, heuristic post-selection rules, and different forms of logical error mitigation. These methods are typically tailored to specific codes and decoders, require additional overhead, and lack a unifying theoretical framework to assess their impact on QEC thresholds.

This method establishes a connection between the R enyi Coherent Information (RCI) and powers of the syndrome probability distribution (SPD). Based on this relation, syndrome resampling (SR) is introduced as a general method to enhance thresholds and substantially improve logical fidelities, particularly in the regime of error rates of current devices. A family of thresholds associated with phase transitions in the RCI is identified, characterising the optimal performance attainable after SR. Unlike previous methods, this approach does not rely on decoding confidence or code-specific information beyond the syndrome statistics. SR is benchmarked in a setting where the SPD can be computed exactly, and a practical scheme to estimate it from syndrome measurements is outlined.

Although the number of required samples can be large, it depends on the noise strength and the desired accuracy of the SPD estimate. Combining SR with PS based on the complementary gap yields additional gains in logical fidelity in experimentally relevant regimes. The RCI and powers of the SPD are defined as follows: A QEC code with parameters [[n, k, d]] is defined as the common eigenspace of the stabilizer generators Si, with i = 1, ., n−k, choosing the codespace corresponding to the +1 eigenspace of all stabilizers.

The logical operators (OX Lj, OZ Lj), with j = 1, ., k, represent the Pauli X and Z acting on the k logical qubits encoded in the QEC code. A maximally entangled state ρ0 RQ is considered between the k logical qubits, supported on the n-qubit physical system Q, and a k-qubit reference system R. This state is a stabilizer state given by the common state of the Si = +1 stabilizers and all products OX,Z Rj OX,Z Lj = +1, equivalently written as ρ0 RQ = |s0, B0⟩⟨s0, B0|, where s0 denotes the trivial syndrome and B0 labels a Bell basis state between reference and logical qubits. The RCI is then defined as I(α) = S(α)(ρQ) −S(α)(ρRQ), where S(α)(ρ) = log(Tr ρα)/(1 −α) is the α−th Renyi entropy of the state and ρQ = TrR (ρRQ) is the reduced density matrix after tracing out the reference system.

In general α ≥0, with the limit α →1 recovering the von Neumann entropy and therefore the standard CI. The noisy map N(ρ) = P P P(E)EρE† is assumed to be Pauli noise such that E ∈P is an element of the n-qubit Pauli group and P(E) is the probability of each Pauli error operator E. Since ρ0 RQ is a pure stabilizer state, ρRQ is a mixed stabilizer state diagonal in the stabilizer basis of the respective error correcting code: ρRQ = X E P(E)Eρ0 RQE† = X s,l P(s)P(l’s)|s, Bl⟩⟨s, Bl |. These states |s, Bl⟩are the common eigenstate of the stabilizers of the QEC code and the products OX,Z Rj OX,Z Lj, each of them denoting a syndrome and logical operator respectively. For example, when k = 1 there are four Bell states corresponding to the logical operators I, OX L, OZ L, and OX L OZ L, such that the label l is in one-to-one correspondence with the logical Pauli operator. The product P(s)P(l’s) = P(s, l) ≡P E∈Es,l P(E), with Es,l denoting the set of errors that generate the same syndrome s and logical quantum number l. P(s) > 0 is the SPD and P(l|s) the conditional probability of each logical operator l given a syndrome s. Since ρRQ is already diagonal, the RCI can be rewritten as: I(α) = 1/(1 −α) log [P(l) P(l’s))α / P(s) P α(s) P(l) P α(l’s)]. The numerical reconstruction of the modified syndrome distribution Qα(s) for the d = 5 unrotated surface code, p = 0.2 and α = 1, 2 is shown; 20 stabilizers are considered, making a total of 220 possible syndromes, with one denoting the trivial syndrome.

The distribution becomes increasingly peaked around the more probable syndromes as α is increased. The argument of the logarithm contains the ratio of two quantities, (P(l) P(l’s))α = 1 and P(l) P α(l’s), averaged over the unnormalized distribution P α(s). This ratio quantifies, on average, how concentrated the conditional distribution P(l|s) is over logical outcomes. Consequently, the equation exhibits a singular behaviour when P(l|s) transitions from being sharply peaked at a single logical outcome to being distributed over multiple logical operators, as expected when crossing the QEC threshold.

As observed in Ref., the RCI displays increasing thresholds with increasing α, suggesting that logical fidelities sampled over the modified SPD Qα(s) = P α(s)/ P(s) P α(s) may show the same threshold behaviour as the RCI. For the toric code under bit-flip noise, the RCI displays a family of thresholds p(α) th for α = 1, 2, 3, 4, ., ∞, with α = 1 denoting the MLD threshold p th ∼0.109 given by the phase transition of the random bond Ising model along the Nishimori line. For α > 1 the RCI is described by different spin models, all the way to α →∞representing the PS threshold. Hence, obtaining the PS threshold α →∞in a QEC experiment implies discarding all non-trivial syndromes.

As suggested by the equation, observing the thresholds for 1 The logical error rates as a function of the bit-flip error rate in the vicinity of the respective threshold for (a) α = 1, (b) α = 2, and (c) α = 3 are shown. The logical error rates are shown for both MWPM and MLD decoding and for multiple code distances, with the insets showing the collapsed data for MLD, listing the extracted threshold pth and critical exponent ν values in Tab. I. The dashed gray lines mark the corresponding RCI threshold values listed in Tab. I. To compute the logical error rate p(α) L associated to Qα(s), syndromes are first sampled from P(s). Then, for each measured syndrome s, a binary random variable Xs is defined, where Xs = 1 if a logical error occurs after error correction and Xs = 0 otherwise, depending on the decoding strategy. The logical error rate is computed as: p(α) L = PN i=1 P α−1(si)Xsi / PN i=1 P α−1(si), where N is the number of samples.

For N →∞this yields the logical error rate sampled from Qα(s), requiring explicit evaluation of the probabilities P(s), which cannot generally be done efficiently. Below we discuss how P(s) can be approximated from syndrome data. However, for the unrotated surface code under bit-flip noise, P(s) can be computed using the efficient MLD. The numerical threshold values and critical exponents for α = 1, 2, 3 obtained from the finite-size scaling collapse are listed in Tab. I. The RCI values are determined from the statistical mechanical mapping of the RCI, with the values for MWPM reported in the SM.

Syndrome resampling achieves substantial reductions in quantum logical error rates

Logical error rates in quantum error correction have been reduced by up to four orders of magnitude using syndrome resampling, representing a substantial leap towards practical quantum computation. This advance overcomes a key limitation of current quantum devices, which struggle to maintain the necessary stability for complex calculations due to inherent noise. Syndrome resampling operates by intelligently re-weighting error diagnoses, known as syndromes, to focus on the most probable error scenarios, thereby improving the accuracy of error correction without altering the quantum hardware itself. Simulations utilising surface codes confirmed the effectiveness of syndrome resampling across both optimal and suboptimal decoding methods. Furthermore, applying this method to existing experimental data from superconducting qubits yielded up to two orders of magnitude reduction in logical error rates without requiring additional measurements or alterations to decoding algorithms.

Prioritising likely error scenarios via syndrome resampling for improved quantum error correction

Scientists are edging closer to reliable quantum computation through advances in error correction, but a practical hurdle remains: accurately diagnosing and mitigating errors without overwhelming the system with complexity. Syndrome resampling offers a clever software solution, re-weighting diagnostic signals to prioritise likely error scenarios. Determining the minimum number of samples needed to build a reliable picture of these probabilities proves challenging, scaling exponentially with system size and error rates in some scenarios. Syndrome resampling offers a practical software improvement applicable to existing quantum error correction experiments without requiring new hardware or complex code modifications.

Scientists reduced quantum logical error rates by up to four orders of magnitude using a technique called syndrome resampling. This method improves the accuracy of quantum error correction by prioritising the most probable error scenarios, effectively enhancing the reliability of quantum computations without needing improved hardware. The research demonstrates that syndrome resampling works with existing quantum error correction experiments and surface codes, and can reduce error rates using current data and decoding algorithms. Researchers found this approach yielded up to two orders of magnitude reduction in logical error rates from existing experimental data.