Every qubit in every quantum computer running today is fighting an enemy it cannot see — and in most cases, it has not even been properly introduced to that enemy. The dominant form of noise in real quantum hardware does not behave the way most noise suppression techniques assume: it remembers the past. Ask a superconducting qubit to perform two gate operations in sequence, and the noise corrupting the second operation may be shaped by what happened during the first — because the surrounding electromagnetic bath has not forgotten. Physicists call this non-Markovian noise, and until July 16, 2026, no experimentally validated method existed to suppress it exponentially without first characterizing it.

That changed when researchers from Southern University of Science and Technology (SUSTech) in Shenzhen and Tsinghua University in Beijing published results in Physical Review Letters demonstrating a protocol that achieves exactly this: exponential suppression of non-Markovian noise using only ancillary qubit resources, with no prior knowledge of the noise whatsoever. The paper — authored by Hongfeng Liu and Xinfang Nie of SUSTech, Zizhao Han and Zhenhuan Liu of Tsinghua University, and senior author Dawei Lu of SUSTech — validates its theoretical guarantees on a five-qubit nuclear magnetic resonance (NMR) processor, confirmed by full process tomography.

The result lands at a precise moment. The broader quantum computing industry spent the first half of 2026 accumulating milestones in quantum error correction: Microsoft and Quantinuum published peer-reviewed confirmation of an 800-fold reduction in error rates on trapped-ion hardware, and Google's Willow processor had already demonstrated below-threshold surface code operation in late 2024. But all of these results share a design assumption that real hardware is now empirically straining against — that the noise they are correcting behaves like a sequence of independent, memoryless events. Non-Markovian noise is the gap those milestones do not close.

What Non-Markovian Noise Is — and Why Standard Approaches Miss It

Quantum computing noise is almost universally modeled as Markovian: at each moment in time, the environment disturbs a qubit independently of what it did before. This assumption is mathematically convenient — it allows noise to be described as a sequence of independent quantum channels — and it underpins the theoretical foundations of quantum error correction, randomized benchmarking, and virtually every practical error mitigation technique developed since the 1990s.

The problem is that real quantum hardware violates this assumption routinely and significantly. In superconducting qubits, silicon spin qubits, and trapped ions, the surrounding physical environment — electromagnetic fluctuations, phonon baths, charge noise sources, nearby nuclear spins — retains correlations across time. A noise event during one gate operation can influence the noise experienced during a later operation, because the bath has not returned to its equilibrium state. Multiple research groups working on real superconducting hardware confirmed in 2020 and 2025 that this is not a theoretical concern: White et al. demonstrated non-Markovian process signatures on IBM hardware with reconstruction infidelities as low as 10⁻³, and a 2025 study at Singapore University of Technology and Design confirmed that specific temporal "streaky" correlations in surface code error syndromes can severely degrade logical error rate scaling in practice.

This creates a structural gap that standard tools cannot address. Full characterization of non-Markovian noise requires process tensor tomography — a generalization of standard quantum process tomography that tracks correlations across multiple time points. The number of parameters scales exponentially with both circuit depth and the degree of non-Markovianity. For any realistic device, it is simply intractable. Standard dynamical decoupling sequences must be tailored to specific noise spectral shapes that shift over time. Probabilistic error cancellation assumes an accurate noise model that is, for non-Markovian noise, prohibitively expensive to obtain. The new protocol sidesteps all of this.

How the Protocol Works: Two Stages, No Noise Model

The protocol combines two techniques — Pauli twirling and purification — in a sequence that, together, achieves something neither can accomplish alone.

Stage 1: Pauli twirling reduces complexity. Pauli twirling is a well-established technique in which random Pauli operators (the single-qubit bit-flip, phase-flip, and combined operators drawn from the set {I, X, Y, Z}) are inserted around gate operations and averaged over. When applied to a Markovian noise channel, Pauli twirling has the effect of eliminating off-diagonal error terms — mapping an arbitrary noise channel to a simpler Pauli channel whose errors have a diagonal structure. The insight in the new protocol is that this simplification also works on non-Markovian noise, and with a particularly useful consequence. After Pauli twirling, the original non-Markovian noise — which could in principle involve correlations across every gate operation in the circuit — is reduced to random Pauli errors occurring at two specific time points within the circuit, with a classical probability distribution over error patterns. The baroque complexity of environmentally correlated quantum noise has been reduced to a classical probability distribution — a far more tractable mathematical object, and one that the second stage can now act on.

Stage 2: Purification exponentially suppresses the errors. Quantum purification is a technique in which multiple copies of a noisy state or channel are combined to extract a higher-fidelity result. Its critical property is that it works for unknown states and processes: because the mechanism operates on the copies themselves rather than on a model of the noise, knowing what the noise looks like is not a prerequisite. The new protocol introduces an ancillary register of the same size as the main computation, along with a single control qubit. A series of controlled-SWAP (CSWAP) operations entangles the main and ancillary registers around the noisy operations. After the circuit completes, the control qubit is measured, and the result is used to post-process the main register output.

The mathematics reveals why this works. Post-selection on the control qubit measurement has the effect of squaring the probability weights of each Pauli error pattern in the classically correlated distribution produced by Stage 1 — and then renormalizing. Because the circuit is in the weak-noise regime (where the probability of no error exceeds the probability of any specific error), squaring the distribution amplifies the error-free component relative to all error terms. This squaring effect — and its iterative generalization to m ancillary copies, in which error terms are raised to the (m+1) power — is what produces exponential suppression. The protocol's only requirement is that the device is operating in the weak-noise regime, which any practical quantum device must be to be worth running at all. This is the definition of "weak noise" — the regime any practical quantum device must operate in to be useful.

Experimental Validation: Five Qubits, Engineered Memory Noise, Process Tomography

The team validated the protocol on a five-qubit NMR processor whose sample consists of two hydrogen-1 nuclear spins and three fluorine-19 nuclear spins in 1-bromo-2,4,5-trifluorobenzene dissolved in the liquid crystal solvent MBBA. NMR is the appropriate platform for this proof-of-principle work for a specific engineering reason: implementing CSWAP gates — the three-qubit operations the protocol requires — is substantially more efficient in liquid-state NMR than on digital superconducting processors, where a CSWAP decomposes into a long sequence of two-qubit gates with their own error contributions.

Non-Markovian noise was deliberately engineered into the experiment by coupling an environment qubit to the main register through a joint Hamiltonian at two separate time points. By tuning the evolution time t, the researchers could dial up the strength of memory noise and observe how the protocol responded. The experiment tested two scenarios: suppression during a standard unitary gate (a Hadamard gate on the main register), and suppression during a non-unitary channel (an engineered partial-SWAP operation that models dissipative dynamics). Most noise suppression techniques apply only to unitary operations; the ability to handle non-unitary channels matters for any quantum algorithm that involves measurement feedback, open-system simulation, or engineered dissipation.

In both cases, as engineered noise strength increased, the reference fidelity (without the protocol) dropped rapidly — while fidelity under the suppression protocol remained markedly higher. Full quantum process tomography — which reconstructs the complete quantum operation performed rather than just measuring output fidelity — confirmed close agreement between experimental results and theoretical predictions across all configurations. The reconstructed process matrices with suppression closely matched the ideal target channel, while those without the protocol showed measurable degradation.

Why This Matters for the Path to Fault Tolerance

The relationship between the new protocol and quantum error correction is specific and important. It is not that the protocol replaces error correction — it complements it in a way that addresses a documented real-device problem.

Standard fault-tolerant error correction architectures — the surface code most prominently — were designed under the assumption that noise is Markovian. When non-Markovian correlations are present, those assumptions break. Correlated errors can spread in ways that error correction's syndrome measurement cannot distinguish from single-qubit events, effectively lowering the fault-tolerance threshold below the physical error rate at which the hardware operates. White et al. demonstrated in 2020 that non-Markovian signatures are detectable on IBM superconducting hardware; Kam et al. confirmed in 2025 that specific temporal correlations can severely degrade surface code logical error rate scaling in practice. The new protocol directly addresses this gap: by suppressing the non-Markovian component of noise before it enters the error correction layer, it reduces the effective error rate that standard surface code schemes must handle — and does so without requiring knowledge of which specific non-Markovian mechanism is causing the problem.

This calibration-free character has a practical dimension that goes beyond the exponential scaling result. In real devices, noise drifts over time. Calibrations that characterize noise accurately in the morning may be stale by afternoon. A protocol that requires no calibration is inherently more robust to this drift: it suppresses non-Markovian noise correctly regardless of whether that noise is a bath of two-level fluctuators, a correlated phonon mode, or something not yet characterized — as long as noise remains weak enough to suppress.

The authors note that the current results are proof-of-principle. Extending the protocol to superconducting qubits or silicon spin qubits — where non-Markovian noise is well documented but CSWAP gates are expensive — is the key engineering challenge. The core theoretical result, however, is platform-agnostic: the exponential suppression guarantee holds for any system in which non-Markovian noise can be Pauli-twirled and the error-free probability dominates.

Publication Details

The paper "Realizing Universal Non-Markovian Noise Suppression" appears in Physical Review Letters, Vol. 137, 030601, published July 16, 2026. The authors are Hongfeng Liu and Xinfang Nie of SUSTech's Department of Physics and State Key Laboratory of Quantum Functional Materials, Zizhao Han of Tsinghua University's Center for Quantum Information (IIIS), and Zhenhuan Liu of Tsinghua's Yau Mathematical Sciences Center, with Dawei Lu of SUSTech as senior author. Funding sources include the National Natural Science Foundation of China, the Guangdong Provincial Quantum Science Strategic Initiative, and the Innovation Program for Quantum Science and Technology.

Frequently Asked Questions

What is non-Markovian noise, and why can't standard quantum error correction handle it?

Non-Markovian noise is quantum noise in which the surrounding environment retains memory of past interactions with the qubit — meaning errors during one gate operation can influence errors during a later operation. Standard quantum error correction codes, including the surface code, were designed assuming noise is Markovian (memoryless): each error event is independent of prior events. When non-Markovian correlations are present, that assumption breaks. Correlated errors can undermine the logical-error-rate scaling the surface code relies on, effectively shifting the fault-tolerance threshold in the wrong direction. White et al. (2020, Nature Communications) and Kam et al. (2025, Quantum Science and Technology) confirmed these effects on real superconducting quantum hardware.

How does the new protocol suppress non-Markovian noise without knowing anything about the noise?

It combines two techniques. First, Pauli twirling — inserting random Pauli operators around gate operations and averaging — converts the complex, correlated non-Markovian noise into a classically correlated distribution of Pauli errors at two time points. This does not require knowing the noise; it works on any noise that satisfies the Pauli twirling conditions. Second, a purification step using controlled-SWAP operations and a single control qubit effectively squares the probability weight of each error term in that distribution and renormalizes. Since the error-free component has the highest probability under weak-noise conditions, squaring amplifies it relative to all error terms — producing exponential suppression. The protocol's only requirement is that the device is operating in the weak-noise regime, which any practical quantum system must be to be useful.

Does this protocol replace quantum error correction, or work alongside it?

It works alongside error correction, not instead of it. The new protocol operates as a pre-layer: it suppresses the non-Markovian component of noise before the error correction layer has to deal with it, reducing the effective error rate that error correction schemes must handle. This is valuable because most error correction architectures assume Markovian noise — and when that assumption is violated by real hardware, their performance degrades in ways the architecture was not designed to handle. The new protocol does not require fault-tolerant infrastructure; it can also stand alone in near-term devices where full error correction is not yet available.

Can this protocol run on today's superconducting quantum computers?

Not in its current form without engineering adaptation. The experiment was performed on a nuclear magnetic resonance (NMR) processor, where the controlled-SWAP (CSWAP) gates the protocol requires can be implemented efficiently using the strong spin-spin couplings of the molecular system. On superconducting processors — where almost all large-scale quantum computers are currently built — a CSWAP decomposes into multiple two-qubit gates, adding its own noise contribution and potentially reducing the benefit. Adapting the protocol to superconducting or silicon spin qubit architectures is the next engineering challenge. The theoretical result is platform-agnostic: the exponential suppression guarantee holds on any platform where non-Markovian noise can be Pauli-twirled and the error-free probability dominates.