---
title: Cycle Error Reconstruction in Quantum Processors
url: https://www.emergentmind.com/topics/cycle-error-reconstruction-cer
type: topic
---

# Cycle Error Reconstruction in Quantum Processors

Searching arXiv for recent papers on Cycle Error Reconstruction and related work.
Cycle Error Reconstruction (CER) is a cycle-centric error-characterization protocol for quantum processors that leverages randomized compiling and Pauli twirling to estimate the error profile of a selected computing cycle, typically a fixed-schedule layer of native instructions or a layer of one- and two-qubit gates executed in parallel. In the CER literature, a “cycle” is the scheduled list of instructions over an arbitrarily large fraction of the chip, partitioned into “easy cycles” of single-qubit gates and a fixed “hard cycle” whose error is the target of reconstruction. Under randomized compiling, the noisy implementation of a dressed cycle is converted, on average, into an effective dressed cycle with a stochastic Pauli error model, enabling estimation of Pauli fidelities, Pauli eigenvalues, and low-weight marginals of the associated Pauli error distribution. In the extended folded variant, CER introduces repeated application of the hard cycle before twirling, which exposes the split between coherent and incoherent contributions to the error profile. The protocol is also referred to as K-body Noise Reconstruction (KNR) in work emphasizing local marginals and polynomial scaling under locality assumptions [2303.17714] [2303.09945].

## 1. Definition and scope

CER is designed to reconstruct the error distribution associated with a selected effective cycle rather than to characterize isolated gates in abstraction from scheduling context. This emphasis is motivated by the observation that error distributions are considerably influenced by the precise gate scheduling across the entire quantum processing unit, so the natural diagnostic object is a cycle rather than a standalone primitive. In the general formalism, a cycle is a fixed-schedule layer of native instructions applied to all qubits in the processor, and two consecutive layers of an arbitrarily chosen “easy” cycle and a fixed “hard” cycle define a dressed cycle [2303.17714].

Within this framework, CER targets either the full Pauli-twirled error model of the cycle or structured reductions of it. One formulation states the goal as estimating, for every Pauli supported on at most \(k\) of the parallel gates in the hard cycle, its marginal error probability
\[
\mu(P_A)=\sum_{Q\in\mathcal P_n:\,Q|_A=P_A} p(Q),
\]
thereby reconstructing all \(k\)-qubit reduced Pauli distributions. Another formulation emphasizes estimation of Pauli eigenvalues \(\lambda_P\), followed by inversion to Pauli error probabilities \(\chi_P\). A further extension, obtained by folding the hard cycle multiple times before twirling, aims not only at total error rates but at a detailed estimate of the coherent contribution to the error profile of a hard computing cycle [2303.17714] [2303.09945].

The scope of CER has expanded from physical-layer diagnostics on IBM-Q 5-qubit devices to coherent-versus-incoherent decomposition on IBM chips, to characterization of a 16-qubit transversal CNOT in trapped ions, and to decoder-informed learning of the dominant Pauli rates relevant to fault-tolerant quantum computation. This suggests that CER functions both as a benchmarking primitive and as a bridge between device-level noise characterization and logical-level performance analysis [2504.11980] [2507.08536].

## 2. Relation to cycle benchmarking and KNR

CER is structurally close to Cycle Benchmarking (CB). CB uses Pauli twirling and randomized compiling to turn an arbitrary cycle error into a stochastic Pauli channel, then extracts Pauli fidelities by fitting an exponential decay in the number of cycle repetitions. In this sense, CER inherits the CB circuit structure, the reliance on repeated-cycle survival signals, and the use of Pauli observables as the relevant diagnostic coordinates. However, CB reports total error rates but does not distinguish coherent versus incoherent parts, and in its standard form it is not organized around reconstructing marginal Pauli distributions [2303.09945].

CER departs from CB in two principal ways. First, it uses the same randomized-compiled, Pauli-twirled setting to estimate many Pauli-sector quantities sufficient to reconstruct marginals of the Pauli error distribution. Second, in the folded extension it introduces an extra hyperparameter \(x=k\), meaning that the hard cycle is folded back on itself \(k\) times before the final Pauli twirl. Varying this hyperparameter allows one to expose the quadratic buildup of coherent errors versus the linear buildup of incoherent errors, thereby generalizing the standard fidelity-decay model [2303.09945].

KNR is described as essentially equivalent to CER when the objective is to learn marginal Pauli error probabilities over small qubit subsets. The common idea is that locality constrains the number of significant parameters, so one can learn \(k\)-body marginals or top-weight terms without reconstructing the exponentially large full Pauli distribution. A common misconception is therefore that CER necessarily requires exhaustive estimation of all \(4^n\) Pauli rates. The literature instead distinguishes exact CER, which yields all \(4^n\) rates but costs \(O(n\,2^n)\) data, from scalable variants that exploit locality or sparsity to recover only low-weight marginals or the largest \(K\) Pauli error rates [2507.08536].

## 3. Experimental protocol and circuit structure

The basic CER protocol begins by separating the target schedule into easy and hard cycles. The hard cycle \(C_{\rm hard}\) is the layer of entangling gates whose error is to be characterized, while the easy cycle is a parallel single-qubit layer used for random Pauli dressings. Under randomized compiling, each noisy cycle \(E_{\rm hard}\circ C_{\rm hard}\) is twirled into an effective stochastic Pauli channel \(E_{\rm RC}\) followed by the ideal hard cycle. Repeating randomized compiling and averaging makes \(E_{\rm RC}\) a diagonal Pauli channel [2303.09945].

For each experimental setting one chooses a Pauli basis \(S\) to prepare and measure, a random randomized-compiling dressing string \(s\), and a number \(m\) of dressed-cycle repetitions. In the folded variant one also chooses the fold number \(x\). The full circuit is described as
\[
\text{Prepare } |\psi\rangle_{S}\rightarrow [\text{easy cycle}] \rightarrow [ (E_{\rm hard}\circ C_{\rm hard})^x + \text{easy cycle}]^m \rightarrow \text{invert SPAM} \rightarrow \text{measure in basis } S.
\]
By construction this implements, up to a final known Pauli, a stochastic Pauli channel so that each shot returns “\(\pm1\)” outcomes from which one estimates a Pauli fidelity \(f_P(m,x,s)\) [2303.09945].

In the more general CER formulation based on CB circuits, the experiment consists of selecting random Pauli dressings \(\{T_j\}\), inserting them before and after repeated applications of the cycle, and wrapping the repeated cycle between a state-preparation Pauli basis and a measurement Pauli basis. The trapped-ion realization makes this explicit through state preparation \(B\), Pauli randomization \(P_0\), repeated compiled Pauli cycles \(P_k\), and final basis rotation \(A\). For even \(m\), \(\mathcal H^m=\mathcal I\), so the ideal map is identity and one always measures in a Pauli basis [2303.17714] [2504.11980].

Data collection is organized over repeated choices of \(m\), random dressings, and, when coherent-versus-incoherent separation is desired, fold numbers \(x\in\{1,2,3,\ldots\}\). Averaging over many random dressing strings gives
\[
f_P(m,x)\simeq \mathbb E_s[f_P(m,x,s)],
\]
and the variation with \(x\) is used to reveal coherent buildup. In the top-\(K\) learning setting, the same Pauli-twirled cycle-benchmarking machinery is used to identify and learn the largest few Pauli error rates \(\{\chi_{P_i}\}_{i=1}^K\), with \(K\ll 4^n\) under plausible sparsity assumptions [2303.09945] [2507.08536].

## 4. Mathematical reconstruction

The central algebraic step in CER is the reduction of a general Markovian noise channel to a Pauli channel under twirling. In one formulation,
\[
\tau[E](\rho)=\frac1{4^n}\sum_{P\in P_n}P\,E(P\rho P)\,P=\sum_{P\in P_n}\chi_P(E)\,P\rho P,
\]
where \(\chi_P(E)\) is the Pauli error probability of \(P\). The experimentally accessible quantity is a Pauli-transfer-matrix eigenvalue \(\lambda_P\), extracted from the survival law
\[
S_P(m)=A_P\,\lambda_P^m.
\]
These eigenvalues are related to the Pauli probabilities by a Walsh–Hadamard transform,
\[
\lambda_Q=\sum_P H_{Q,P}\,\chi_P,\qquad
\chi_P=\frac1{4^n}\sum_{Q\in P_n}H_{P,Q}\,\lambda_Q,
\]
which provides the inversion from observed decay parameters to Pauli error rates [2507.08536].

The general CER formulation on small supports uses Pauli fidelities
\[
f(Q)=2^{-n}\mathrm{Tr}[Q\,S(Q)],
\]
for a Pauli channel \(S\), together with character orthogonality to obtain
\[
p(P)=\frac1{4^n}\sum_{Q\in\mathcal P_n}\chi_P(Q)\,f(Q),
\]
and, for marginals on a support \(A\),
\[
\mu(P_A)=\frac1{4^{|A|}}\sum_{Q_A\in\mathcal P_A}\chi_{Q_A}(P_A)\,f(Q_A).
\]
When the hard cycle normalizes the Pauli group on the chosen support, CER exploits conjugation orbits and reconstructs orbit-marginals from orbit-averaged fidelities rather than from each Pauli separately [2303.17714].

The folded extension refines this picture by introducing a small-error expansion for the fidelity of the folded cycle:
\[
f_P(x)\simeq 1-\delta f_P^{\rm decoh}\,x-\delta f_P^{\rm coh}\,x^2+O(x^3).
\]
Here \(\delta f_P^{\rm decoh}\) and \(\delta f_P^{\rm coh}\) are the marginal decoherent and coherent infidelities of the hard-cycle component that commutes with \(C_{\rm hard}\). Repeating the folded cycle \(m\) times yields
\[
f_P(m,x)=A_P[f_P(x)]^m
\simeq A_P\exp[-m(\delta f_P^{\rm decoh}x+\delta f_P^{\rm coh}x^2)+m\epsilon(x)],
\]
where \(A_P\) absorbs SPAM errors and \(\epsilon(x)\) is a small linear-in-\(x\) cross-term arising from residual interactions with the easy cycles. In practice one fits
\[
f_P(m,x)\simeq A_P\exp[-m(a_Px+b_Px^2)] + c_P m,
\]
with \(b_P\) revealing the coherent contribution and \(a_P\) the incoherent one [2303.09945].

In the trapped-ion transversal-CNOT study, the corresponding observable is an orbital eigenvalue, because Paulis decay through an orbit under the hard cycle. The measured decay of \(\langle P\rangle(m)\) yields \(\lambda_{P^{\circlearrowright \mathcal H}}\), and an inverse Walsh–Hadamard transform over the marginal subset produces the corresponding orbit-marginal probabilities. A constrained least-squares projection is then used to enforce nonnegativity and normalization when sampling noise would otherwise produce slight negatives [2504.11980].

## 5. Assumptions, scaling, and limitations

CER rests on a set of explicit approximations. The folded coherent-error analysis assumes Markovian, time-homogeneous errors generated by a time-independent Lindbladian; error locality, so that Hamiltonian and Lindblad operators are \(k\)-local on the device connectivity graph; small error rates, allowing truncation of Taylor and Baker–Campbell–Hausdorff expansions at second order; and a perfect or idealized Pauli twirl via randomized compiling, so that residual coherent cross-terms remain small. The general marginal-reconstruction formulation similarly assumes that after randomized compiling each effective dressed cycle has purely Pauli-stochastic error, that different errors are uncorrelated across cycles up to negligible second-order terms, and that a reduced-weight error model applies [2303.09945] [2303.17714].

Under these assumptions, the protocol is scalable. One resource estimate states that the number of Pauli marginals \(N\) grows at most polynomially in \(n\) under local noise, the total number of fit parameters is \(\sim 4N\), and the numbers of sequence lengths \(m\) and folds \(x\) are independent of \(n\), \(O(10)\). The number of random Pauli-compiled circuits per \((m,x)\) is \(O(1\ldots 30)\) to achieve concentration, and the number of shots per circuit only needs to deliver \(1\)–\(2\%\) precision, again independent of \(n\). In the low-weight-marginal formulation, the total number of runs scales as
\[
O\!\bigl(\tfrac1\epsilon\bigr)\times {s\choose k}\times O(\log\tfrac1\delta),
\]
which is polynomial in the number of supports \(s\) for fixed \(k\) [2303.09945] [2303.17714].

The literature also spells out the limitations. Strong time-correlations such as \(1/f\) noise can bias estimates under the Markovian-repeat assumption. Additive estimation of Pauli fidelities within each orbit is less robust to SPAM than CB-style orbital averages. Highly non-Pauli or coherent leakage errors require larger orbited groups or alternative twirling sets, such as \(D_4\) dihedral groups. In the decoder-oriented formulation, exact CER yields all \(4^n\) rates but costs \(O(n\,2^n)\) data; top-\(K\) learning reduces this burden only under sparsity assumptions. A plausible implication is that CER is most effective when the experimental regime already supports high-quality randomized compiling and when the dominant noise admits a compact Pauli description [2303.17714] [2507.08536].

## 6. Physical-layer and logical-layer demonstrations

Numerical and hardware demonstrations establish the range of phenomena CER can resolve. In the folded coherent-error study, numerical benchmarks simulated a single spectator qubit idling next to a train of CNOTs with \(T_1,T_2\) from ibmq\_montreal and an imposed coherent \(Z\)-rotation of rate \(0.002\) per CNOT. Using \(x\in\{1,3,5,7,9\}\) and \(m\in\{4,8,12,16,32\}\), the method recovered both incoherent rates \(\delta f^{\rm decoh}\) and coherent rates \(\delta f^{\rm coh}\) to within their statistical uncertainties. Proof-of-concept experiments on ibmq\_guadalupe, ibmq\_montreal, and ibmq\_manila targeted a spectator ancilla adjacent to a CNOT hard cycle and measured a pronounced coherent \(Z\)-biased error, \(\delta f_Z^{\rm coh}\approx 3\times 10^{-3}\), versus \(X,Y\) coherent rates \(\lesssim 10^{-4}\); up to \(\sim 80\)–\(90\%\) of the bias was attributed to coherent \(ZZ\)-coupling crosstalk [2303.09945].

The earlier CER implementation on IBM-Q 5-qubit devices applied the protocol to “burlington,” “essex,” “ourense” and “vigo,” reconstructing single-support and nearest-neighbor marginals across different topologies. The reported findings included low entangler error, with dominant two-qubit cycles having marginal Pauli error \(<1\%\); spectator crosstalk, where idle qubits adjacent to a busy two-qubit gate routinely showed weight-1 \(Z\) errors at the \(1\)–\(2\%\) level; and negligible higher-weight errors on disjoint pairs or triples. These data were then used to guide stochastic calibration, discussed below [2303.17714].

A later application characterized a 16-qubit transversal CNOT in a trapped-ion quantum computer. There, the hard cycle was
\[
\mathcal H=\prod_{i=0}^6 \mathrm{CNOT}_{i,i+9},
\]
executed serially due to the single motional bus but logically treated as one “parallel” cycle. CER learned all 1-CNOT marginals and 2-CNOT marginals across two disjoint pairs. Typical single-CNOT marginals on the nonadjacent pair \((1,14)\) were
\[
p(II)\approx 96.5\%,\quad p(ZI)\approx 1.2\%,\quad p(IZ)\approx 0.9\%,\quad p(XX)\approx 0.45\%.
\]
Adjacent pairs showed up to twice larger correlated \(XX\) and \(ZX\) orbits due to optical crosstalk. For the full transversal CNOT, the 1-CNOT marginals increased mainly in the dephasing orbits, with \(\mu(ZI)\approx 2.5\%\) and \(\mu(IZ)\approx 2.3\%\), while weight-2 orbits remained \(\le 1.0\%\). Using a nearest-neighbor Gibbs Random Field built from CER marginals, the reconstructed 14-qubit joint distribution yielded an uncorrectable logical-error probability
\[
P_{\rm uncorr}=0.088\pm 0.004\quad (2\sigma),
\]
compared to a total gate infidelity of \(0.335\pm 0.003\); a consistency check via standard cycle-benchmarking sampling of random eigenvalues gave \(0.362\pm 0.052\) [2504.11980].

These demonstrations collectively show that CER is not confined to average error-rate reporting. It can isolate spectator crosstalk, expose coherent \(Z\)-bias, identify short-range correlations, and separate correctable from uncorrectable contributions in a logical primitive. This suggests a role for CER as a context-sensitive diagnostic for both physical calibration and logical benchmarking [2303.09945] [2504.11980].

## 7. Calibration, mitigation, and decoder-informed use

CER has been used directly to drive calibration. In the IBM-Q 5-qubit study, the dominant spectator \(Z\) errors identified by CER defined a calibration target for stochastic calibration (SC), a fast compilation-based calibration method intended to identify and suppress local coherent error sources occurring in an effective cycle of interest. Three local virtual-\(Z\) phases on idle qubits were tuned by maximizing
\[
O(\theta_0,\theta_2,\theta_4)=f(X_0|\theta_0)+f(X_2|\theta_2)+f(X_4|\theta_4),
\]
with each curve fit to a quadratic to find the optimum phases. After SC, spectator \(Z\) marginals dropped by up to a factor of \(5\), and overall cycle infidelity improved by \(30\%\). The broader work reports up to a \(5\)-fold improvement of circuit performance [2303.17714].

CER has also been proposed as an input to mitigation and calibration strategies that distinguish coherent from incoherent error sources. Because the folded extension quantifies coherent and incoherent contributions separately, it can guide targeted calibration of coherent \(ZZ\)-coupling crosstalk while leaving incoherent components to other suppression tools. The source text explicitly notes potential applications ranging from benchmarking NISQ devices, calibrating pulse-level corrections for coherent crosstalk, to designing hybrid mitigation strategies that attack incoherent errors via zero-noise extrapolation and coherent errors via unitary compensation [2303.09945].

A separate line of work uses CER-derived Pauli rates for maximum-likelihood decoding. There, conventional decoders based only on the average infidelity \(\epsilon\) are contrasted with decoders informed by the top-\(K\) CER rates plus the Uncorrelated Split Search (USS) heuristic, which recursively fills in missing Pauli probabilities from a restricted dataset. For the level-2 concatenated Steane code (\(n=7\)), representative results in a CG1D ensemble show that \(K=163\), corresponding to \(1\%\) of the Pauli error rates, yields median gains of \(9.8\times\), \(9.3\times\), \(8.5\times\), and \(7.8\times\) across bins with \(\langle\epsilon\rangle=6.6\times 10^{-3}, 0.03, 0.08,\) and \(0.19\), respectively. The paper further states that with CER data constituting merely \(1\%\) of the Pauli error rates in the system, decoding achieves a \(10X\) gain in performance compared to the case where decoding is based solely on the fidelity of the underlying noise process, and that a typical violin plot shows most instances achieving \(15\)–\(30\times\) gain when \(K=163\). The corresponding total-variation-distance data show that USS reduces the model mismatch fivefold on average [2507.08536].

Taken together, these developments place CER at the intersection of characterization, calibration, and fault-tolerance analysis. It provides a cycle-level description of effective Pauli noise, can be extended to estimate coherent and incoherent components separately, and supplies structured noise information that can be propagated upward to logical primitives and decoders. A plausible implication is that CER is valuable precisely because it preserves the schedule dependence of noise while still admitting scalable reconstruction under locality or sparsity assumptions [2504.11980] [2507.08536].

Source: https://www.emergentmind.com/topics/cycle-error-reconstruction-cer