Error Budget Decomposition

• Error budget decomposition is the structured partition of a global error metric into subcomponents for allocation, attribution, and optimization.
• It employs various mathematical forms such as additive conservation, simplex allocation, and quadrature models to maintain composability with system-level targets.
• Its applications span quantum compilation, forecasting, and optical instrumentation, enabling informed design decisions and improved performance inference.

Error budget decomposition is the structured partition of a top-level performance, risk, or uncertainty quantity into constituent terms that can be propagated, allocated, optimized, or verified. In the cited literature, the budgeted object ranges from a document-level differential-privacy parameter, a forecast-error functional, and a fault-tolerant failure tolerance to two-qubit-gate infidelity, wavefront error, line-spread broadening, and astrometric baseline uncertainty. What unifies these otherwise disparate uses is the requirement that local terms remain composable with respect to a system-level target and interpretable enough to support design or inference decisions (Meisenbacher et al., 1 May 2026, Gourieroux et al., 2024, Sete et al., 2024, Ronggon et al., 17 Apr 2026, Donovan et al., 2021).

1. Functions and scope of decomposition

Error budget decomposition serves at least three distinct functions in the cited work. First, it is an allocation mechanism, in which a fixed global allowance is divided among subproblems before execution. Differentially private text obfuscation fixes a document-level privacy budget and distributes it across decomposed text chunks; fault-tolerant quantum-compilation papers distribute a total tolerated error across logical operations, T-state distillation, and rotation synthesis; and a Toffoli compiler selects per-gate decompositions to minimize a calibrated two-qubit-infidelity budget under contextual soundness constraints (Meisenbacher et al., 1 May 2026, Ronggon et al., 17 Apr 2026, Forster et al., 2 Sep 2025, Bartkiewicz et al., 30 Jun 2026).

Second, it is an attribution mechanism, in which an observed or predicted system error is decomposed into physically interpretable contributions. The parametric-resonance CZ-gate study separates total gate infidelity into incoherent and coherent channels; OGRE decomposes the line-spread function into optic, grating, alignment, detector, and jitter terms; MORFEO decomposes total residual wavefront error into high-order, low-order, focus, reference-loop, relay, calibration, telescope, and other categories; and the Roman Coronagraph Instrument expresses flux-ratio error in calibration, photometric, and differential-contrast branches (Sete et al., 2024, Donovan et al., 2021, Agapito et al., 5 Jun 2026, Nemati et al., 2023).

Third, it is an optimization or decision framework, where the decomposition itself becomes a design variable. The text-obfuscation study explicitly evaluates how text chunking and privacy-budget distribution alter empirical privacy-utility trade-offs under the same document-level $\varepsilon$; the fermionic $k$-RDM work chooses between unbiased direct estimation and biased cumulant reconstruction to minimize practical downstream error; and EBS maps disturbance-level wavefront terms into exposure-time consequences, allowing multivariate trades over stability, sensitivity, throughput, and detector performance (Meisenbacher et al., 1 May 2026, Takemori et al., 2023, Steiger et al., 9 Jan 2026).

2. Canonical mathematical forms

The literature does not use a single algebra for error budget decomposition. Instead, several recurrent forms appear, each tied to a different notion of what is being conserved or approximated.

Form	Representative rule	Representative use
Additive conservation	$\sum_i \varepsilon_i=\varepsilon$	Document-level DP budget split across chunks (Meisenbacher et al., 1 May 2026)
Simplex allocation	$$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$	FTQC error distribution across logical, T-state, and rotation budgets (Forster et al., 2 Sep 2025)
Additive surrogate cost	$B=\sum_g \mathcal I(d_g)$	Toffoli decomposition selection using calibrated two-qubit infidelities (Bartkiewicz et al., 30 Jun 2026)
Quadrature / RSS	$\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$	OGRE detector-plane blur and resolving-power budgeting (Donovan et al., 2021)
Variance summation	$$\sigma_{\mathrm{tot}}^2=\sigma_{HO}^2+\sigma_{LO}^2+\sigma_{Focus}^2+\cdots$$	MORFEO top-level WFE budget (Agapito et al., 5 Jun 2026)
Update decomposition	$$\gamma(h\mid I_t)=\sum_{k=0}^{h-1}\gamma(k,h\mid I_t)$$	FRED/FEKD/FELD in nonlinear forecasting (Gourieroux et al., 2024)
Empirical tradeoff proxy	$RG=\frac{U_p}{U_o}-\frac{P_p}{P_o}$	Privacy-utility comparison in DP text obfuscation (Meisenbacher et al., 1 May 2026)
Empirical accuracy ratio	$$r_k=\frac{|{}^kD-{}^k\hat D_{\rm cum}|}{|{}^kD-{}^k\hat D|}$$	Biased-versus-unbiased $k$0-RDM comparison (Takemori et al., 2023)

Additive rules dominate when the budget is a conserved allowance or a first-order proxy. In document-level DP text rewriting, normalized per-word scores are scaled by the target $k$1, then re-aggregated to chunk budgets so that the composed document-level accounting remains fixed. In Toffoli-network optimization, the exact multiplicative survival model $k$2 is acknowledged, but the selector uses the first-order additive surrogate $k$3, justified in the small-$k$4 regime as preserving decomposition rankings (Meisenbacher et al., 1 May 2026, Bartkiewicz et al., 30 Jun 2026).

Quadrature models dominate in wavefront and imaging systems, where independent or approximately independent blur and variance terms are expressed in common units and combined by RSS. OGRE explicitly assumes independent LSF contributions and combines them in quadrature; MORFEO writes the total WFE as a sum of squared category terms; the nlCWFS budget similarly assumes random, independent terms and adds spatial-domain and temporal-domain contributions in quadrature (Donovan et al., 2021, Agapito et al., 5 Jun 2026, Potier et al., 2023).

A third class replaces variance entirely. The forecast-error literature introduces additive decompositions of expected log-relative forecast error and density or Laplace-transform updates, thereby extending FEVD-style reasoning to nonlinear and non-Gaussian settings. This is not a small modification of FEVD, but a change in the object being decomposed: from conditional variance of a point forecast to update-by-update informational loss for positive transforms, predictive densities, or conditional Laplace transforms (Gourieroux et al., 2024).

3. Statistical, inferential, and forecasting uses

In forecasting, error budget decomposition is generalized beyond second moments by the family of measures introduced in “Forecast Relative Error Decomposition” (Gourieroux et al., 2024). The paper starts from the standard FEVD identity and then proposes FRED, FEKD, and FELD, which decompose total forecast error into nonnegative contributions from successive forecast updates. FEKD operates on predictive densities and can vary with the evaluation point $k$5, while FELD operates on conditional Laplace transforms and can accommodate multivariate combinations $k$6. The significance of this move is that the budget can now reflect serial and cross-sectional nonlinear dependence, tail behavior, and cases where FEVD is undefined, such as the Cauchy AR(1) discussed in the paper.

In approximate probabilistic structure learning, “Minimum Error Tree Decomposition” (1304.1103) uses local inconsistency itself as the budgeted quantity. The foundational local error term is $k$7, derived from the exact latent-tree condition $k$8. Pair-pair, pair-tree, tree-tree, and node-tree merge scores are built from that quantity, usually by taking maxima over relevant quadruples, and a greedy search repeatedly selects the currently minimum-error merge. Here the decomposition is neither exact nor globally optimized; it is a heuristic management of approximation error under noisy correlation measurements.

In astronomical dynamical inference, the DiskMass Survey treats decomposition as a full propagation of random and systematic errors from observables to derived physical quantities (Bershady et al., 2010). The central relation $k$9 propagates spectroscopy, SVE deprojection, scale-height calibration, gas corrections, photometry, and inclination into dynamical surface density, stellar mass-to-light ratio, and disk maximality. The paper concludes that for individual galaxies random and systematic errors are both at roughly the 25–30% level, while survey quartiles reduce to about 10%. It also draws a methodological distinction between observationally robust quantities and strongly model-coupled ones: disk maximality $\sum_i \varepsilon_i=\varepsilon$0 is treated as robust, whereas baryon fraction $\sum_i \varepsilon_i=\varepsilon$1 is said not to be a well-defined observational quantity because it remains coupled to the halo mass model.

4. Quantum information, compilation, and control

In fault-tolerant quantum compilation, decomposition is explicitly an allocation problem. “A Game Theoretic Approach for Optimizing Quantum Error Budget Distribution” (Ronggon et al., 17 Apr 2026) models logical error correction, T-state distillation, and rotation synthesis as three players sharing a normalized total error budget, with strategies $\sum_i \varepsilon_i=\varepsilon$2 constrained to the simplex and a shared cost $\sum_i \varepsilon_i=\varepsilon$3. An iterated best-response algorithm converges by monotonic descent of the common potential, and on 433 MQT benchmarks the paper reports a 30.22% average reduction in physical resource requirements relative to uniform allocation, with a peak improvement of 97.81%. “Improving Hardware Requirements for Fault-Tolerant Quantum Computing by Optimizing Error Budget Distributions” (Forster et al., 2 Sep 2025) uses the same three-way split, $\sum_i \varepsilon_i=\varepsilon$4, but learns circuit-specific allocations by supervised learning on resource-estimator outputs; it reports improvements for over 75% of circuits, with 15.6% average and 77.7% maximum reduction in estimated space-time cost.

“Context-Verified, Error-Budget-Aware Decomposition Selection for Toffoli Networks” (Bartkiewicz et al., 30 Jun 2026) uses a different notion of budget. Its optimization target is the additive hardware budget $\sum_i \varepsilon_i=\varepsilon$5, where $\sum_i \varepsilon_i=\varepsilon$6 sums calibrated two-qubit-gate infidelities inside a chosen Toffoli decomposition. That optimization is constrained by a separate soundness condition: relative-phase or approximate decompositions are admissible only if a context-conditioned equivalence or bounded-deviation check certifies them on the reachable subspace. The paper reports that blind pattern-matched relative-phase substitution is unsafe: 66 library rewrites are flagged as non-equivalent without context checks, count-greedy substitution corrupts 6 of 12 benchmark circuits, and the full verified pass yields 0 errors while still reducing two-qubit cost.

The fermionic $\sum_i \varepsilon_i=\varepsilon$7-RDM study treats decomposition as an empirical balance among statistical error, hardware/noise-induced error, cumulant-truncation bias, and QSE subspace-truncation error (Takemori et al., 2023). Direct $\sum_i \varepsilon_i=\varepsilon$8-RDM estimation is unbiased but has $\sum_i \varepsilon_i=\varepsilon$9 shot complexity; cumulant reconstruction from lower-order RDMs is biased but effectively $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$0. The comparison is operationalized through the accuracy ratio $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$1, and the main conclusion is that a biased estimator can outperform the unbiased one at realistic shot budgets and under hardware noise.

The parametric-resonance CZ-gate paper is a post hoc attribution budget (Sete et al., 2024). It decomposes measured two-qubit-gate infidelity into incoherent and coherent channels, with $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$2 and $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$3 obtained from under-gate $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$4, white-noise dephasing, and $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$5 dephasing formulas. For the representative 64 ns gate, the estimated total $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$6 agrees with the interleaved-RB value $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$7. The paper further states that incoherent channels account for approximately 83% of the total CZ error on average, with leakage the second-largest contributor.

5. Optical, wavefront, and astronomical instrumentation

In optical instrumentation, error budget decomposition is typically expressed in a common image-plane or phase unit and then rolled up to a science-facing metric. OGRE is exemplary: the line-spread-function budget translates optic PSF, grating resolution, grating/stack/module alignment, forward-assembly misalignments, detector placement, and in-flight jitter into detector-plane widths, combines independent contributions by RSS, and converts the resulting dispersion blur to resolving power through $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$8 (Donovan et al., 2021). The final RSS total is $$\varepsilon_L+\varepsilon_T+\varepsilon_R=\varepsilon_{\mathrm{tot}}$$9, corresponding to $B=\sum_g \mathcal I(d_g)$0, against a design requirement $B=\sum_g \mathcal I(d_g)$1 and goal $B=\sum_g \mathcal I(d_g)$2. The budget is validated with 1000 raytrace simulations: a worst-case simultaneous-$B=\sum_g \mathcal I(d_g)$3 study gives median $B=\sum_g \mathcal I(d_g)$4, while randomized misalignment simulations give median $B=\sum_g \mathcal I(d_g)$5. At the astrometric end, the GRAVITY baseline poster shows how baseline-related errors fit into a larger astrometric budget: because $B=\sum_g \mathcal I(d_g)$6, a 0.5 mm baseline error on a 100 m baseline corresponds to 5 ppm and therefore about $B=\sum_g \mathcal I(d_g)$7as for a $B=\sum_g \mathcal I(d_g)$8 pair separation (Lacour et al., 2014).

Adaptive-optics budgets use the same logic but with WFE as the common currency. The nlCWFS paper constructs a spatial-domain wavefront-error budget including algorithmic error, photon noise, finite bit depth, read noise, vibration, non-common-path tip-tilt, higher-frequency non-common-path aberrations, DM fitting error, and servo lag, under an approximate independence assumption (Potier et al., 2023). It shows that nlCWFS can dominate SHWFS on spatial-domain terms yet lose overall when reconstruction latency drives servo-lag error; in the $B=\sum_g \mathcal I(d_g)$9 optimistic case, the spatial quadrature sum is 34 nm for nlCWFS versus 66 nm for SHWFS, but total WFE becomes 97 nm versus 67 nm once servo lag is included. MORFEO adopts a larger hierarchical WFE budget, $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$0, with scenario totals of 245 nm, 318 nm, 449 nm, and 288 nm for its four requirement cases (Agapito et al., 5 Jun 2026). The paper stresses an organizational transition from early undifferentiated reserve terms to a bottom-up decomposed budget tied to E2E simulations, thermal analyses, optical tolerancing, calibration assumptions, and explicit contingency.

Roman CGI and EBS extend instrumentation budgeting from internal performance rollups to mission-facing detectability (Nemati et al., 2023, Steiger et al., 9 Jan 2026). Roman’s analytical model uses flux-ratio noise as the top-level metric and decomposes $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$1 into calibration and count-domain terms, $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$2, with three top-level branches: calibration errors, photometry noise, and differential flux ratio with post processing. The contrast-stability branch is especially distinctive: residual differential intensity is expressed through mean-field and variance-change terms of the coherent focal-plane field, and those are converted into FRN. EBS then generalizes this logic to future coronagraph missions by taking a spatio-temporal WFE array $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$3, a WFS&C transmission array $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$4, and a coronagraph sensitivity array $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$5, forming residual WFE $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$6, mapping it to contrast instability, and passing the resulting systematic speckle floor into EXOSIMS for exposure-time calculations. The significance is that wavefront-error decomposition is no longer only a subsystem exercise; it directly changes $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$7 and therefore yield-relevant observability.

6. Recurring principles, misconceptions, and open issues

A first recurring result is that uniform allocation is usually a weak baseline rather than a principled optimum. The DP text-obfuscation study explicitly states that “simple uniform budget distribution rarely produces superior results” and shows statistically significant effects of both decomposition method and distribution method on relative gain, with decomposition $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$8, distribution $\Delta x_{\mathrm{tot}}=\sqrt{\sum_i \Delta x_i^2}$9, and no strong interaction (Meisenbacher et al., 1 May 2026). The FTQC allocation papers make the same point in a different idiom: uniform splitting of a total tolerated error across logical, T-state, and rotation budgets is convenient, but it ignores heterogeneous cost sensitivities and therefore produces suboptimal physical-resource overhead (Ronggon et al., 17 Apr 2026, Forster et al., 2 Sep 2025).

A second recurring result is that equivalent top-level budgets do not imply equivalent empirical performance. Under the same document-level $$\sigma_{\mathrm{tot}}^2=\sigma_{HO}^2+\sigma_{LO}^2+\sigma_{Focus}^2+\cdots$$0, different chunking and token-importance heuristics yield materially different privacy, utility, and trade-off profiles in text obfuscation; under the same two-qubit-error model, different Toffoli decompositions are admissible or inadmissible depending on circuit context; under the same shot budget, a biased cumulant estimator can produce smaller total downstream error than an unbiased direct estimator; and under the same WFE total, PSF morphology and science utility can still differ because the spatial and temporal distribution of residuals matters (Meisenbacher et al., 1 May 2026, Bartkiewicz et al., 30 Jun 2026, Takemori et al., 2023, Steiger et al., 9 Jan 2026).

A third principle is methodological: most real error budgets are approximate, hybrid, or partially empirical rather than exact decompositions derived from a single theorem. The Toffoli selector uses an additive first-order surrogate although the exact per-decomposition survival model is multiplicative; OGRE and MORFEO rely on RSS under approximate independence assumptions; the parametric-resonance gate analysis assumes additive channel contributions only in the short-gate perturbative regime; the fermionic $$\sigma_{\mathrm{tot}}^2=\sigma_{HO}^2+\sigma_{LO}^2+\sigma_{Focus}^2+\cdots$$1-RDM paper does not provide a closed-form total-error theorem; and the Roman analytical model, while validated against an integrated model, still shows an annulus-dependent discrepancy with an overall RMS of 16% (Bartkiewicz et al., 30 Jun 2026, Donovan et al., 2021, Agapito et al., 5 Jun 2026, Sete et al., 2024, Takemori et al., 2023, Nemati et al., 2023).

A final implication is that current research increasingly treats decomposition itself as an optimization target rather than a bookkeeping afterthought. This suggests further movement toward end-to-end learned or certified allocation schemes, richer reachability models for context-dependent circuit rewrites, and tighter integration of subsystem instability models with mission-level exposure-time or yield codes (Meisenbacher et al., 1 May 2026, Bartkiewicz et al., 30 Jun 2026, Steiger et al., 9 Jan 2026).