Efficient Nonlinear Uncertainty Quantification for Spaceflight Leveraging Nonlinear Expansions

Abstract: This paper provides a comparative study of modern uncertainty quantification (UQ) methods. To greatly enhance real-time performance, both differential algebra (DA) and a directional differential algebra (DDA) approach are employed. This can enable fast UQ in the case of non-Gaussian statistics. Higher-order moments, namely skew and kurtosis, can be computed quickly by several means. This motivates their implementation in an analytic approximation of the confidence bounds for the so-called "banana-shaped" non-Gaussian distributions encountered often in nonlinear astrodynamics problems. This method improves greatly on a linear covariance approach, with only 5x its runtime in numerical tests, even before DA methods are employed. Test problems in this work include a restricted three-body cislunar example and an Earth-return aerocapture example.

• The paper shows that precomputed differential algebra and directional differential algebra flow maps reduce repeated uncertainty-evaluation costs, with DDA-based Monte Carlo running about 200 times faster than direct Monte Carlo while retaining better accuracy than linear covariance propagation.
• The paper demonstrates that a four-moment banana confidence contour captures strongly non-Gaussian aerocapture dispersion more effectively than ellipses, covering 99.55% of samples versus 88.6% for the linear-covariance 3σ ellipse in the headline case.
• The paper establishes a practical speed–accuracy tradeoff: full DA preserves higher-order and mixed-moment fidelity, whereas DDA uses far fewer terms but loses off-direction accuracy, and the proposed contour method still lacks formal coverage guarantees for multimodal or severely distorted distributions.

Overview

This paper, by Burnett and Boone (2605.24147), presents a comparative study of uncertainty quantification (UQ) methods for nonlinear astrodynamics problems, with two principal contributions. First, it demonstrates that precomputed differential algebra (DA) flow maps—and a reduced variant termed directional differential algebra (DDA)—can dramatically lower the cost of repeated UQ evaluations for Monte Carlo (MC), sigma-point, polynomial chaos expansion (PCE), and Gaussian mixture model (GMM) methods. Second, it applies an analytic "banana" confidence boundary construction, driven by third- and fourth-order central moments computed via the conjugate unscented transform (CUT4), to non-Gaussian distributions arising in cislunar and aerocapture problems. The central quantitative claim is that the banana contour method improves substantially on linear covariance (LinCov) propagation at roughly 5× its runtime, even before DA acceleration is applied.

Methodological framework

The paper reviews the standard UQ hierarchy: MC sampling as the high-fidelity benchmark; LinCov based on first-order state transition matrix (STM) transport; the unscented transform (UT) with $2N+1$ sigma points; CUT4, which augments the sigma set with conjugate off-axis points ($1+2N+2^N$ points) to recover third- and fourth-order mixed central moments deterministically; non-intrusive PCE, from which mean and covariance follow algebraically from expansion coefficients; and GMM propagation with moment matching across mixture components.

The DA layer supplies a Taylor (jet transport) representation of the flow map $\bm{\varphi}_{t_f,t_0}$, so that repeated statistical evaluations reduce to polynomial evaluations rather than numerical integrations. The key combinatorial observation is that a full order-$j$ expansion in $N$ variables contains $K(N,j)=\binom{N+j}{N}-1$ monomials, which motivates the DDA formulation: the perturbation is decomposed as $$\delta\bm{X}=\hat{\bm{\gamma}}^*\chi+L\bm{\epsilon}$$ along a prioritized direction $\hat{\bm{\gamma}}^*$, retaining full nonlinear dependence in $\chi$ but only first-order terms transverse to it. This reduces the retained non-constant terms from combinatorial growth to $N+j-1$. The preferential direction is chosen as the dominant eigenvector of the Cauchy–Green tensor $1+2N+2^N$0, i.e., the direction of maximal linearized stretching—an approach analogous to directional state transition tensors.

The paper is explicit about the cost of this reduction: DDA discards all higher-order off-direction terms, degrading recovery of higher-order mixed moments and off-diagonal covariance accuracy relative to full DA, though it remains more accurate than LinCov.

Non-Gaussian confidence boundaries

The second contribution adapts the analytic banana contour of Burnett and Boone: treating the non-Gaussian confidence boundary as a perturbation of the Gaussian ellipse in whitened principal-axis coordinates. Two corrections are applied—a quadratic transverse bend with coefficient $1+2N+2^N$1 capturing skew-induced bending, and a Cornish–Fisher long-axis asymmetry correction $1+2N+2^N$2—using projected scalar moments $1+2N+2^N$3, $1+2N+2^N$4, and $1+2N+2^N$5 obtained by projecting the polynomial map onto the principal axes before forming any moment tensors. This projection trick avoids constructing unused entries of the full third- and fourth-order moment tensors, addressing the well-known monomial explosion when computing high-order moments directly from DA maps.

The authors concede plainly that the first four moments do not uniquely determine a non-Gaussian confidence boundary; the construction is a minimum-complexity approximation that has performed well empirically but carries no general guarantee.

Numerical results

The CR3BP trade study uses a JPL SSD southern $1+2N+2^N$6 halo orbit ($1+2N+2^N$7, period ≈ 13.62 days), initialized 0.25 TU past apolune and propagated 0.9 TU under anisotropic initial covariance aligned with the maximum-stretching direction. Representative runtimes on an Apple M4 Max are:

Method	Operation	Runtime
MC (direct)	$1+2N+2^N$8 samples	12.786 s
LinCov	STM/covariance	0.005 s
UT / CUT4	sigma points	0.017 / 0.097 s
PCE	total runtime	2.928 s
DA / DDA map construction	third-order maps	0.292 / 0.139 s
DDA+MC	$1+2N+2^N$9 map calls	0.061 s
DDA+UT	UT map calls	$\bm{\varphi}_{t_f,t_0}$0 s

Once the map is built, DDA+MC evaluates $\bm{\varphi}_{t_f,t_0}$1 samples roughly 200× faster than direct MC. Accuracy against the nonlinear UT reference shows the expected tradeoff: full DA+UT achieves mean/covariance errors of $\bm{\varphi}_{t_f,t_0}$2/$\bm{\varphi}_{t_f,t_0}$3, while DDA+UT degrades to $\bm{\varphi}_{t_f,t_0}$4/$\bm{\varphi}_{t_f,t_0}$5—still better than LinCov at $\bm{\varphi}_{t_f,t_0}$6/$\bm{\varphi}_{t_f,t_0}$7. The covariance error metric penalizes off-diagonal discrepancies most, confirming that directional truncation primarily harms mixed-moment fidelity. The paper also notes that CUT4 and PCE give similar higher-moment predictions here, with CUT4's advantage being determinism versus the sampling-based PCE used.

The aerocapture study considers Earth return at $\bm{\varphi}_{t_f,t_0}$8 km/s with exponential atmosphere ($\bm{\varphi}_{t_f,t_0}$9 km, $j$0 kg/m²). In the headline case, the LinCov $j$1 ellipse captures only 88.6% of 2000 MC samples while the banana contour captures 99.55%, against an expected two-dimensional $j$2 coverage near 98.9%—slightly conservative but geometrically far more representative. Across seven dispersion cases, the banana method outperforms LinCov in all cases and the CUT4 covariance ellipse in five of seven, with runtimes of 0.077–0.229 s versus 12–36 s for 4000-sample MC. Notably, these results were achieved without DA acceleration; the banana characterization took only ~4.7× the LinCov runtime.

Limitations and open questions

Several limitations are stated or implicit. The banana contour is a four-moment approximation valid for moderate non-Gaussian deformation; strongly multimodal or heavily sheared distributions fall outside its modeling assumptions, and no formal coverage guarantee is offered. DDA sacrifices off-direction higher-order fidelity, and the paper does not quantify when multiple retained directions become necessary—the mixing-term bookkeeping for that case is mentioned but not developed. The GMM comparison uses manually tuned weights without optimized splitting or risk allocation, so its runtime entry is not representative of a tuned implementation. Exploitation of sparsity and redundancy in high-order moment computations from DA maps is deferred to future work. Finally, the connection to chance-constrained guidance—where map-based UQ would supply confidence-boundary information for stochastic extensions of DA-based convex guidance—is proposed but not demonstrated.

Conclusion

The paper provides a systematic runtime-and-accuracy accounting of modern UQ methods for spaceflight, showing that precomputed DA/DDA flow maps make repeated nonlinear UQ evaluations orders of magnitude cheaper than repeated dynamics integration, and that a four-moment analytic banana contour recovers non-elliptical confidence geometry in aerocapture-class problems at near-LinCov cost. The main open question left by the work is whether these fast, moment-based confidence boundaries can be integrated reliably into onboard chance-constrained guidance loops, where coverage guarantees—not merely empirical coverage percentages—are required.