⚡ electrical engineering

Efficient Nonlinear Uncertainty Quantification for Spaceflight Leveraging Nonlinear Expansions

This paper presents a comparative study of modern uncertainty quantification methods using differential algebra and directional differential algebra to efficiently compute higher-order moments and improve confidence bounds for non-Gaussian distributions in nonlinear spaceflight problems, demonstrating significantly better accuracy than linear covariance approaches with only a fivefold increase in runtime.

Original authors: Ethan R. Burnett, Spencer Boone

Published 2026-05-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ethan R. Burnett, Spencer Boone

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict where a spaceship will end up after a long, bumpy journey through space. The problem is that space isn't empty; it's full of invisible tugs, gravitational pulls from moons, and atmospheric friction that can twist and stretch your path in unpredictable ways.

This paper is about a new, faster way to draw a "safety map" for that spaceship, especially when the path gets really weird and non-linear.

Here is the breakdown of their approach using simple analogies:

The Problem: The "Rubber Sheet" Effect

Usually, when scientists predict where a ship will go, they assume the uncertainty (the "maybe" part of the path) looks like a perfect, round balloon. If you push it, it stays round. This is called a Linear Covariance approach.

But in deep space, things don't stay round. The forces of gravity and atmosphere can stretch that balloon into a long, thin noodle, or even bend it into a banana shape. If you try to cover a banana with a round balloon, you either miss the ends of the banana (danger!) or cover a huge empty area (wasteful).

The Old Way: The "Slow Motion" Camera

To see the real shape of the banana, the old method was Monte Carlo. Imagine you have a camera that takes a picture of the ship's path. To understand the shape, you have to film the journey 10,000 times, each time starting with a tiny, random nudge. Then you look at all 10,000 photos to see the pattern.

Pros: Very accurate.
Cons: It takes forever. If you need to make a decision right now (like in real-time spaceflight), waiting for 10,000 simulations is too slow.

The New Solution: The "Magic Map" (Differential Algebra)

The authors propose a smarter way. Instead of filming the journey 10,000 times, they build a Magic Map (using a math tool called Differential Algebra or DA).

Think of this map as a detailed instruction manual that describes exactly how any tiny nudge will affect the ship's path. Once you write this manual (which takes a little time), you can instantly predict the outcome for thousands of different nudges without re-running the physics engine. It's like having a weather forecast that tells you exactly how a breeze will ripple a pond, rather than waiting for the wind to actually blow and watching the ripples.

The "Directional" Shortcut

Building the full Magic Map for every possible direction is still heavy. So, the authors introduced a trick called Directional Differential Algebra (DDA).

Imagine the spaceship is being stretched mostly in one direction (like pulling a piece of taffy). The authors say, "Let's only write the detailed instructions for that one pulling direction, and just make a quick guess for the side-to-side wiggles."

The Result: The map becomes much smaller and faster to use, with only a tiny loss in accuracy. It's like focusing your camera lens only on the most important part of the scene.

The "Banana" Shape

Once they have this fast map, they use it to calculate something called higher-order moments (which are fancy math words for "skewness" and "curvature").

Skewness: Is the banana bent to the left or right?
Kurtosis: Is the banana thick in the middle or thin?

By calculating these, they can draw a Banana-Shaped Safety Zone.

In their tests, the old "round balloon" method (Linear Covariance) missed the ship's actual location 11% of the time.
Their new "banana contour" method caught the ship 99.5% of the time, and it did it almost as fast as the old round method.

The Bottom Line

The paper shows that by using these "Magic Maps" and focusing on the most important directions, we can:

Speed things up: We can calculate complex safety zones in a fraction of a second, making them usable for real-time decisions.
Be more accurate: We can draw safety zones that actually look like the weird, bent shapes the ship's path takes, rather than forcing them into simple circles.

They tested this on two scenarios: a spaceship orbiting between the Earth and Moon, and a spaceship diving into Earth's atmosphere to slow down. In both cases, their new method was much faster than the old "take 10,000 photos" method and much more accurate than the "assume it's a circle" method.

Technical Summary: Efficient Nonlinear Uncertainty Quantification for Spaceflight Leveraging Nonlinear Expansions

Problem Statement
Reliable and efficient uncertainty quantification (UQ) remains a persistent challenge in modern astrodynamics, particularly within nonlinear and weakly stable regimes such as cislunar flight, close-proximity operations, and atmospheric entry/aerocapture. In these environments, initial Gaussian distributions often evolve into non-Gaussian shapes due to perturbations, nonlinear dynamics, and resonances. Standard methods face a trade-off: high-fidelity Monte Carlo (MC) methods are computationally prohibitive for rapid iteration or onboard deployment, while first-order linear covariance propagation is efficient but fails to capture nonlinear uncertainty transport, potentially misrepresenting the distribution's geometry. Intermediate methods like the Unscented Transform (UT) and Polynomial Chaos Expansion (PCE) offer a balance but still require repeated dynamics evaluations or complex bookkeeping for higher-order moments.

Methodology
This paper presents a comparative study of modern UQ methods, focusing on enhancing real-time performance through the use of Differential Algebra (DA) and a novel "Directional Differential Algebra" (DDA) approach.

Differential Algebra (DA) and Directional DA (DDA):
- DA: Instead of re-integrating full nonlinear dynamics for every sample, DA computes high-order Taylor expansions of the flow map. This allows for repeated UQ evaluations via algebraic operations on the expansion coefficients.
- DDA: To mitigate the combinatorial explosion of monomial terms in full DA expansions, the authors employ DDA. This technique selects independent variables aligned with physically important directions (specifically, the direction of maximum linearized stretching). It retains full nonlinear dependence along this preferential direction while applying only first-order corrections to transverse coordinates. This significantly reduces the number of retained terms and computational overhead.
Higher-Order Moment Recovery:
The study evaluates various methods (MC, LinCov, UT, Conjugate UT [CUT4], PCE, and Gaussian Mixture Models [GMM]) for their ability to recover not just mean and covariance, but also higher-order central moments (skewness and kurtosis). CUT4 and PCE are highlighted for their ability to extract these moments at a lower cost than large-sample MC.
Analytic "Banana" Approximation:
Leveraging the computed third- and fourth-order moments, the authors implement an analytic approximation for non-Gaussian confidence boundaries. Treating the non-Gaussian distribution as a perturbation of a Gaussian ellipse, they apply corrections based on skewness (long-axis asymmetry) and kurtosis (transverse bending). This results in a "banana-shaped" contour that captures moderate non-Gaussian deformation effects without requiring explicit reconstruction of the full probability density function.

Key Results
Numerical studies were conducted on two test cases: a restricted three-body cislunar halo orbit and an Earth-return aerocapture scenario.

Computational Efficiency:
- In the cislunar test, direct MC required ~12.8 seconds for $10^4$ samples.
- Pre-computed DA and DDA maps introduced a one-time construction cost (0.292s and 0.139s, respectively) but drastically reduced evaluation times. For instance, evaluating 10,000 samples via DDA+MC took only 0.061s.
- DDA was consistently faster than full DA, though it incurred a slight increase in covariance error due to the truncation of off-direction terms.
- The analytic banana contour method for the aerocapture problem required 0.227s, which is approximately 4.7 times the runtime of the linear covariance (LinCov) method (0.048s) but significantly faster than the 18.6s required for 2,000 MC samples.
Accuracy and Coverage:
- Cislunar Orbit: DDA+UT showed higher relative covariance error compared to full DA+UT, confirming the trade-off between speed and off-direction fidelity. However, it still outperformed LinCov.
- Aerocapture: The LinCov 3 $\sigma$ boundary captured only 88.6% of MC samples, failing to represent the sheared distribution. In contrast, the banana contour captured 99.55% of samples, providing a much more accurate spatial characterization of the non-Gaussian distribution with minimal computational overhead. Across seven aerocapture test cases, the banana method consistently outperformed LinCov and reduced over-conservatism compared to CUT4 covariance ellipses.

Significance and Claims
The paper claims that the proposed methodology offers a practical pathway for efficient, high-fidelity UQ in nonlinear spaceflight problems. By combining DDA for rapid flow map evaluation with an analytic banana approximation for confidence boundary reconstruction, the authors demonstrate that:

Real-time performance can be achieved for non-Gaussian UQ, with runtimes only marginally higher than linear covariance methods (e.g., ~5 $\times$ LinCov runtime) while capturing critical nonlinear deformations.
Higher-order moments (skew and kurtosis) are essential for accurately describing distributions in nonlinear regimes and can be computed quickly using methods like CUT4 or DDA.
The banana contour approximation provides a fully analytic confidence boundary that is significantly more accurate than linear covariance approaches for highly nonlinear problems, making it suitable for applications requiring chance-constrained guidance or rapid trade studies.

The authors suggest that these results support the integration of map-based UQ into stochastic extensions of fast localized guidance methods, enabling the supply of necessary confidence boundary information for chance-constrained variants without the prohibitive cost of repeated nonlinear propagation.