[Paper Review] The Geometry of Limit State Function Graphs and Subset Simulation

Karl Breitung|arXiv (Cornell University)|May 12, 2017

Probabilistic and Robust Engineering Design13 references3 citations

TL;DR

This paper investigates the geometric structure of limit state functions (LSFs) and reveals that subset simulation (SuS) can produce inaccurate results when underlying assumptions—such as unimodal failure domains or single dominant design points—are violated. The study demonstrates that SuS may fail to capture complex failure domain topologies, advocating for integration with FORM/SORM to detect structural features before simulation.

ABSTRACT

In the last fifteen the subset sampling method has often been used in reliability problems as a tool for calculating small probabilities. This method is extrapolating from an initial Monte Carlo estimate for the probability content of a failure domain found by a suitable higher level of the original limit state function. Then iteratively conditional probabilities are estimated for failures domains decreasing to the original failure domain. But there are assumptions not immediately obvious about the structure of the failure domains which must be fulfilled that the method works properly. Here examples are studied that show that at least in some cases if these premises are not fulfilled, inaccurate results may be obtained. For the further development of the subset sampling method it is certainly desirable to find approaches where it is possible to check that these implicit assumptions are not violated. Also it would be probably important to develop further improvements of the concept to get rid of these limitations.

Motivation & Objective

• To identify implicit geometric assumptions in subset simulation (SuS) that, when violated, lead to inaccurate failure probability estimates.
• To demonstrate through counterexamples that SuS may fail in cases with multiple or non-convex failure domains, even with large sample sizes.
• To argue that SuS, as a standalone method, lacks diagnostic capability to detect structural flaws in limit state functions.
• To propose that SuS should be complemented by FORM/SORM analysis to detect design points and curvature information before simulation.
• To advocate for a paradigm shift in structural reliability from pure probability computation to failure domain structure detection.

Proposed method

• Uses analytical and numerical examples to test SuS performance on limit state functions with complex geometries, including multiple design points and non-convex failure domains.
• Applies the standard subset simulation algorithm with iterative conditional probabilities: P(F) = ∏P(F_{k+1}|F_k), where F_k = {g(u) < a_k} and a_k → 0.
• Compares SuS results with FORM/SORM approximations to identify discrepancies arising from unmodeled geometric complexity.
• Employs Monte Carlo Markov Chain (MCMC) sampling to generate conditional samples in each subset level.
• Analyzes the complementary cumulative distribution function (CCDF) of the LSF g(u) to assess whether SuS captures all significant failure contributions.
• Proposes using FORM/SORM as a pre-processing step to detect multiple beta points and curvature, thus validating SuS applicability.

Experimental results

Research questions

• RQ1 Under what geometric conditions does subset simulation fail to accurately estimate small failure probabilities?
• RQ2 How do multiple or non-convex failure domains affect the convergence and accuracy of subset simulation?
• RQ3 To what extent can subset simulation be trusted when the underlying limit state function has complex topological features not captured by the method?
• RQ4 Can FORM/SORM analysis serve as a diagnostic tool to detect structural deficiencies in failure domains before applying subset simulation?
• RQ5 Is there a need to shift from probability-focused reliability analysis to structure-aware analysis of limit state surfaces?

Key findings

• Subset simulation can produce misleading results when the failure domain has multiple design points or non-convex geometry, even with large sample sizes.
• The method assumes a single dominant failure mode and unimodal structure, which may not hold in realistic high-dimensional problems.
• SuS lacks internal diagnostics to detect if all significant failure contributions have been captured, leading to undetected errors.
• When the limit state function is a black box, SuS cannot detect structural features like multiple beta points or curvature, unlike FORM/SORM.
• The paper concludes that SuS should not be used as a standalone method; instead, it should be preceded by FORM/SORM to verify the geometric structure of the failure domain.
• Without such validation, SuS risks being applied to problems where it is fundamentally unsuited, potentially requiring full Monte Carlo effort to achieve accuracy.