PaperlayerView exact revision on arXiv

Source-linked AI summary

Optimal Uncertainty Quantification

Houman Owhadi, Clint Scovel, Timothy John Sullivan, Mike McKerns, Michael Ortiz

arXiv:1009.0679v3math.PRcs.ITmath.ITmath.STphysics.data-anstat.TH

TL;DR

The paper frames certification as bounding the probability that a system response exceeds a safety threshold. It develops reductions for uncertainty optimization; for one admissible set, extremizers have at most two support points, while the treatment of optimal bounds with sample data is deferred.

  • Problem

    Certification asks whether a system response will remain below a safety threshold with probability at least 1−ǫ.

    Source: §1.1 · p. 4

  • Method

    The paper reduces OUQ optimization to finite-dimensional problems, including a case with extremizers supported on at most two points.

    Source: §4 · p. 34, §5.1.3 · p. 48

  • Results

    For the admissible set A1, extremizers have push-forward measures supported on at most two points.

    Source: §4 · p. 34

  • Takeaways & Limitations

    Reduction theorems make planning experiment campaigns over admissible scenarios tractable.

    Source: §2.2 · p. 29

    The notion of an optimal bound with sample data is not immediately clear, and its full treatment is deferred to future work.

    Source: §3.3 · p. 34

Abstract

from arXiv · show

We propose a rigorous framework for Uncertainty Quantification (UQ) in which the UQ objectives and the assumptions/information set are brought to the forefront. This framework, which we call \emph{Optimal Uncertainty Quantification} (OUQ), is based on the observation that, given a set of assumptions and information about the problem, there exist optimal bounds on uncertainties: these are obtained as values of well-defined optimization problems corresponding to extremizing probabilities of failure, or of deviations, subject to the constraints imposed by the scenarios compatible with the assumptions and information. In particular, this framework does not implicitly impose inappropriate assumptions, nor does it repudiate relevant information. Although OUQ optimization problems are extremely large, we show that under general conditions they have finite-dimensional reductions. As an application, we develop \emph{Optimal Concentration Inequalities} (OCI) of Hoeffding and McDiarmid type. Surprisingly, these results show that uncertainties in input parameters, which propagate to output uncertainties in the classical sensitivity analysis paradigm, may fail to do so if the transfer functions (or probability distributions) are imperfectly known. We show how, for hierarchical structures, this phenomenon may lead to the non-propagation of uncertainties or information across scales. In addition, a general algorithmic framework is developed for OUQ and is tested on the Caltech surrogate model for hypervelocity impact and on the seismic safety assessment of truss structures, suggesting the feasibility of the framework for important complex systems. The introduction of this paper provides both an overview of the paper and a self-contained mini-tutorial about basic concepts and issues of UQ.

1 Introduction

The introduction presents Optimal Uncertainty Quantification (OUQ) as optimizing uncertainty bounds over scenarios consistent with available information and assumptions. It previews finite-dimensional reductions, optimal concentration inequalities, the computational framework and its hypervelocity-impact application, and cross-scale uncertainty questions.

1.1 The UQ problem

  • OUQ seeks sharp bounds by optimizing over admissible scenarios consistent with available information, rather than imposing assumptions that the information does not warrant.The admissible set A captures uncertainty about the unknown response function and probability measure; the resulting bounds are optimal for the assumptions and information encoded in A.Source: §1.1 · pp. 4–6

1.2 Motivating physical example and outline of the paper

  • The introduction motivates OUQ with a Caltech hypervelocity-impact surrogate and identifies seismic safety assessment of structures as a more complex application.The SPHIR example models perforation area for a steel ball impacting a steel plate; the paper considers both known and imperfectly known response functions.Source: §1.2 · pp. 6–7

1.2.1 Formulation of the admissible set and reduction theorems.

  • For linear generalized-moment constraints, OUQ problems can reduce to finite searches over products of finitely supported measures, with explicit support-size bounds.For the SPHIR example with known H, each input marginal needs at most two support points; with generalized-moment constraints on functions, the AMcD search reduces to functions and product measures on {0, 1}^3.Source: §1.2.1 · pp. 9–10

1.2.2 Optimal concentration inequalities.

  • Optimal McDiarmid bounds have analytic solutions for m = 1, 2, and 3 and a recursive formula for general m, while optimal Hoeffding bounds have analytic solutions.The optimal McDiarmid and Hoeffding bounds coincide for m = 2 but may differ for m = 3, so information about response-function linearity can have dimension-dependent effects.Source: §1.2.2 · p. 10
  • Under incomplete information, reducing uncertainty in one input may not improve the worst-case bound; in the two-variable case, the bound is independent of D2 when D2 + a ≤ D1.The stated SPHIR bound likewise does not change when obliquity uncertainty is reduced. In hierarchical systems, uncertainties or information may fail to propagate across scales.Source: §1.2.2 · p. 11
  • 43.7% is the least upper bound on SPHIR non-perforation probability under the stated imperfect-information set, compared with the non-optimal 66.4% bound.The bound reflects the information in AMcD: independence, input ranges and component-wise oscillations, and a mean perforation area between 5.5 and 7.5 mm2.Source: §1.2.2 · p. 11

1.2.3 Computational framework.

  • 37.9% is the numerical algorithm’s output, not a theorem; the reduced optimization is finite-dimensional, and simulations show the SPHIR maximizer’s marginals collapse to sparse supports.In the simulations, velocity and obliquity each collapse to a single Dirac mass, while thickness is supported at the extremes of its range; the algorithm has no proven convergence theorem.Source: Figure 1.3 · p. 12, §1.2.3 · p. 14
  • Wider searches still find two-point-support maximizers, suggesting reduced OUQ extremizers act as attractors and can reveal hidden reduction properties numerically.The proposed algorithm handles arbitrary constraints through coagulation and fragmentation of probability distributions.Source: §1.2.3 · p. 14
  • The simulations show singular extremizers whose support points identify system weak points; for U(AH), reducing thickness uncertainty is suggested to decrease the optimal bound on non-perforation.Reducing maximum obliquity and velocity range is not expected to decrease that bound.Source: §1.2.3 · p. 14
  • The proposed OUQ experimental-design loop predicts optimal experiment-result bounds under safe or unsafe assumptions, with best experiments minimizing overlap between predicted safe and unsafe ranges.The prediction problem is described as well-posed and as benefiting from reduction properties similar to certification.Source: §1.2.3 · pp. 14–15

1.2.4 Outline of the paper.

  • The paper outlines OUQ’s formal framework, finite-dimensional reductions, concentration inequalities, numerical and structural-safety applications, elliptic-PDE analysis, and a critique of common UQ practice.It notes that incomplete information can prevent uncertainty or information from propagating across scales, and presents OUQ as an opening gambit rather than a definitive solution.Source: §1.2.4 · p. 15

2 Optimal Uncertainty Quantification

This section presents Optimal Uncertainty Quantification (OUQ) as a way to bound failure probabilities over scenarios compatible with stated information and assumptions. It also describes how the OUQ loop uses available information to guide experiments that reduce worst-case uncertainty, including sequential minimax experiment campaigns.

2.1 First description

  • OUQ treats information and assumptions as defining admissible scenarios and optimizes over that feasible set rather than distorting or using only part of the available information.The assumptions may only approximate reality, but the set is intended to contain the real system.Source: §2.1 · p. 16
  • For a valid set A, L(A) and U(A) are optimal bounds: U(A) ≤ ε certifies safety, ε < L(A) de-certifies it, and intermediate bounds leave safety undecided.At the boundary L(A) = ε, the optimal decision can depend on whether a minimizing scenario exists; the paper describes a conservative “cannot decide” rule without that information.Source: §2.1 · pp. 17–18
  • In the simple example, certification can avoid evaluating an expensive response function, and the maximizers’ simple structure motivates later finite-dimensional reductions.The example’s optimum is achieved by a probability measure supported on two Dirac masses.Source: §2.1 · p. 19
  • Adding constraints shrinks the admissible set, lowers U(A), and raises L(A); knowing the median velocity is at the range midpoint reduced the non-perforation upper bound to 30.0%.The bounds are optimal relative to the information set, so a gap from a result based on exact knowledge reflects less-selective information rather than a loose bound.Source: §2.1 · pp. 20–21

2.2 The Optimal UQ loop

  • When bounds do not certify or de-certify, the OUQ loop uses physical laws, expert opinion, and data to form A, then seeks more information to sharpen bounds and tests their sensitivity to assumptions.New information imposes constraints that reduce A; sensitivity analysis can identify assumptions worth investigating.Source: §2.2 · pp. 22–23
  • Experiments are most useful when the possible functional-value intervals for safe and unsafe scenarios are nearly disjoint; substantial overlap makes a decisive certification less likely.The framework can also use multivalued experiment outcomes and accommodate measurement noise.Source: §2.2 · p. 24
  • For the proposed upper-half velocity-mass experiment, larger acceptable failure probabilities made it more discriminating, but its safe and unsafe intervals still overlapped, so it was not ideal.The example’s value of “best” depends on the safety threshold ε.Source: §2.2 · pp. 24–25
  • Selecting the next experiment is computational and uses prior information as constraints; it requires neither new experiments nor exact knowledge of the response function G.The optimization may evaluate many admissible functions compatible with the information about G, but does not require new evaluations of G.Source: §2.2 · p. 26
  • The most predictive experiment minimizes the worst-case decrease in the prediction interval, and E4 is identified as the most predictive experiment in the example.The criterion considers every possible outcome of an experiment, even though its precise outcome is unknown beforehand.Source: §2.2 · p. 27
  • Campaigns of experiments can be planned in advance as minmax games against the Universe, with reduction theorems making the search over admissible scenarios tractable.The framework is proposed as a potential alternative to scientific investigation based mainly on intuition, expert judgment, and guessing.Source: §2.2 · pp. 27, 29
  • In the sequential example, the first selected experiment measures thickness variance and the second measures thickness mean; an accurate optional model may reduce experiments, while an inaccurate one does not produce erroneous bounds.The candidate measurements concern the mean or variance of thickness, obliquity, or velocity.Source: §2.2 · p. 29

3 Generalizations and Comparisons

OUQ recasts prediction, verification, and validation as certification problems constrained by available information. Its comparisons emphasize that UQ methods rely on different assumptions, while incorporating random sample data raises delicate questions about what constitutes an optimal bound.

3.1 Prediction, extrapolation, verification and validation

  • OUQ formulates prediction, verification, and validation as certification problems by representing each task through information and assumptions that constrain admissible scenarios.Prediction seeks a minimum-width interval containing G(X) with probability at least 1 − ε; validation constrains the relation between a model F and response G, while verification addresses possible code bugs and discretization approximations.Source: §3.1 · pp. 29–30

3.2 Comparisons with other UQ methods

  • For rare-event certification, Monte Carlo can require prohibitive sample counts: ε = 10^-9 would require O(10^18) samples in the aviation example.The method uses independent samples and can incorporate additional information in a limited fashion through importance or weighted sampling.Source: §3.2 · pp. 30–31
  • Many UQ applications do not satisfy the regularity, known-distribution, sampling, or low-dimensionality conditions required by quasi-Monte Carlo and stochastic expansion methods.Quasi-Monte Carlo requires regularity and predetermined sampling, while stochastic expansions assume a sufficiently regular G and known distribution for X.Source: §3.2 · p. 31
  • Sensitivity analysis identifies critical inputs, but in its general form it does not target rigorous upper bounds on certification failure probabilities.Single-parameter oscillations can yield bounds through McDiarmid’s inequality, and partitioning along maximum-oscillation directions can sharpen them.Source: §3.2 · pp. 31–32
  • Bayesian inference can incorporate sampled data through Bayes’ rule, but the paper notes that priors may be arbitrary or unavailable and that small ε can demand many samples.It also warns that Bayesian convergence may fail in infinite-outcome settings.Source: §3.2 · p. 32
  • Deterministic worst-case methods optimize over response functions, whereas OUQ also optimizes over admissible probability measures; worst-case methods may be too pessimistic for practical use.The paper argues that weak assumptions on probability measures can produce reliable and practical failure-probability bounds.Source: §3.2 · pp. 32–33
  • OUQ comparisons depend on available information: methods may require assumptions that applications lack or discard information that is available.For the hypervelocity impact example, the compared methods would require arbitrary assumptions on probability densities or would discard information about input means or independence.Source: §3.2 · p. 33

3.3 OUQ with random sample data

  • With random sample data, defining an optimal bound on P[G(X) ≥ a] is delicate, and the paper defers full treatment to future work.The authors note connections to information theory and nonparametric statistics.Source: §3.3 · pp. 33–34

4 Reduction of OUQ Optimization Problems

The section reduces broad classes of OUQ problems from infinite-dimensional searches to finite-dimensional ones, using finite-support measures and, under additional conditions, functions on finite discrete spaces. It applies these reductions to McDiarmid-type bounds, including optimal solutions for small input dimensions and an asymptotic formula for general dimension.

4 Reduction of OUQ Optimization Problems

  • When admissible scenarios satisfy independence and linear inequality constraints on generalized moments, OUQ searches can be reduced to products of finite convex combinations of Dirac masses.The section also provides a priori upper bounds on the number of required masses.Source: §4 · p. 35

4.3 Application to McDiarmid’s inequality

  • The McDiarmid problem admits analytic solutions for m = 1, 2, and 3, plus an asymptotic formula for general m.For m = 2, when D2 is sufficiently smaller than D1, the optimal bound depends only on D1, so reducing D2 further does not improve the inequality.Source: §4 · p. 35

4.1 Reduction of OUQ

  • For optimization over product measures with n′ joint and n_i marginal linear moment constraints, optima can be achieved with each marginal supported on at most n′ + n_i + 1 points.The result applies on very general product spaces; the paper notes that indicator criteria satisfy its integrability condition.Source: §4.1 · p. 37

4.2 Generalized moments of the response function

  • When the criterion and all constraints are functions of f, the search further reduces to functions and product measures on finite discrete spaces, eliminating dependence on the Dirac masses’ coordinate positions.The reduction does not apply when constraints include functions that are not compositions with f.Source: §4.2 · p. 39

4.3 Application to McDiarmid’s inequality

  • For McDiarmid’s assumptions, the paper formulates the sharp upper bound as an OUQ optimization problem and reduces it to finite discrete searches.The resulting bound is optimal in the sense that admissible scenarios can approach it arbitrarily closely.Source: §4 · p. 35, §4.3 · p. 41
  • The reduction to a search over subsets of {0, 1}^m uses the lattice structure of the function space and provides an explicit algorithm for producing a reduced candidate.The paper contrasts this with more general constraints, for which finding such a candidate requires solving a separate problem.Source: §4.3 · p. 42

5 Optimal Concentration Inequalities

The section derives optimal concentration bounds under McDiarmid’s and Hoeffding’s assumptions, with explicit solutions in several dimensions and a recursive result for general dimension. These bounds show that reducing uncertainty in some inputs may leave the overall bound unchanged.

5.1 Explicit solutions under the assumptions of McDiarmid’s inequality

  • The authors describe their McDiarmid bounds as new and say prior optimized concentration bounds were suboptimal because they used the moment-generating-function technique.Source: §5.1 · p. 43

5.1.1 Explicit solutions in dimensions one and two

  • If a ≤ D1 − D2, decreasing D2 does not improve the two-dimensional OUQ bound; reducing D1 is then needed to reduce global uncertainty.Source: §5.1.1 · p. 44
  • When D1 ≥ D2 and D1 + D2 = D, the McDiarmid-bound ratio is largest at D2 = 0 and smallest at D1 = D2.The respective values are 1 − a/D and (1 − a/D)^2.Source: §5.1.1 · p. 44

5.1.2 Explicit solution in dimension three

  • For the SPHIR surrogate, the sub-diameters yield F2 = 0.253 and F1 = 0.437; reducing the obliquity sub-diameter does not change the 43.7% bound F1.Source: §5.1.2 · p. 47

5.1.3 Solution in dimension m

  • In arbitrary dimension, the McDiarmid bound has the stated tail formula, but no analytic solution is given for m ≥ 4 when a < the sum of the first m − 2 sub-diameters.Source: §5.1.3 · p. 47

5.2 Explicit solutions under the assumptions of Hoeffding’s inequality

  • The paper gives analytic solutions under Hoeffding’s assumptions and finds that the Hoeffding and McDiarmid optimal inequalities coincide for m = 2, while they may differ for m = 3.Source: §5.2 · pp. 48–49

6 Computational Implementation

The computational section tests whether reduced-support optimizers emerge in OUQ searches and outlines an algorithm and implementation for solving these optimization problems. In the hypervelocity-impact example, expanded searches retain small-support marginals, motivating a coagulation–fragmentation search strategy.

6.1 Extreme points of reduced OUQ problems are attractors

  • Even when the hypervelocity-impact search permits larger supports, the velocity and obliquity marginals collapse to single Dirac masses while thickness retains support at its range extremes.The wider search is more computationally intensive, and these observed patterns suggest that reduced-support extreme points act as attractors.Source: §6.1 · p. 49

6.2 Coagulation–Fragmentation algorithm for OUQ

  • The proposed coagulation–fragmentation procedure increases the allowed support when an optimizer uses all available points, and terminates when its support is smaller than the allowed maximum.That termination treats the observed collapse as evidence of a hidden reduction, not as a general guarantee.Source: §6.2 · p. 52
  • The procedure is a sketch that must be adapted to the application; with independence constraints, decisions should use marginal support cardinalities rather than the joint measure’s support.Source: §6.2 · p. 52

6.3 The OUQ algorithm in the mystic framework

  • The mystic implementation uses differential evolution with a modified Lagrange-multiplier method to satisfy constraints during optimization.The inner optimization aims to keep iterates admissible, avoiding the filtering of invalid evaluations.Source: §6.3 · p. 53
  • The framework parameterizes measure objects separately from the optimizer’s parameter lists, allowing the measure basis to be changed, including to Gaussian measures.Source: §6.3 · p. 54

7 Application to the Seismic Safety Assessment of Structures

The section applies OUQ to assess the safety of truss structures exposed to ground motion under limited information about potential earthquakes. It emphasizes the challenge of combining uncertain, high-dimensional inputs with complex, time-dependent structural responses and consequential safety decisions.

7 Application to the Seismic Safety Assessment of Structures

  • The application assesses truss safety under ground motion while accounting for imperfect knowledge of high-dimensional random inputs and a complex, time-dependent system response.Source: §7 · p. 54
  • The stated objective is to assess safety using the maximum earthquake magnitude and focal distance, with limited information and as few assumptions as possible.Source: §7 · p. 55

7.1 Formulation in the time domain

The section formulates seismic certification as a time-domain OUQ problem, representing ground motion through uncertain earthquake-source impulses and an uncertain transfer function. Reduction results convert the infinite-dimensional optimization into a finite-dimensional search, which is used to estimate truss failure probabilities across earthquake magnitudes.

7.1.1 Formulation of the problem

  • For certification, the truss is assumed to respond purely elastically, and member strains are obtained from the structure’s response to ground-motion acceleration.The response maps ground acceleration to member strains, which are checked against yield strains.Source: §7.1.1 · pp. 55–56

7.1.2 Formulation of the information set

  • Ground motion is modeled using independent, not necessarily identically distributed boxcar impulses with zero-mean bounded amplitudes, alongside a random transfer function of unknown distribution.Impulse durations have bounded support and constrained means; the source and transfer function capture uncertain inputs and propagation.Source: §7.1.2 · pp. 56–57

7.1.3 The OUQ optimization problem

  • The failure-probability bound is posed as an optimization over admissible response functions and probability measures consistent with the specified input constraints.Source: §7.1.3 · p. 58

7.1.4 Reduction of the optimization problem

  • Reduction replaces the unrestricted measures on impulse amplitudes and durations with two-point measures, while treating the unconstrained transfer coefficients c deterministically.The resulting problem has 8B + q = 180 dimensions.Source: §7.1.4 · p. 58

7.1.5 Numerical results

  • At ML = 6.5, the estimated failure probability is 0.631 at maximum ground acceleration 0.892 m · s−2, near the sharp transition in the vulnerability curve.The transition becomes smoother when the admissible-set description includes more independent variables.Source: §7.1.5 · p. 59
  • For the maximizing measure at amax = 0.892 m · s−2, the two weakest truss elements exceed the yield strain of 0.00119048.The two reported strains are 0.00142317 and 0.00125928.Source: §7.1.5 · p. 59
  • The extremal amplitude-mass positions and weights vary with earthquake magnitude, suggesting that large-amplitude impulses may maximize failure probabilities for strong earthquakes.The corresponding duration and transfer-coefficient masses show no discernible trend.Source: §7.1.5 · p. 61
  • For transition-region cases around ML = 6.5, convergence takes 30–50 iterations, while extreme ground-acceleration cases can require only one iteration.Source: §7.1.5 · p. 61

7.2 OUQ and critical excitation.

The section argues that seismic certification must balance worst-case response with realistic probabilistic assumptions. Because earthquake distributions are poorly determined by sparse records, OUQ may offer robust yet practical certification, provided the information set is reliable and narrow.

7.2 OUQ and critical excitation.

  • Sparse earthquake time histories make probability-distribution choices partly arbitrary, potentially undermining certification of critical structures and rare events.The authors suggest OUQ's weak assumptions could support robust, practical methods without the excessive pessimism of unconstrained critical excitation, but require a reliable, narrow information set.Source: §7.2 · pp. 62–63
  • Without constraints, worst-case excitation concentrates earthquake energy in structural resonance modes, while independent ground-motion variables would need to conspire in rare events to do so.When the transfer function and mean values are unknown, admissible scenarios can instead include strongly correlated ground motion that focuses energy in those modes.Source: §7.2 · p. 63

7.3 Alternative formulation in the frequency domain

The frequency-domain formulation uses weak information about seismic ground motion to optimize structural failure probabilities over all compatible distributions. Reduction and strong-duality methods make the resulting bounds computationally tractable, while the plotted vulnerability curve depends on the chosen information set.

7.3.1 Formulation of the information set

  • The admissible distributions constrain the mean power-spectrum shape using the Matsuda–Asano shape function and Livermore earthquake parameters.The formulation sets W = 100, τd = 20 s, ωg = 6.24 Hz, and ξg = 0.662; it bounds Fourier coefficients using amax from Esteva’s expression.Source: §7.3.1 · pp. 63–64

7.3.2 OUQ objectives

  • Failure is defined as any time-dependent axial strain leaving the safe domain, and the objective is to find its minimum and maximum probability over admissible measures.The admissible set consists of measures with the specified average power spectra; unlike filtered white noise, OUQ evaluates best- and worst-case scenarios across that set.Source: §7.3.2 · pp. 64–65

7.3.3 Reduction of the optimization problem with Dirac masses

  • With 6W moment constraints, the extrema can be attained using weighted sums of at most 6W + 1 Dirac masses.This reduction replaces optimization over arbitrary probability measures with a finite-dimensional representation.Source: §7.3.3 · p. 65

7.3.4 Reduction of the optimization problem based on strong duality

  • Strong duality converts the moment-constrained bounds into finite-dimensional problems, with the upper-bound problem convex and the lower-bound problem concave.The dual formulation does not identify extremal measures, but uses fewer optimization variables than the Dirac-mass reduction.Source: §7.3.4 · pp. 65–66, §7.3.5 · p. 66
  • Figure 7.5 plots optimal upper and lower failure-probability bounds versus Richter magnitude at R = 25 km; the maximum-probability curve differs from Sub-figure 7.1(b) because it uses a different information set.Failure is defined as displacement Y(t) not belonging to safe region S for all t in [0, τd].Source: §7.3.4 · p. 66, Figure 7.5 · p. 66

7.3.5 Numerical results

  • The strong-duality reduction is enabled by an identified response, no independence constraints, and information limited to 6W scalar moment constraints.Narrowing the gap between the failure-probability bounds requires richer information, such as constraints on correlations between Fourier-mode amplitudes.Source: §7.3.5 · p. 67
  • The nested Differential Evolution computation converges in about 15 hours on average, using roughly 2,000 outer iterations and 35,000–50,000 function evaluations.Each function evaluation entails an inner optimization over x on a high-performance computing cluster.Source: §7.3.5 · p. 67

8 Application to Transport in Porous Media

The porous-media application optimizes deviation probabilities for elliptic PDE solutions when distributions of random coefficients are imperfectly known. Its results show that a dominant uncertainty source can prevent other uncertainties from propagating into the output probability, including across scales.

8 Application to Transport in Porous Media

  • The application addresses imperfectly known coefficient distributions; its illustrative multiscale admissible set assumes κ2 is periodic with small period δ.Classical treatments often assume the coefficient distributions are known, whereas the paper notes that randomness alone may not fully characterize uncertainty when those distributions are unknown.Source: §8 · pp. 67–68
  • The analysis considers a stochastic elliptic PDE with random microstructure κ and positive source f, interpreting u as pressure or hydraulic head in a porous medium.The target is the least upper bound on the probability of unsafe supercritical pressure at an interior point.Source: §8 · pp. 67–68
  • When source uncertainty dominates by D2 ≥ a + D1, microstructure uncertainty does not propagate to the probability that log u(x0, ω) deviates from its mean.The result is expressed as equality between the optimal bound with both uncertainties and the bound with microstructure uncertainty set to zero.Source: §8 · p. 69
  • When background-coefficient uncertainty dominates by D1 ≥ a + D2, uncertainty in the periodic microstructure κ2 does not propagate to the deviation probability for log u(x0, ω).The theorem's bound is unchanged when the κ2 uncertainty is set to zero.Source: §8 · p. 69
  • The example suggests that incomplete information can prevent uncertainty or information from propagating across modules or scales, leaving global uncertainty unchanged below certain accuracy thresholds.The authors explain that scales or modules may not communicate some types of information; reducing global uncertainty may require decreasing local dominant uncertainties.Source: §8 · p. 69

9 Conclusions

The conclusion frames UQ as lacking a coherent general framework, with risks from both discarded information and excessive conservatism. OUQ foregrounds objectives and available information, while the authors describe it as a starting point and identify unresolved theoretical and computational questions.

9 Conclusions

  • UQ methods may impose assumptions that applications cannot support or discard relevant information, while inconsistent UQ objectives and communication leave the field's problem ill-posed.The conclusion contrasts the risks of discarding information with those of over-conservative safety certification.Source: §9 · p. 70
  • OUQ is presented as an opening gambit, with reduction theorems, optimal concentration inequalities, experiment selection, and numerical evidence offered as topics for further inquiry.The authors explicitly say OUQ is not a definitive answer to the UQ problem.Source: §9 · p. 70
  • Finding admissible states can yield rigorous lower bounds on U(A), while duality can yield upper bounds when the associated Lagrangians can be computed.The conclusion leaves open whether useful classes of problems permit rigorous estimation or bounding of those Lagrangians.Source: §9 · pp. 70–71
  • The reduction theorems are limited to linear constraints on probability-distribution marginals, and the algorithmic framework lacks general convergence theorems.The authors also identify sample data as making OUQ optimization more complex and raise the reducibility of resulting optimization problems as an open question.Source: §9 · p. 71

10 Appendix: Proofs

The appendix proves finite-dimensional reductions for Section 4 and establishes Section 5 results through optimization and symmetry arguments, including interior-optimum conditions and reductions of the admissible set. It also compares uncertainty bounds and function classes, including Hoeffding- and McDiarmid-type bounds in selected low-dimensional cases.

10.1 Proofs for Section 4

  • Theorem 4.1 reduces product-measure optimization to finite-support marginals by an extreme-point and Choquet argument, with successive replacements losing at most an arbitrary ε.The proof replaces each marginal in turn with a measure supported on finitely many points, then lets ε tend to zero.Source: §10.1 · pp. 72–73
  • Theorem 4.9’s proof clips h at a, then replaces it with h_C to preserve feasibility and the event probability.The lattice properties ensure the replacements remain admissible while leaving the criterion α[h ≥ a] unchanged.Source: §10.1 · p. 75

10.2 Proofs for Section 5

  • Proposition 5.7’s proof derives the optimizer through interior stationarity or boundary comparisons by induction; zero D_i values reduce the problem’s dimension.The stated results remain valid when one or more D_i equal zero.Source: §10.2 · pp. 76–77
  • Hypercube symmetry yields two optimizer properties: symmetric C0 permits α_i = 1, and reflections across central face-normal hyperplanes preserve optimizers.Source: §10.2 · p. 77
  • For m = 2, the bounds are equal for C = {(1, 1)}; for m = 3, U(AHfd) = U(AMcD) when F1 ≥ F2 and U(AHfd) < U(AMcD) when F1 < F2.The m = 2 equality follows from linearity; for m = 3, the comparison depends on the candidate configurations’ linearity or nonlinearity.Source: Figure 10.1 · p. 78, Figure 10.2 · p. 79
  • The F2 case is established when Ψ′(α) = 0 has an α ∈ (0, 1) with α3 ∈ (0, 1).The stationary-point equation is equivalent to equation (5.6), and the interior condition leads to the definition of F2.Source: §10.2 · p. 82
  • For C3, comparison with (5.3) gives U(AC3) ≤ U(AC1), by Proposition 5.7.Source: §10.2 · p. 83
  • The proof of Proposition 5.9 reduces the optimum of (5.8) to a set C containing one hypercube vertex, yielding Proposition 5.7’s explicit formula.For sets with multiple vertices, the proof uses infeasibility or symmetry-based reduction to a lower dimension.Source: §10.2 · p. 83
  • For Theorem 5.13, F1 ≤ U(A_Hfd) ≤ max(F1, F2); equality with U(A_McD) implies F1 ≥ F2.The proof uses a linear maximizer associated with F1 and shows that equality requires an optimum supported by a set C containing one hypercube vertex.Source: §10.2 · p. 83
Loading 1009.0679v3…