Conditional Density of the Environment (CDE)

• CDE is a probabilistic framework that defines environmental conditions leading to extreme structural responses by conditioning on return-level or failure events.
• It combines metocean models, short-term response simulations, and Bayes’ theorem to compute normalized and un-normalized densities for reliability analysis.
• CDE also acts as a diagnostic tool to assess and calibrate environmental contour methods for improved offshore structural design.

The Conditional Density of the Environment (CDE) is a metocean reliability construct that characterizes which environmental states are responsible for extreme structural response. In offshore-structure applications, the environment is represented by a vector $X\in\mathbb R^p$ of long-term variables for a sea-state, and the response is a structural maximum such as base shear over that sea-state. Two closely related formulations appear in the recent literature: a normalized density of environment conditional on a return-level response, $f_{X\mid R_L}(x\mid r_P)$, and an un-normalised density $$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$ associated with failure or threshold exceedance. In both cases, CDE identifies the regions of environment space that drive rare response events and supports fully probabilistic structural assessment, while also providing a diagnostic for the adequacy of environmental contour methods (Speers et al., 2024, Speers et al., 22 Sep 2025).

1. Mathematical object and conditioning conventions

In the return-level formulation, one observes a sea-state of length $L$, an environmental vector $X\in\mathbb R^p$, and a maximum structural response $R_L$ in that sea-state. If $r_P$ denotes the level corresponding to a return period $P$ for the annual maximum response $R_A$, with

$r_P=F_{R_A}^{-1}(1-1/P),$

then the CDE at return period $f_{X\mid R_L}(x\mid r_P)$0 is the joint density of the environment conditional on observing that $f_{X\mid R_L}(x\mid r_P)$1-year response,

$f_{X\mid R_L}(x\mid r_P)$2

In the paper’s worked metocean example, $f_{X\mid R_L}(x\mid r_P)$3 and $f_{X\mid R_L}(x\mid r_P)$4, where $f_{X\mid R_L}(x\mid r_P)$5 is storm-peak significant wave height and $f_{X\mid R_L}(x\mid r_P)$6 is wave steepness (Speers et al., 2024).

In the failure-threshold formulation, for a univariate structural response $f_{X\mid R_L}(x\mid r_P)$7 and critical threshold $f_{X\mid R_L}(x\mid r_P)$8, the CDE is defined as the un-normalised density

$f_{X\mid R_L}(x\mid r_P)$9

This quantity is proportional to the environmental density conditional on failure, and it is explicitly not normalized to integrate to one (Speers et al., 22 Sep 2025).

Formulation	Conditioning target	Expression
Return-level CDE	$$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$0	$$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$1
Failure-conditioned CDE	$$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$2	$$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$3

The two formulations are closely related. The return-level version is a normalized conditional density at an extreme response level; the threshold version is an un-normalised density over environment space whose integral gives failure probability. The literature also notes an exceedance-conditioned analogue,

$$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$4

and states that conditioning at the return level $$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$5 and conditioning on exceedance are closely related in the extreme regime (Speers et al., 2024).

2. Bayes representation and the link to failure probability

The normalized CDE is obtained by Bayes’ theorem: $$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$6 Here $$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$7 is the metocean model, $$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$8 is the short-term density of structural response given environment, and $$\tilde f_X(x;r_{\rm Cr})=P\{R>r_{\rm Cr}\mid X=x\}f_X(x)$$9 is the marginal response density (Speers et al., 2024).

The failure-threshold version is

$L$0

so it is the environment density weighted by conditional failure probability. Under this definition, the overall failure probability is

$L$1

The literature states this relation explicitly: the probability of structural failure is obtained by simply integrating the CDE over the environment space (Speers et al., 22 Sep 2025).

These expressions separate the problem into two factors: how frequently an environment occurs, and how likely that environment is to induce extreme response. That factorization is central to the role of CDE as both a descriptive object and a computational target. The normalized density identifies the environmental regimes most associated with a specified return-level response, while the un-normalised density directly supports failure-probability estimation.

3. Construction from metocean and short-term response models

A full CDE analysis in the metocean setting combines three inferential blocks: a joint environmental model $L$2, a short-term structural response model $L$3, and Bayes’ formula. In the metocean framework used to study contour methods, each environmental margin is fitted above a high threshold $L$4 with a Generalized Pareto distribution, transformed to a standard Laplace scale by the probability integral transform, and then modeled jointly using the Heffernan–Tawn conditional extremes model. For each coordinate $L$5, when $L$6,

$L$7

where $L$8 is a residual vector independent of $L$9. Parameters are estimated by likelihood under a working Gaussian assumption for $X\in\mathbb R^p$0, and the empirical kernel density of the residuals is then used for residual simulation (Speers et al., 2024).

Large samples from the tail of $X\in\mathbb R^p$1 are generated by simulating Laplace-scale vectors from the fitted conditional model, patching together regions where different coordinates exceed thresholds, and inverting the marginal transforms back to the original environmental variables. The environmental density is then approximated by bin-counting or a small-ball volume argument. For a small cell $X\in\mathbb R^p$2 containing $X\in\mathbb R^p$3, if $X\in\mathbb R^p$4 of the $X\in\mathbb R^p$5 draws fall in $X\in\mathbb R^p$6, then

$X\in\mathbb R^p$7

The short-term response model is built by decomposing the $X\in\mathbb R^p$8-hour maximum response into the maximum of $X\in\mathbb R^p$9 independent single-wave responses: $R_L$0 with

$R_L$1

The single-wave response distribution is obtained by integrating over crest height $R_L$2: $R_L$3 where

$R_L$4

In practice, the method samples conditioning crests from a biasing proposal $R_L$5, with $R_L$6, generates conditional waves, extracts single-wave maxima $R_L$7, and forms the importance-sampling estimator

$R_L$8

Then

$R_L$9

and its density is obtained by numerical differentiation around $r_P$0 (Speers et al., 2024).

This construction is explicitly described as a full “forward” probabilistic model. Its significance is methodological as well as numerical: the CDE is not postulated directly, but assembled from an extreme-value model for the environment and a physics-informed short-term response model.

4. Diagnostic use against environmental contour methods

Environmental contour methods, especially IFORM contours, attempt to identify design environments without explicit structural modelling. The CDE framework was introduced in part to test whether those contours actually characterize the environments responsible for extreme response. In the metocean study, the estimated $r_P$1 is plotted in the $r_P$2-plane and compared with collections of IFORM contours based on different hierarchical models for $r_P$3, including Lognormal, Weibull, GEV, and Gamma forms with linear or quadratic parameter trends (Speers et al., 2024).

The formal comparison uses an overlap metric

$r_P$4

The interpretation is stated explicitly: $r_P$5 indicates over-conservatism, meaning that the contour encloses parts of the environment that rarely produce the $r_P$6-year response, while $r_P$7 indicates non-conservatism, meaning that the contour misses the main contributors to extreme response (Speers et al., 2024).

For three example “stick” structures—A with uniform drag/inertia, B with deck-region load, and C with near-seabed load—the study reports that no single IFORM contour remains conservative for all three structures. Structure C shifts the CDE region to higher $r_P$8 but lower $r_P$9, so contours that are conservative for A and B become non-conservative for C. The paper’s abstract states the general conclusion more directly: none of the IFORM environmental contours considered characterise CDE well for three example structures (Speers et al., 2024).

The diagnostic implication is structural specificity. A contour derived only from the environment model need not align with the environmental states most relevant to a particular response mechanism. CDE makes that mismatch visible because it is conditioned on the actual response event of interest.

5. Estimation strategies for failure-conditioned environment densities

A subsequent methodological development studies two explicit procedures for estimating the failure-conditioned CDE and the associated failure probability: IS-PT, which combines parallel tempering MCMC with importance sampling, and AGE, which combines adaptive Gaussian emulation with Bayesian quadrature (Speers et al., 22 Sep 2025).

For IS-PT, the target is the unnormalised density

$P$0

The method runs $P$1 parallel Metropolis–Hastings chains targeting $P$2, with $P$3, using Gaussian within-chain proposals and periodic swaps between adjacent temperatures. The temperatures and proposal scales are tuned automatically to target a prespecified swap acceptance rate, following Vousden, Farr and Mandel (2016) as implemented in pyPESTO. After burn-in, samples from the $P$4 chain are smoothed with a Gaussian kernel density estimate

$P$5

with bandwidth $P$6 chosen via Scott’s rule. This yields an importance-sampling proposal, and failure probability is estimated by

$P$7

where $P$8. The paper states that this estimator is unbiased (Speers et al., 22 Sep 2025).

AGE instead models the log-CDE,

$P$9

as a Gaussian process with a Matérn kernel plus white noise. Given training data $R_A$0, the posterior mean and covariance are denoted $R_A$1 and $R_A$2. The failure probability is then approximated by Bayesian quadrature through

$R_A$3

To choose additional training points, AGE maximizes an acquisition function

$R_A$4

where $R_A$5 is an exploration term and $R_A$6 is an exploitation term. An alternative criterion replaces $R_A$7 with $R_A$8 (Speers et al., 22 Sep 2025).

The paper evaluates both approaches on a synthetic multimodal example and a monopile case study. For the synthetic case, the true probability is $R_A$9. IS-PT with $r_P=F_{R_A}^{-1}(1-1/P),$0, $r_P=F_{R_A}^{-1}(1-1/P),$1, and $r_P=F_{R_A}^{-1}(1-1/P),$2 gives total cost $r_P=F_{R_A}^{-1}(1-1/P),$3, RMSE $r_P=F_{R_A}^{-1}(1-1/P),$4, and bias $r_P=F_{R_A}^{-1}(1-1/P),$5. AGE with $r_P=F_{R_A}^{-1}(1-1/P),$6, $r_P=F_{R_A}^{-1}(1-1/P),$7, $r_P=F_{R_A}^{-1}(1-1/P),$8, and $r_P=F_{R_A}^{-1}(1-1/P),$9 gives cost $f_{X\mid R_L}(x\mid r_P)$00, RMSE $f_{X\mid R_L}(x\mid r_P)$01, and bias $f_{X\mid R_L}(x\mid r_P)$02. AGE with $f_{X\mid R_L}(x\mid r_P)$03, $f_{X\mid R_L}(x\mid r_P)$04, also at cost $f_{X\mid R_L}(x\mid r_P)$05, gives RMSE $f_{X\mid R_L}(x\mid r_P)$06 and bias $f_{X\mid R_L}(x\mid r_P)$07. For the monopile study, the true probability is $f_{X\mid R_L}(x\mid r_P)$08. IS-PT gives RMSE $f_{X\mid R_L}(x\mid r_P)$09 and bias $f_{X\mid R_L}(x\mid r_P)$10, whereas AGE with $f_{X\mid R_L}(x\mid r_P)$11, $f_{X\mid R_L}(x\mid r_P)$12, at cost $f_{X\mid R_L}(x\mid r_P)$13 gives RMSE $f_{X\mid R_L}(x\mid r_P)$14 and bias $f_{X\mid R_L}(x\mid r_P)$15, and AGE with $f_{X\mid R_L}(x\mid r_P)$16, $f_{X\mid R_L}(x\mid r_P)$17, at cost $f_{X\mid R_L}(x\mid r_P)$18 gives RMSE $f_{X\mid R_L}(x\mid r_P)$19 and bias $f_{X\mid R_L}(x\mid r_P)$20 (Speers et al., 22 Sep 2025).

These results separate reliability from efficiency. The paper states that IS-PT provides reliable results for both applications for lesser compute cost than naive integration, whereas AGE can provide a further reduction in computational cost when the exploration–exploitation weight $f_{X\mid R_L}(x\mid r_P)$21 is known, but is less reliable when $f_{X\mid R_L}(x\mid r_P)$22 is unknown (Speers et al., 22 Sep 2025).

6. Relation to broader conditional density estimation

Outside offshore reliability, conditional density estimation ordinarily means estimating the full density $f_{X\mid R_L}(x\mid r_P)$23 rather than a conditional mean. That distinction matters when the conditional distribution is multi-modal, skewed, or heteroscedastic, and when sharp predictive intervals are needed instead of intervals derived from a point predictor (Izbicki et al., 2017). The metocean CDE inherits the same probabilistic logic, but with the direction of conditioning chosen to isolate environment states associated with extreme response.

A related extreme-value antecedent appears in work on multivariate regularly varying vectors $f_{X\mid R_L}(x\mid r_P)$24, where the goal is to approximate

$f_{X\mid R_L}(x\mid r_P)$25

when $f_{X\mid R_L}(x\mid r_P)$26 is large. That framework uses the angular measure $f_{X\mid R_L}(x\mid r_P)$27 of the radial–angular decomposition to encode extremal dependence and produces a predictive distribution for an unobserved environmental component given large observed values (Cooley et al., 2013). This suggests that the metocean CDE can be viewed as a specialized tail-conditioned density problem in which the conditioned event is extreme structural response rather than a large coordinate observation.

The broader CDE literature also emphasizes that full densities support calibration diagnostics, interval forecasts, and uncertainty propagation. In the metocean setting, the analogous role is structural: the density over environment space identifies which combinations of wave variables materially contribute to rare response, and therefore which environmental regimes matter for design and failure analysis.

7. Advantages, limitations, and design implications

The metocean CDE literature lists three principal advantages. First, CDE arises naturally from the full “forward” probabilistic model, so no ad hoc assumptions about failure-boundary shape are needed. Second, it identifies precisely which combinations of environmental variables $f_{X\mid R_L}(x\mid r_P)$28 drive extreme responses for a given structural archetype. Third, it provides a diagnostic to calibrate or assess any approximate design-contour method (Speers et al., 2024).

The limitations are equally explicit. Computing CDE can require fitting a conditional-extremes model, simulating large tail samples, and performing importance-sampling wave–structure simulations, all of which incur non-trivial CPU time. Threshold choices, kernel-smoothing bandwidths, and the importance-sampling proposal parameter $f_{X\mid R_L}(x\mid r_P)$29 introduce practitioner judgment and statistical uncertainty. Moreover, the approach requires an explicit structural model, such as the Morison-load setting studied in the contour analysis, which contour methods were partly designed to avoid (Speers et al., 2024). In the sequential-design setting, AGE additionally requires balancing exploration and exploitation through a typically-unknown weight parameter $f_{X\mid R_L}(x\mid r_P)$30; when $f_{X\mid R_L}(x\mid r_P)$31 is unknown, IS-PT is reported to be more reliable, especially for multimodal CDEs (Speers et al., 22 Sep 2025).

The design recommendations follow directly from those findings. Where feasible, the literature recommends adopting the full probabilistic forward approach and computing CDE as a by-product of $f_{X\mid R_L}(x\mid r_P)$32-year response estimation. If contour methods must be used for computational or regulatory reasons, the recommendation is to calibrate the contour to the specific structural archetype so that $f_{X\mid R_L}(x\mid r_P)$33. A further recommendation is to incorporate uncertainty in environmental model form into contour estimation, or to move to asymptotically justified tail methods such as the conditional-extremes copula (Speers et al., 2024).

In that sense, CDE is not merely another reliability summary. It is a structurally conditioned density over environment space that exposes which parts of the metocean regime are relevant to extreme response, quantifies failure probability through integration in its un-normalised form, and provides a principled basis for evaluating whether environment-only approximations are capturing the events that matter for design.