1Blurb:: Uncertainty quantification using polynomial chaos expansions 2 3Description:: 4The polynomial chaos expansion (PCE) is a general framework for 5the approximate representation of random response functions in terms 6of finite-dimensional series expansions in standardized random variables 7 8\f[R = \sum_{i=0}^P \alpha_i \Psi_i(\xi) \f] 9 10where \f$\alpha_i\f$ is a deterministic coefficient, \f$\Psi_i\f$ is a 11multidimensional orthogonal polynomial and \f$\xi\f$ is a vector of 12standardized random variables. An important distinguishing feature of 13the methodology is that the functional relationship between random 14inputs and outputs is captured, not merely the output statistics as in 15the case of many nondeterministic methodologies. 16 17<!-- 18Groups 1 and 2, plus the optional keywords \c p_refinement and 19\c fixed_seed relate 20to the specification of a PCE method. In addition, this method 21treats variables that are not aleatoric-uncertain different, 22despite the \ref variables-active keyword. 23 24Group 3, and the remainder of the optional keywords relate 25to the output of the method. 26--> 27 28<b> Basis polynomial family (Group 1) </b> 29 30Group 1 keywords are used to select the type of basis, 31\f$\Psi_i\f$, of the expansion. Three approaches may be employed: 32 33\li Wiener: employs standard normal random variables in a transformed 34 probability space, corresponding to Hermite orthogonal basis 35 polynomials (see \ref method-polynomial_chaos-wiener). 36 37\li Askey: employs standard normal, standard uniform, standard 38 exponential, standard beta, and standard gamma random variables in a 39 transformed probability space, corresponding to Hermite, Legendre, 40 Laguerre, Jacobi, and generalized Laguerre orthogonal basis 41 polynomials, respectively (see \ref method-polynomial_chaos-askey). 42 43\li Extended (default if no option is selected): The Extended option 44 avoids the use of any nonlinear variable transformations by 45 augmenting the Askey approach with numerically-generated orthogonal 46 polynomials for non-Askey probability density functions. Extended 47 polynomial selections replace each of the sub-optimal Askey basis 48 selections <!-- with numerically-generated polynomials that are 49 orthogonal to the prescribed probability density functions --> 50 for bounded normal, lognormal, bounded lognormal, loguniform, 51 triangular, gumbel, frechet, weibull, and bin-based histogram. 52 53For supporting correlated random variables, certain fallbacks must 54be implemented. <!-- The selection of Wiener versus Askey versus 55Extended is partially automated and partially under the user's control. --> 56- The Extended option is the default and supports only 57 Gaussian correlations. <!--- This default can 58 be overridden by the user by supplying the keyword \c askey to request 59 restriction to the use of Askey bases only or by supplying the keyword 60 \c wiener to request restriction to the use of exclusively Hermite 61 bases. --> 62- If needed to support prescribed correlations (not under user 63 control), the Extended and Askey options will fall back to the Wiener 64 option <EM>on a per variable basis</EM>. If the prescribed 65 correlations are also unsupported by Wiener expansions, then %Dakota 66 will exit with an error. 67 68Refer to \ref topic-variable_support for additional information on 69supported variable types, with and without correlation. 70 71<b> Coefficient estimation approach (Group 2) </b> 72 73To obtain the coefficients \f$\alpha_i\f$ of the expansion, seven 74options are provided: 75 76<ol> 77<li> multidimensional integration by a tensor-product of Gaussian 78 quadrature rules (specified with \c quadrature_order, and, 79 optionally, \c dimension_preference). 80<li> multidimensional integration by the Smolyak sparse grid method 81 (specified with \c sparse_grid_level and, optionally, 82 \c dimension_preference) 83<li> multidimensional integration by Stroud cubature rules 84 and extensions as specified with \c cubature_integrand. 85<li> multidimensional integration by Latin hypercube sampling 86 (specified with \c expansion_order and \c expansion_samples). 87<li> linear regression (specified with \c expansion_order and 88 either \c collocation_points or \c collocation_ratio), using 89 either over-determined (least squares) or under-determined 90 (compressed sensing) approaches. 91<li> orthogonal least interpolation (specified with 92 \c orthogonal_least_interpolation and \c collocation_points) 93<li> coefficient import from a file (specified with \c 94 import_expansion_file). The expansion can be comprised of a 95 general set of expansion terms, as indicated by the multi-index 96 annotation within the file. 97</ol> 98 99It is important to note that, for polynomial chaos using a single 100model fidelity, \c quadrature_order, \c sparse_grid_level, and \c 101expansion_order are scalar inputs used for a single expansion 102estimation. These scalars can be augmented with a \c 103dimension_preference to support anisotropy across the random dimension 104set. This differs from the use of sequence arrays in advanced use 105cases such as multilevel and multifidelity polynomial chaos, where 106multiple grid resolutions can be employed across a model hierarchy. 107 108<b> Active Variables </b> 109 110The default behavior is to form expansions over aleatory 111uncertain continuous variables. To form expansions 112over a broader set of variables, one needs to specify 113\c active followed by \c state, \c epistemic, \c design, or \c all 114in the variables specification block. 115 116For continuous design, continuous state, and continuous 117epistemic uncertain variables included in the expansion, 118Legendre chaos bases are used to model the bounded intervals for these 119variables. However, these variables are not assumed to have any 120particular probability distribution, only that they are independent 121variables. Moreover, when probability integrals are evaluated, only 122the aleatory random variable domain is integrated, leaving behind a 123polynomial relationship between the statistics and the remaining 124design/state/epistemic variables. 125 126<b> Covariance type (Group 3) </b> 127 128These two keywords are used to specify how this method computes, stores, 129and outputs the covariance of the responses. In particular, the diagonal 130covariance option is provided for reducing post-processing overhead and 131output volume in high dimensional applications. 132 133<b> Optional Keywords regarding method outputs </b> 134 135Each of these sampling specifications refer to sampling on the PCE 136approximation for the purposes of generating approximate statistics. 137\li \c sample_type 138\li \c samples 139\li \c seed 140\li \c fixed_seed 141\li \c rng 142\li \c probability_refinement 143\li \c distribution 144\li \c reliability_levels 145\li \c response_levels 146\li \c probability_levels 147\li \c gen_reliability_levels 148 149which should be distinguished from simulation sampling for generating 150the PCE coefficients as described in options 4, 5, and 6 above 151(although these options will share the \c sample_type, \c seed, and \c 152rng settings, if provided). 153 154When using the \c probability_refinement control, the number of 155refinement samples is not under the user's control (these evaluations 156are approximation-based, so management of this expense is less 157critical). This option allows for refinement of probability and 158generalized reliability results using importance sampling. 159 160 161<b> Usage Tips </b> 162 163If \e n is small (e.g., two or three), then tensor-product Gaussian 164quadrature is quite effective and can be the preferred choice. For 165moderate to large \e n (e.g., five or more), tensor-product quadrature 166quickly becomes too expensive and the sparse grid and regression 167approaches are preferred. <!-- For large \e n (e.g., more than ten), 168point collocation may begin to suffer from ill-conditioning and sparse 169grids are generally recommended. --> Random sampling for coefficient 170estimation is generally not recommended due to its slow convergence 171rate. <!--, although it does hold the advantage that the simulation 172budget is more flexible than that required by the other approaches.--> 173For incremental studies, approaches 4 and 5 support reuse of previous 174samples through the \ref method-sampling-sample_type-incremental_lhs 175and \ref model-surrogate-global-reuse_points 176specifications, respectively. 177 178In the quadrature and sparse grid cases, growth rates for nested and 179non-nested rules can be synchronized for consistency. For a 180non-nested Gauss rule used within a sparse grid, linear 181one-dimensional growth rules of \f$m=2l+1\f$ are used to enforce odd 182quadrature orders, where \e l is the grid level and \e m is the number 183of points in the rule. The precision of this Gauss rule is then 184\f$i=2m-1=4l+1\f$. For nested rules, order growth with level is 185typically exponential; however, the default behavior is to restrict 186the number of points to be the lowest order rule that is available 187that meets the one-dimensional precision requirement implied by either 188a level \e l for a sparse grid (\f$i=4l+1\f$) or an order \e m for a 189tensor grid (\f$i=2m-1\f$). This behavior is known as "restricted 190growth" or "delayed sequences." To override this default behavior in 191the case of sparse grids, the \c unrestricted keyword can be used; it 192cannot be overridden for tensor grids using nested rules since it also 193provides a mapping to the available nested rule quadrature orders. An 194exception to the default usage of restricted growth is the \c 195dimension_adaptive \c p_refinement \c generalized sparse grid case 196described previously, since the ability to evolve the index sets of a 197sparse grid in an unstructured manner eliminates the motivation for 198restricting the exponential growth of nested rules. 199 200<b> Additional Resources </b> 201 202%Dakota provides access to PCE methods through the NonDPolynomialChaos 203class. Refer to the Uncertainty Quantification Capabilities chapter of 204the Users Manual \cite UsersMan and the Stochastic Expansion Methods 205chapter of the Theory Manual \cite TheoMan for additional information 206on the PCE algorithm. 207 208<b> Expected HDF5 Output </b> 209 210If Dakota was built with HDF5 support and run with the 211\ref environment-results_output-hdf5 keyword, this method 212writes the following results to HDF5: 213 214- \ref hdf5_results-se_moments 215- \ref hdf5_results-pdf 216- \ref hdf5_results-level_mappings 217- \ref hdf5_results-vbd 218 219 220 221Topics:: 222 223Examples:: 224\verbatim 225method, 226 polynomial_chaos 227 sparse_grid_level = 2 228 samples = 10000 seed = 12347 rng rnum2 229 response_levels = .1 1. 50. 100. 500. 1000. 230 variance_based_decomp 231\endverbatim 232 233Theory:: 234 235Faq:: 236See_Also:: method-stoch_collocation, method-function_train 237