Download Active Subspaces: Emerging Ideas for Dimension Reduction in by Paul G. Constantine PDF

By Paul G. Constantine

ISBN-10: 1611973856

ISBN-13: 9781611973853

Scientists and engineers use desktop simulations to check relationships among a model's enter parameters and its outputs. besides the fact that, thorough parameter stories are difficult, if now not very unlikely, whilst the simulation is pricey and the version has a number of inputs. To allow reviews in those situations, the engineer may possibly try and decrease the measurement of the model's enter parameter house. energetic subspaces are an rising set of measurement aid instruments that establish very important instructions within the parameter house. This publication describes innovations for locating a model's energetic subspace and proposes equipment for exploiting the diminished measurement to permit another way infeasible parameter reviews. Readers will locate new rules for size relief, easy-to-implement algorithms, and several other examples of lively subspaces in action.

Parameter reports are in all places in computational technological know-how. complicated engineering simulations needs to run a number of instances with diversified inputs to successfully research the relationships among inputs and outputs. reviews like optimization, uncertainty quantification, and sensitivity research produce subtle characterizations of the input/output map. yet thorough parameter reports are tougher whilst every one simulation is dear and the variety of parameters is huge. In perform, the engineer could try and restrict a learn to crucial parameters, which successfully reduces the size of the parameter research.

Show description

Read Online or Download Active Subspaces: Emerging Ideas for Dimension Reduction in Parameter Studies PDF

Similar analysis books

Topological Nonlinear Analysis II: Degree, Singularity and Variations

The most objective of the current quantity is to offer a survey of a few of the main major achievements acquired through topological tools in nonlin­ ear research over the last 3 many years. it's meant, not less than partially, as a continuation of Topological Nonlinear research: measure, Singularity and Varia­ tions, released in 1995.

Computer assisted vegetation analysis

There are numerous books and laptop courses dealing glance forward instead of wondering the earlier. it is a with facts research. it'd be effortless to count number a minimum of a handbook of contemporary perspectives that advanced within the examine of hundred, but few of those might convey purposes in plants. This e-book is meant to stress the recent crops technological know-how.

Extra info for Active Subspaces: Emerging Ideas for Dimension Reduction in Parameter Studies

Sample text

5. 80). (c) and (d) show the estimates and bootstrap intervals on the distance between the estimated active subspace and the true active subspace. (a) and (c) are computed with the multiplier α = 2 when choosing M ; (b) and (d) use α = 10. The gap between the first and second eigenvalues is significant as judged by the gap between the bootstrap intervals. 79) where the xi ’s are independent standard normal random variables, and the pairs {γi , φi (s)} are the eigenpairs of the correlation operator (s, t) = exp −β−1 s − t 1 .

And Compute the norm of the total variance, σ 2 := X 2j . 7. Assume ∇x f ≤ L for all x ∈ ˆ −C ≥ C C t ≤ σ 2 /R, t ≥ σ 2 /R. Then for ∈ (0, 1], ≤ 2m exp −3M λ1 8L2 2 . 3. Computing the active subspace 31 Proof. 44) ≥ M t⎦ ≤ 2θ, where θ upper-bounds both probabilities. The final result of the proof is the upper bound θ. Note that ∇x f ∇x f T − C ρ d x = C − ∇x f ∇x f T ρ d x = 0. 6. 47) (∇x f ∇x f T − C)2 ρ d x ≤ M L2 C − C 2 ≤ M C L2 I − C ≤ M λ 1 L2 . The last line follows from the fact that λ1 ≤ L2 .

N ) with n < m, and W1 contains the first n eigenvectors. Define the new variables y and z by y = W1T x ∈ n , z = W2T x ∈ m−n . 9) 24 Chapter 3. Discover the Active Subspace Any x ∈ m can be expressed in terms of y and z, x = W W T x = W1 W1T x + W2 W2T x = W1 y + W2 z. 10) once more. We exploit this decomposition several times in Chapter 4. In particular, f (x) = f (W1 y + W2 z) implies that we can compute f ’s gradient with respect to y and z. These partial derivatives are directional derivatives of f along the directions defined by the eigenvectors.

Download PDF sample

Rated 4.78 of 5 – based on 39 votes