Derivative based global sensitivity analysis and its entropic link

Jiannan Yang

Derivative based global sensitivity analysis and its entropic link

Physics

Research output: Working paper › Preprint

Abstract

Distribution-based global sensitivity analysis (GSA), such as variance-based and entropy-based approaches, can provide quantitative sensitivity information. However, they can be expensive to evaluate and are thus limited to low dimensional problems. Derivative-based GSA, on the other hand, require much fewer model evaluations. It is known that derivative-based GSA is closely linked to variance-based total sensitivity index, while its relationship with the entropy-based measure is unclear. To fill this gap, we introduce a log-derivative based functional to demonstrate that the entropy-based and derivative-based sensitivity measures are strongly connected. In particular, we give proofs that, similar to the case with variance-based GSA, there is an inequality relationship between entropy-based and derivative-based important measures. Both analytical and numerical verifications are provided. Examples show that the derivative-based methods give similar variable rankings as entropy-based index and can thus be potentially used as a proxy for both variance-based and entropy-based distribution-type GSA.

Original language	Undefined/Unknown
Number of pages	11
Publication status	Published - 1 Oct 2023

Bibliographical note

11 page, 3 figures, 4 tables

Keywords

math.NA
cs.NA
math.PR
stat.CO

Access to Document

2310.00551v1

Cite this

@techreport{fef81b3d1e404f05b9670f5ae57e0ed2,

title = "Derivative based global sensitivity analysis and its entropic link",

abstract = " Distribution-based global sensitivity analysis (GSA), such as variance-based and entropy-based approaches, can provide quantitative sensitivity information. However, they can be expensive to evaluate and are thus limited to low dimensional problems. Derivative-based GSA, on the other hand, require much fewer model evaluations. It is known that derivative-based GSA is closely linked to variance-based total sensitivity index, while its relationship with the entropy-based measure is unclear. To fill this gap, we introduce a log-derivative based functional to demonstrate that the entropy-based and derivative-based sensitivity measures are strongly connected. In particular, we give proofs that, similar to the case with variance-based GSA, there is an inequality relationship between entropy-based and derivative-based important measures. Both analytical and numerical verifications are provided. Examples show that the derivative-based methods give similar variable rankings as entropy-based index and can thus be potentially used as a proxy for both variance-based and entropy-based distribution-type GSA. ",

keywords = "math.NA, cs.NA, math.PR, stat.CO",

author = "Jiannan Yang",

note = "11 page, 3 figures, 4 tables",

year = "2023",

month = oct,

day = "1",

language = "Undefined/Unknown",

type = "WorkingPaper",

}

TY - UNPB

T1 - Derivative based global sensitivity analysis and its entropic link

AU - Yang, Jiannan

N1 - 11 page, 3 figures, 4 tables

PY - 2023/10/1

Y1 - 2023/10/1

N2 - Distribution-based global sensitivity analysis (GSA), such as variance-based and entropy-based approaches, can provide quantitative sensitivity information. However, they can be expensive to evaluate and are thus limited to low dimensional problems. Derivative-based GSA, on the other hand, require much fewer model evaluations. It is known that derivative-based GSA is closely linked to variance-based total sensitivity index, while its relationship with the entropy-based measure is unclear. To fill this gap, we introduce a log-derivative based functional to demonstrate that the entropy-based and derivative-based sensitivity measures are strongly connected. In particular, we give proofs that, similar to the case with variance-based GSA, there is an inequality relationship between entropy-based and derivative-based important measures. Both analytical and numerical verifications are provided. Examples show that the derivative-based methods give similar variable rankings as entropy-based index and can thus be potentially used as a proxy for both variance-based and entropy-based distribution-type GSA.

AB - Distribution-based global sensitivity analysis (GSA), such as variance-based and entropy-based approaches, can provide quantitative sensitivity information. However, they can be expensive to evaluate and are thus limited to low dimensional problems. Derivative-based GSA, on the other hand, require much fewer model evaluations. It is known that derivative-based GSA is closely linked to variance-based total sensitivity index, while its relationship with the entropy-based measure is unclear. To fill this gap, we introduce a log-derivative based functional to demonstrate that the entropy-based and derivative-based sensitivity measures are strongly connected. In particular, we give proofs that, similar to the case with variance-based GSA, there is an inequality relationship between entropy-based and derivative-based important measures. Both analytical and numerical verifications are provided. Examples show that the derivative-based methods give similar variable rankings as entropy-based index and can thus be potentially used as a proxy for both variance-based and entropy-based distribution-type GSA.

KW - math.NA

KW - cs.NA

KW - math.PR

KW - stat.CO

M3 - Preprint

BT - Derivative based global sensitivity analysis and its entropic link

ER -