convex functions (i)vector composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom...

57
Convex Functions (I) Lijun Zhang [email protected] http://cs.nju.edu.cn/zlj

Upload: others

Post on 10-Mar-2020

6 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Convex Functions (I)

Lijun [email protected]://cs.nju.edu.cn/zlj

Page 2: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Outline Basic Properties

Definition First-order Conditions, Second-order Conditions Jensenโ€™s inequality and extensions Epigraph

Operations That Preserve Convexity Nonnegative Weighted Sums Composition with an affine mapping Pointwise maximum and supremum Composition Minimization Perspective of a function

Summary

Page 3: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Outline Basic Properties

Definition First-order Conditions, Second-order Conditions Jensenโ€™s inequality and extensions Epigraph

Operations That Preserve Convexity Nonnegative Weighted Sums Composition with an affine mapping Pointwise maximum and supremum Composition Minimization Perspective of a function

Summary

Page 4: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Convex Function

is convex if is convex

Page 5: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Convex Function

is convex if is convex

is strictly convex if

Page 6: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Convex Function

is convex if is convex

is concave if is convex is convex

Affine functions are both convex and concave, and vice versa.

Page 7: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Extended-value Extensions

The extended-value extension of is

Example

Page 8: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Extended-value Extensions

The extended-value extension of is

Example Indicator Function of a Set

Page 9: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Zeroth-order Condition

Definition High-dimensional space

A function is convex if and only if it is convex when restricted to any line that intersects its domain. , is convex is convex One-dimensional space

Page 10: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

First-order Conditions

is differentiable. Then is convex if and only if is convex For all

๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅ ๐›ป๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

First-order Taylor approximation

Page 11: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

First-order Conditions

is differentiable. Then is convex if and only if is convex For all

Local Information Global Information

is strictly convex if and only if

๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅ ๐›ป๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅ ๐›ป๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

Page 12: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Proof

is convex

Necessary condition:๐‘“ ๐‘ฅ ๐‘ก ๐‘ฆ ๐‘ฅ 1 ๐‘ก ๐‘“ ๐‘ฅ ๐‘ก๐‘“ ๐‘ฆ , 0 ๐‘ก 1

โ‡’ ๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅโ†’

๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅ ๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

Sufficient condition:๐‘ง ๐œƒ๐‘ฅ 1 ๐œƒ ๐‘ฆ

๐‘“ ๐‘ฅ ๐‘“ ๐‘ง ๐‘“ ๐‘ง ๐‘ฅ ๐‘ง๐‘“ ๐‘ฆ ๐‘“ ๐‘ง ๐‘“ ๐‘ง ๐‘ฆ ๐‘ง

โ‡’ ๐‘“ ๐‘ฅ ๐‘“ ๐‘ง 1 ๐œƒ ๐‘“ ๐‘ง ๐‘ฅ ๐‘ฆ๐‘“ ๐‘ฆ ๐‘“ ๐‘ง ๐œƒ๐‘“ ๐‘ง ๐‘ฅ ๐‘ฆ

โ‡’ ๐œƒ๐‘“ ๐‘ฅ 1 ๐œƒ ๐‘“ ๐‘ฆ ๐‘“ ๐‘ง โ‡’ ๐‘“ ๐œƒ๐‘ฅ 1 ๐œƒ ๐‘ฆ ๐œƒ๐‘“ ๐‘ฅ 1 ๐œƒ ๐‘“ ๐‘ฆ

Page 13: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Proof

is convex

is convex

๐‘“ is convex โ‡’ is convex โ‡’ ๐‘” 1 ๐‘” 0๐‘” 0 โ‡’ ๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅ ๐›ป๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

๐‘” ๐‘ก ๐‘“ ๐‘ก๐‘ฆ 1 ๐‘ก ๐‘ฅ , ๐‘”โ€ฒ ๐‘ก ๐›ป๐‘“ ๐‘ก๐‘ฆ 1 ๐‘ก ๐‘ฅ ๐‘ฆ ๐‘ฅ

๐‘“ is convex

๐‘” is convex

First-order condition of ๐‘”

First-order condition of ๐‘“

Page 14: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Proof

is convex

is convex

โ‡’ ๐‘” ๐‘ก ๐‘” ๐‘ก ๐‘” ๐‘ก ๐‘ก ๐‘ก โ‡’ ๐‘” ๐‘ก is convex โ‡’๐‘“ is convex

๐‘” ๐‘ก ๐‘“ ๐‘ก๐‘ฆ 1 ๐‘ก ๐‘ฅ ,

๐‘“ ๐‘ก๐‘ฆ 1 ๐‘ก ๐‘ฅ ๐‘“ ๐‘ก๐‘ฆ 1 ๐‘ก ๐‘ฅ ๐›ป๐‘“ ๐‘ก๐‘ฆ 1 ๐‘ก ๐‘ฅ ๐‘ฆ ๐‘ฅ ๐‘ก ๐‘ก

๐‘”โ€ฒ ๐‘ก ๐›ป๐‘“ ๐‘ก๐‘ฆ 1 ๐‘ก ๐‘ฅ ๐‘ฆ ๐‘ฅ

๐‘“ is convex

๐‘” is convex

First-order condition of ๐‘”

First-order condition of ๐‘“

Page 15: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Second-order Conditions

is twice differentiable. Then is convex if and only if is convex For all ,

Attention is strictly convex is strict convex

is strict convex but is convex is necessary,

Page 16: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Functions on is convex on ,

is convex on when or ,

and concave for

, for , is convex on

is concave on

Negative entropy is convex on

Page 17: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Functions on Every norm on is convex

Quadratic-over-linear: dom ๐‘“ ๐‘ฅ, ๐‘ฆ โˆˆ ๐‘ ๐‘ฆ 0

/ is concave on

is concave on

Page 18: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Functions on Every norm on is convex

๐‘“ ๐‘ฅ is a norm on ๐‘

๐‘“ ๐œƒ๐‘ฅ 1 ๐œƒ ๐‘ฆ ๐‘“ ๐œƒ๐‘ฅ ๐‘“ 1 ๐œƒ ๐‘ฆ

๐œƒ๐‘“ ๐‘ฅ 1 ๐œƒ ๐‘“ ๐‘ฆ

๐‘“ ๐œƒ๐‘ฅ 1 ๐œƒ ๐‘ฆ max ๐œƒ๐‘ฅ 1 ๐œƒ ๐‘ฆ

๐œƒmax ๐‘ฅ 1 ๐œƒ max ๐‘ฆ

Page 19: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Functions on

๐›ป ๐‘“ ๐‘ฅ, ๐‘ฆ ๐‘ฆ ๐‘ฅ๐‘ฆ๐‘ฅ๐‘ฆ ๐‘ฅ

๐‘ฆ๐‘ฅ ๐‘ฆ

๐‘ฅ โ‰ฝ 0

Page 20: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Functions on

Page 21: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Functions on

๐›ป ๐‘“ ๐‘ฅ๐Ÿ

๐Ÿ ๐‘ง diag ๐‘ง ๐‘ง๐‘ง

๐‘ง ๐‘’ , โ€ฆ ๐‘’

๐‘ฃ ๐›ป ๐‘“ ๐‘ฅ ๐‘ฃ๐Ÿ

โˆ‘ ๐‘ง โˆ‘ ๐‘ฃ ๐‘ง

โˆ‘ ๐‘ฃ ๐‘ง 0

Cauchy-Schwarz inequality: ๐‘Ž ๐‘Ž ๐‘ ๐‘

๐‘Ž ๐‘

Page 22: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Functions on is concave on

๐‘” ๐‘ก ๐‘“ ๐‘ ๐‘ก๐‘‰ , ๐‘ ๐‘ก๐‘‰ โ‰ป 0, ๐‘ โ‰ป 0

๐‘” ๐‘ก log det ๐‘ ๐‘ก๐‘‰

log det ๐‘ ๐ผ ๐‘ก๐‘ ๐‘‰๐‘ ๐‘

โˆ‘ log 1 ๐‘ก๐œ† log det ๐‘

๐œ† , โ€ฆ ๐œ† are the eigenvalues of ๐‘ ๐‘‰๐‘

๐‘” ๐‘ก โˆ‘ , ๐‘” ๐‘ก โˆ‘

det ๐ด๐ต det ๐ด det ๐ต https://en.wikipedia.org/wiki/Determinant

Page 23: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Sublevel Sets

-sublevel set

is convex is convex is convex is convex

-superlevel set

is concave convex

is convex

๐ถ ๐‘ฅ โˆˆ dom ๐‘“ ๐‘“ ๐‘ฅ ๐›ผ

๐ถ ๐‘ฅ โˆˆ dom ๐‘“ ๐‘“ ๐‘ฅ ๐›ผ

Page 24: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Epigraph

Graph of function

Epigraph of function

Page 25: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Epigraph

Epigraph of function

Hypograph

Conditions is convex is convex is concave is convex

Page 26: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Example

Matrix Fractional Function

Quadratic-over-linear:

Schur complement condition is convex Recall Example 2.10 in the book

๐‘“ ๐‘ฅ, ๐‘Œ ๐‘ฅ ๐‘Œ ๐‘ฅ, dom ๐‘“ ๐‘ ๐’

Page 27: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Application of Epigraph

First order Condition ๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅ ๐›ป๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

๐‘ฆ, ๐‘ก โˆˆ epi ๐‘“ โ‡’ ๐‘ก ๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅ ๐›ป๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

Page 28: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Application of Epigraph

First order Condition ๐‘“ ๐‘ฆ ๐‘“ ๐‘ฅ ๐›ป๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

๐‘ฆ, ๐‘ก โˆˆ epi ๐‘“ โ‡’ ๐‘ก ๐‘“ ๐‘ฅ ๐›ป๐‘“ ๐‘ฅ ๐‘ฆ ๐‘ฅ

๐‘ฆ, ๐‘ก โˆˆ epi ๐‘“ โ‡’ ๐›ป๐‘“ ๐‘ฅ1

๐‘ฆ๐‘ก

๐‘ฅ๐‘“ ๐‘ฅ 0

Page 29: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Jensenโ€™s Inequality

Basic inequality

points

Page 30: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Jensenโ€™s Inequality

Infinite points

๐‘“ ๐‘ฅ ๐„๐‘“ ๐‘ฅ ๐‘ง , ๐‘ง is a zero-mean noisy

Hรถlderโ€™s inequality

Page 31: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Outline Basic Properties

Definition First-order Conditions, Second-order Conditions Jensenโ€™s inequality and extensions Epigraph

Operations That Preserve Convexity Nonnegative Weighted Sums Composition with an affine mapping Pointwise maximum and supremum Composition Minimization Perspective of a function

Summary

Page 32: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Nonnegative Weighted Sums Finite sums ๐‘ค 0, ๐‘“ is convex ๐‘“ ๐‘ค ๐‘“ โ‹ฏ ๐‘ค ๐‘“ is convex

Infinite sums ๐‘“ ๐‘ฅ, ๐‘ฆ is convex in ๐‘ฅ, โˆ€๐‘ฆ โˆˆ ๐’œ, ๐‘ค ๐‘ฆ 0 ๐‘” ๐‘ฅ ๐‘“ ๐‘ฅ, ๐‘ฆ ๐‘ค ๐‘ฆ๐’œ ๐‘‘๐‘ฆ is convex

Epigraph interpretation ๐ž๐ฉ๐ข ๐‘ค๐‘“ ๐‘ฅ, ๐‘ก |๐‘ค๐‘“ ๐‘ฅ ๐‘ก

๐ผ 00 ๐‘ค ๐ž๐ฉ๐ข ๐‘“ ๐‘ฅ, ๐‘ค๐‘ก |๐‘“ ๐‘ฅ ๐‘ก

๐ž๐ฉ๐ข ๐‘ค๐‘“ ๐ผ 00 ๐‘ค ๐ž๐ฉ๐ข ๐‘“

Page 33: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Composition with an affine mapping

Affine Mapping

If is convex, so is

If is concave, so is

๐‘” ๐‘ฅ ๐‘“ ๐ด๐‘ฅ ๐‘

Page 34: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Pointwise Maximum

is convex

is convex with

max ๐‘“ ๐œƒ๐‘ฅ 1 ๐œƒ ๐‘ฆ , ๐‘“ ๐œƒ๐‘ฅ 1 ๐œƒ ๐‘ฆmax ๐œƒ๐‘“ ๐‘ฅ 1 ๐œƒ ๐‘“ ๐‘ฆ , ๐œƒ๐‘“ ๐‘ฅ 1

๐œƒ ๐‘“ ๐‘ฆ๐œƒ max ๐‘“ ๐‘ฅ , ๐‘“ ๐‘ฅ 1 ๐œƒ max ๐‘“ ๐‘ฆ , ๐‘“ ๐‘ฆ๐œƒ๐‘“ ๐‘ฅ 1 ๐œƒ ๐‘“ ๐‘ฆ

๐‘“ , โ€ฆ ๐‘“ is convex โ‡’ ๐‘“ ๐‘ฅ max ๐‘“ ๐‘ฅ , โ€ฆ ๐‘“ ๐‘ฅ

Page 35: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Piecewise-linear functions

Sum of largest components

is convex

Pointwise maximum of !! !

linear functions

Page 36: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Pointwise Supremum

is convex in

is convex with

โˆˆ๐’œ

Epigraph interpretation โˆˆ๐’œ

Intersection of convex sets is convex Pointwise infimum of a set of

concave functions is concave

โˆˆ๐’œ

Page 37: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Support function of a set

โˆˆ

Distance to farthest point of a set

โˆˆ

Page 38: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Maximum eigenvalue of a symmetric matrix

Norm of a matrix is maximum singular value

of

Page 39: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Representation

Almost every convex function can be expressed as the pointwise supremum of a family of affine functions.

โŸน ๐‘“ ๐‘ฅ sup ๐‘” ๐‘ฅ |๐‘” affine, ๐‘” z ๐‘“ ๐‘ง โˆ€๐‘ง

๐‘“: ๐‘ โ†’ ๐‘ is convex and dom ๐‘“ ๐‘

Page 40: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Compositions

Definition

Chain Rule

๐‘“ ๐‘ฅ โ„Ž ๐‘” ๐‘ฅ

๐›ป ๐‘“ ๐‘ฅ โ„Ž ๐‘” ๐‘ฅ ๐›ป ๐‘” ๐‘ฅ โ„Ž ๐‘” ๐‘ฅ ๐›ป๐‘” ๐‘ฅ ๐›ป๐‘” ๐‘ฅ

Page 41: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Scalar Composition

and are twice differentiable

is convex, if

โ„Ž is convex and nondecreasing ๐‘” is convex

โ„Ž is convex and nonincreasing, ๐‘” is concave

๐‘“ ๐‘ฅ โ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒ ๐‘ฅ โ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒโ€ฒ ๐‘ฅ

Page 42: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Scalar Composition

and are twice differentiable

is concave, if

โ„Ž is concave and nondecreasing ๐‘” is concave

โ„Ž is concave and nonincreasing, ๐‘” is convex

๐‘“ ๐‘ฅ โ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒ ๐‘ฅ โ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒโ€ฒ ๐‘ฅ

Page 43: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Without differentiability assumption Without domain condition

โ„Ž ๐‘ฅ 0 with dom โ„Ž 1,2 , which is convex and non-decreasing

๐‘” ๐‘ฅ ๐‘ฅ with dom ๐‘” ๐‘, which is convex

dom ๐‘“ 2, 1 โˆช 1, 2

๐‘“ ๐‘ฅ โ„Ž ๐‘” ๐‘ฅ 0

Scalar Composition

Page 44: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Without differentiability assumption Without domain condition

โ„Ž is convex, โ„Ž is nondecreasing, and ๐‘” is convex โ‡’ ๐‘“ is convex

โ„Ž is convex, โ„Ž is nonincreasing, and ๐‘” is concave โ‡’ ๐‘“ is convex

The conditions for concave are similar

Scalar Composition

Page 45: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Extended-value Extensions

Page 46: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples is convex is convex is concave and positive is

concave is concave and positive is

convex is convex and nonnegative and

is convex is convex is convex on

Page 47: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Vector Composition

โ„Ž and ๐‘” are twice differentiable dom ๐‘” ๐‘, dom โ„Ž ๐‘

๐‘“ โ„Ž โˆ˜ ๐‘” โ„Ž ๐‘” ๐‘ฅ , โ€ฆ , ๐‘” ๐‘ฅ

๐‘“ ๐‘ฅ ๐‘” ๐‘ฅ ๐›ป โ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒ ๐‘ฅ ๐›ปโ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒโ€ฒ ๐‘ฅ

๐‘“โ€ฒ ๐‘ฅ ๐›ปโ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒ ๐‘ฅ

Page 48: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Vector Composition

โ„Ž and ๐‘” are twice differentiable dom ๐‘” ๐‘, dom โ„Ž ๐‘

๐‘“ is convex, if ๐‘“ ๐‘ฅ 0 โ„Ž is convex, โ„Ž is nondecreasing in each

argument, and ๐‘” are convex โ„Ž is convex, โ„Ž is nonincreasing in each

argument, and ๐‘” are concave

๐‘“ โ„Ž โˆ˜ ๐‘” โ„Ž ๐‘” ๐‘ฅ , โ€ฆ , ๐‘” ๐‘ฅ

๐‘“ ๐‘ฅ ๐‘” ๐‘ฅ ๐›ป โ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒ ๐‘ฅ ๐›ปโ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒโ€ฒ ๐‘ฅ

Page 49: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Vector Composition

โ„Ž and ๐‘” are twice differentiable dom ๐‘” ๐‘, dom โ„Ž ๐‘

๐‘“ is concave, if ๐‘“ ๐‘ฅ 0 โ„Ž is concave, โ„Ž is nondecreasing in each

argument, and ๐‘” are concave

The general case is similar

๐‘“ โ„Ž โˆ˜ ๐‘” โ„Ž ๐‘” ๐‘ฅ , โ€ฆ , ๐‘” ๐‘ฅ

๐‘“ ๐‘ฅ ๐‘” ๐‘ฅ ๐›ป โ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒ ๐‘ฅ ๐›ปโ„Ž ๐‘” ๐‘ฅ ๐‘”โ€ฒโ€ฒ ๐‘ฅ

Page 50: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples โ„Ž ๐‘ง ๐‘ง โ‹ฏ ๐‘ง , ๐‘ง โˆˆ ๐‘ , ๐‘” , โ€ฆ , ๐‘” is convex

โ‡’ โ„Ž โˆ˜ ๐‘” is convex

โ„Ž ๐‘ง log โˆ‘ ๐‘’ , ๐‘” , โ€ฆ , ๐‘” is convex โ‡’ โ„Ž โˆ˜๐‘” is convex

โ„Ž ๐‘ง โˆ‘ ๐‘ง/ on ๐‘ is concave for 0 ๐‘

1, and its extension is nondecreasing. If ๐‘” is concave and nonnegative โ‡’ โ„Ž โˆ˜ ๐‘” is concave

Page 51: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Minimization is convex in is convex ๐‘” ๐‘ฅ inf

โˆˆ๐‘“ ๐‘ฅ, ๐‘ฆ is convex if ๐‘” ๐‘ฅ

โˆž, โˆ€ ๐‘ฅ โˆˆ dom ๐‘” dom ๐‘” ๐‘ฅ ๐‘ฅ, ๐‘ฆ โˆˆ dom ๐‘“ for some ๐‘ฆ โˆˆ ๐ถ

Proof by Epigraph epi ๐‘” ๐‘ฅ, ๐‘ก | ๐‘ฅ, ๐‘ฆ, ๐‘ก โˆˆ epi ๐‘“ for some ๐‘ฆ โˆˆ ๐ถ The projection of a convex set is convex.

Page 52: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples Schur complement ๐‘“ ๐‘ฅ, ๐‘ฆ ๐‘ฅ ๐ด๐‘ฅ 2๐‘ฅ ๐ต๐‘ฆ ๐‘ฆ ๐ถ๐‘ฆ ๐ด ๐ต

๐ต ๐ถ โ‰ฝ 0, ๐ด, ๐ถ is symmetric โ‡’ ๐‘“ ๐‘ฅ, ๐‘ฆ is convex ๐‘” ๐‘ฅ inf ๐‘“ ๐‘ฅ, ๐‘ฆ ๐‘ฅ ๐ด ๐ต๐ถ ๐ต ๐‘ฅ is convex

โ‡’ ๐ด ๐ต๐ถ ๐ต โ‰ฝ 0, ๐ถ is the pseudo-inverse of ๐ถ

Distance to a set ๐‘† is a convex nonempty set,๐‘“ ๐‘ฅ, ๐‘ฆ โ€–๐‘ฅ ๐‘ฆโ€– is

convex in ๐‘ฅ, ๐‘ฆ ๐‘” ๐‘ฅ dist ๐‘ฅ, ๐‘† inf

โˆˆโ€–๐‘ฅ ๐‘ฆโ€–

Page 53: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Examples

Affine domain โ„Ž ๐‘ฆ is convex ๐‘” ๐‘ฅ inf โ„Ž ๐‘ฆ |๐ด๐‘ฆ ๐‘ฅ is convex

Proof

๐‘“ ๐‘ฅ, ๐‘ฆ โ„Ž ๐‘ฆ if ๐ด๐‘ฆ ๐‘ฅ โˆž otherwise

๐‘“ ๐‘ฅ, ๐‘ฆ is convex in ๐‘ฅ, ๐‘ฆ ๐‘” is the minimum of ๐‘“ over ๐‘ฆ

Page 54: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Perspective of a function

defined as

is the perspective of dom ๐‘” ๐‘ฅ, ๐‘ก |๐‘ฅ/๐‘ก โˆˆ dom ๐‘“, ๐‘ก 0 ๐‘“ is convex โ‡’ ๐‘” is convex

Proof

Perspective mapping preserve convexity

๐‘ฅ, ๐‘ก, ๐‘  โˆˆ epi ๐‘” โ‡” ๐‘ก๐‘“๐‘ฅ๐‘ก ๐‘ 

โ‡” ๐‘“๐‘ฅ๐‘ก

๐‘ ๐‘ก

โ‡” ๐‘ฅ/๐‘ก, ๐‘ /๐‘ก โˆˆ epi ๐‘“

Page 55: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Example

Euclidean norm squared

Composition with an Affine function is convex

is convex

Page 56: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Outline Basic Properties

Definition First-order Conditions, Second-order Conditions Jensenโ€™s inequality and extensions Epigraph

Operations That Preserve Convexity Nonnegative Weighted Sums Composition with an affine mapping Pointwise maximum and supremum Composition Minimization Perspective of a function

Summary

Page 57: Convex Functions (I)Vector Composition รž รœ โ„Žand ๐‘”are twice differentiable dom ๐‘” รœ๐‘,dom โ„Ž๐‘ รž ๐‘“is convex, if ๐‘“ รฑ รฑ๐‘ฅ0 โ„Žis convex, โ„Žis nondecreasing

Summary Basic Properties

Definition First-order Conditions, Second-order Conditions Jensenโ€™s inequality and extensions Epigraph

Operations That Preserve Convexity Nonnegative Weighted Sums Composition with an affine mapping Pointwise maximum and supremum Composition Minimization Perspective of a function