22
Outline Math 408A Convergence of Backtracking Methods Math 408A Convergence of Backtracking Methods

Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Math 408AConvergence of Backtracking Methods

Math 408A Convergence of Backtracking Methods

Page 2: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Math 408A Convergence of Backtracking Methods

Page 3: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity

We say that F : Rn → Rm is Lipschitz continuous relative to aset D ⊂ Rn if there exists a constant K ≥ 0 such that

‖F (x)− F (y)‖ ≤ K‖x − y‖

for all x , y ∈ D.

Fact: Lipschitz continuity implies uniform continuity.

Examples:

1. f (x) = x−1 is continuous on (0, 1), but it is not uniformlycontinuous on (0, 1).

2. f (x) =√

x is uniformly continuous on [0, 1], but it is notLipschitz continuous on [0, 1].

Math 408A Convergence of Backtracking Methods

Page 4: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity

We say that F : Rn → Rm is Lipschitz continuous relative to aset D ⊂ Rn if there exists a constant K ≥ 0 such that

‖F (x)− F (y)‖ ≤ K‖x − y‖

for all x , y ∈ D.

Fact: Lipschitz continuity implies uniform continuity.

Examples:

1. f (x) = x−1 is continuous on (0, 1), but it is not uniformlycontinuous on (0, 1).

2. f (x) =√

x is uniformly continuous on [0, 1], but it is notLipschitz continuous on [0, 1].

Math 408A Convergence of Backtracking Methods

Page 5: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity

We say that F : Rn → Rm is Lipschitz continuous relative to aset D ⊂ Rn if there exists a constant K ≥ 0 such that

‖F (x)− F (y)‖ ≤ K‖x − y‖

for all x , y ∈ D.

Fact: Lipschitz continuity implies uniform continuity.

Examples:

1. f (x) = x−1 is continuous on (0, 1), but it is not uniformlycontinuous on (0, 1).

2. f (x) =√

x is uniformly continuous on [0, 1], but it is notLipschitz continuous on [0, 1].

Math 408A Convergence of Backtracking Methods

Page 6: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity

We say that F : Rn → Rm is Lipschitz continuous relative to aset D ⊂ Rn if there exists a constant K ≥ 0 such that

‖F (x)− F (y)‖ ≤ K‖x − y‖

for all x , y ∈ D.

Fact: Lipschitz continuity implies uniform continuity.

Examples:

1. f (x) = x−1 is continuous on (0, 1), but it is not uniformlycontinuous on (0, 1).

2. f (x) =√

x is uniformly continuous on [0, 1], but it is notLipschitz continuous on [0, 1].

Math 408A Convergence of Backtracking Methods

Page 7: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity

We say that F : Rn → Rm is Lipschitz continuous relative to aset D ⊂ Rn if there exists a constant K ≥ 0 such that

‖F (x)− F (y)‖ ≤ K‖x − y‖

for all x , y ∈ D.

Fact: Lipschitz continuity implies uniform continuity.

Examples:

1. f (x) = x−1 is continuous on (0, 1), but it is not uniformlycontinuous on (0, 1).

2. f (x) =√

x is uniformly continuous on [0, 1], but it is notLipschitz continuous on [0, 1].

Math 408A Convergence of Backtracking Methods

Page 8: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity and the Derivative

Fact:If F ′ exists and is continuous on a compact convex setD ⊂ Rm, then F is Lipschitz continuous on D.

Proof: By the MVT

‖F (x)− F (y)‖ ≤

(sup

z∈[x ,y ]‖F ′(z)‖

)‖x − y‖.

Lipschitz continuity is almost but not quite a differentiabilityhypothesis. The Lipschitz constant provides bounds on rate ofchange. For example, every norm is Lipschitz continuous, but notdifferentiable at the origin.

Math 408A Convergence of Backtracking Methods

Page 9: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity and the Derivative

Fact:If F ′ exists and is continuous on a compact convex setD ⊂ Rm, then F is Lipschitz continuous on D.

Proof: By the MVT

‖F (x)− F (y)‖ ≤

(sup

z∈[x ,y ]‖F ′(z)‖

)‖x − y‖.

Lipschitz continuity is almost but not quite a differentiabilityhypothesis. The Lipschitz constant provides bounds on rate ofchange. For example, every norm is Lipschitz continuous, but notdifferentiable at the origin.

Math 408A Convergence of Backtracking Methods

Page 10: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity and the Derivative

Fact:If F ′ exists and is continuous on a compact convex setD ⊂ Rm, then F is Lipschitz continuous on D.

Proof: By the MVT

‖F (x)− F (y)‖ ≤

(sup

z∈[x ,y ]‖F ′(z)‖

)‖x − y‖.

Lipschitz continuity is almost but not quite a differentiabilityhypothesis. The Lipschitz constant provides bounds on rate ofchange.

For example, every norm is Lipschitz continuous, but notdifferentiable at the origin.

Math 408A Convergence of Backtracking Methods

Page 11: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity and the Derivative

Fact:If F ′ exists and is continuous on a compact convex setD ⊂ Rm, then F is Lipschitz continuous on D.

Proof: By the MVT

‖F (x)− F (y)‖ ≤

(sup

z∈[x ,y ]‖F ′(z)‖

)‖x − y‖.

Lipschitz continuity is almost but not quite a differentiabilityhypothesis. The Lipschitz constant provides bounds on rate ofchange. For example, every norm is Lipschitz continuous, but notdifferentiable at the origin.

Math 408A Convergence of Backtracking Methods

Page 12: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

The Quadratic Bound Lemma (QBL)

Let F : Rn → Rm be such that F ′ is Lipschitz continuous on theconvex set D ⊂ Rn. Then

‖F (y)− (F (x) + F ′(x)(y − x))‖ ≤ K

2‖y − x‖2

for all x , y ∈ D where K is a Lipschitz constant for F ′ on D.

Math 408A Convergence of Backtracking Methods

Page 13: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Lipschitz Continuity and the Quadratic Bound Lemma

Proof:

F (y)− F (x)− F ′(x)(y − x) =∫ 1

0F ′(x + t(y − x))(y − x)dt − F ′(x)(y − x)

=∫ 1

0[F ′(x + t(y − x))− F ′(x)](y − x)dt

‖F (y)− (F (x) + F ′(x)(y − x))‖ = ‖∫ 1

0[F ′(x + t(y − x))− F ′(x)](y − x)dt‖

≤∫ 1

0‖(F ′(x + t(y − x)− F ′(x))(y − x)‖dt

≤∫ 1

0‖F ′(x + t(y − x))− F ′(x)‖ ‖y − x‖dt

≤∫ 1

0Kt‖y − x‖2dt

= K2 ‖y − x‖2.

Math 408A Convergence of Backtracking Methods

Page 14: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Theorem: Convergence of Backtracking

Let f : Rn → R and x0 ∈ R be such that f is differentiable on Rn with∇f Lipschitz continuous on an open convex set containing the set{x : f (x) ≤ f (x0)}. Let {xk} be the sequence satisfying xk+1 = xk

if∇f (xk) = 0; otherwise,

xk+1 = xk + tkdk , where dk satisfies f ′(xk ; dk) < 0,

and tk is chosen by the backtracking stepsize selection method.

Thenone of thefollowing statements must be true:

(i) There is a k0 such that ∇f ′(xk0) = 0.

(ii) f (xk)↘ −∞(iii) The sequence {‖dk‖} diverges (‖dk‖ → ∞).

(iv) For every subsequence J ⊂ N for which {dk : k ∈ J} is bounded, wehave

limk∈J

f ′(xk ; dk) = 0.

Math 408A Convergence of Backtracking Methods

Page 15: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Theorem: Convergence of Backtracking

Let f : Rn → R and x0 ∈ R be such that f is differentiable on Rn with∇f Lipschitz continuous on an open convex set containing the set{x : f (x) ≤ f (x0)}. Let {xk} be the sequence satisfying xk+1 = xk

if∇f (xk) = 0; otherwise,

xk+1 = xk + tkdk , where dk satisfies f ′(xk ; dk) < 0,

and tk is chosen by the backtracking stepsize selection method. Thenone of thefollowing statements must be true:

(i) There is a k0 such that ∇f ′(xk0) = 0.

(ii) f (xk)↘ −∞(iii) The sequence {‖dk‖} diverges (‖dk‖ → ∞).

(iv) For every subsequence J ⊂ N for which {dk : k ∈ J} is bounded, wehave

limk∈J

f ′(xk ; dk) = 0.

Math 408A Convergence of Backtracking Methods

Page 16: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Theorem: Convergence of Backtracking

Let f : Rn → R and x0 ∈ R be such that f is differentiable on Rn with∇f Lipschitz continuous on an open convex set containing the set{x : f (x) ≤ f (x0)}. Let {xk} be the sequence satisfying xk+1 = xk

if∇f (xk) = 0; otherwise,

xk+1 = xk + tkdk , where dk satisfies f ′(xk ; dk) < 0,

and tk is chosen by the backtracking stepsize selection method. Thenone of thefollowing statements must be true:

(i) There is a k0 such that ∇f ′(xk0) = 0.

(ii) f (xk)↘ −∞(iii) The sequence {‖dk‖} diverges (‖dk‖ → ∞).

(iv) For every subsequence J ⊂ N for which {dk : k ∈ J} is bounded, wehave

limk∈J

f ′(xk ; dk) = 0.

Math 408A Convergence of Backtracking Methods

Page 17: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Theorem: Convergence of Backtracking

Let f : Rn → R and x0 ∈ R be such that f is differentiable on Rn with∇f Lipschitz continuous on an open convex set containing the set{x : f (x) ≤ f (x0)}. Let {xk} be the sequence satisfying xk+1 = xk

if∇f (xk) = 0; otherwise,

xk+1 = xk + tkdk , where dk satisfies f ′(xk ; dk) < 0,

and tk is chosen by the backtracking stepsize selection method. Thenone of thefollowing statements must be true:

(i) There is a k0 such that ∇f ′(xk0) = 0.

(ii) f (xk)↘ −∞

(iii) The sequence {‖dk‖} diverges (‖dk‖ → ∞).

(iv) For every subsequence J ⊂ N for which {dk : k ∈ J} is bounded, wehave

limk∈J

f ′(xk ; dk) = 0.

Math 408A Convergence of Backtracking Methods

Page 18: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Theorem: Convergence of Backtracking

Let f : Rn → R and x0 ∈ R be such that f is differentiable on Rn with∇f Lipschitz continuous on an open convex set containing the set{x : f (x) ≤ f (x0)}. Let {xk} be the sequence satisfying xk+1 = xk

if∇f (xk) = 0; otherwise,

xk+1 = xk + tkdk , where dk satisfies f ′(xk ; dk) < 0,

and tk is chosen by the backtracking stepsize selection method. Thenone of thefollowing statements must be true:

(i) There is a k0 such that ∇f ′(xk0) = 0.

(ii) f (xk)↘ −∞(iii) The sequence {‖dk‖} diverges (‖dk‖ → ∞).

(iv) For every subsequence J ⊂ N for which {dk : k ∈ J} is bounded, wehave

limk∈J

f ′(xk ; dk) = 0.

Math 408A Convergence of Backtracking Methods

Page 19: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Theorem: Convergence of Backtracking

Let f : Rn → R and x0 ∈ R be such that f is differentiable on Rn with∇f Lipschitz continuous on an open convex set containing the set{x : f (x) ≤ f (x0)}. Let {xk} be the sequence satisfying xk+1 = xk

if∇f (xk) = 0; otherwise,

xk+1 = xk + tkdk , where dk satisfies f ′(xk ; dk) < 0,

and tk is chosen by the backtracking stepsize selection method. Thenone of thefollowing statements must be true:

(i) There is a k0 such that ∇f ′(xk0) = 0.

(ii) f (xk)↘ −∞(iii) The sequence {‖dk‖} diverges (‖dk‖ → ∞).

(iv) For every subsequence J ⊂ N for which {dk : k ∈ J} is bounded, wehave

limk∈J

f ′(xk ; dk) = 0.

Math 408A Convergence of Backtracking Methods

Page 20: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Corollaries

Corollary 1:If the sequences {dk} and {f (xk)} are bounded, then

limk→∞

f ′(xk ; dk) = 0.

Corollary 2: If dk = −∇f ′(xk)/‖∇f (xk)‖ is the Cauchy directionfor all k, then every accumulation point, x, of the sequence {xk}satisfies ∇f (x) = 0.

Math 408A Convergence of Backtracking Methods

Page 21: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Corollaries

Corollary 1:If the sequences {dk} and {f (xk)} are bounded, then

limk→∞

f ′(xk ; dk) = 0.

Corollary 2: If dk = −∇f ′(xk)/‖∇f (xk)‖ is the Cauchy directionfor all k, then every accumulation point, x, of the sequence {xk}satisfies ∇f (x) = 0.

Math 408A Convergence of Backtracking Methods

Page 22: Math 408A Convergence of Backtracking Methodsburke/crs/408/... · Math 408A Convergence of Backtracking Methods. Outline Theorem: Convergence of Backtracking Let f : Rn!R and x 0

Outline

Corollaries

Corollary 3: Let us further assume that f is twice continuouslydifferentiable and that there is a β > 0 such that, for all u ∈ Rn,β‖u‖2 < uT∇2f (x)u on {x : f (x) ≤ f (x0)}. If the BasicBacktracking algorithm is implemented using the Newton searchdirections,

dk = −∇2f (xk)−1∇f (xk),

then every accumulation point, x, of the sequence {xk} satisfies∇f (x) = 0.

Math 408A Convergence of Backtracking Methods