minimization problem with smooth components
play

Minimization Problem with Smooth Components Yu. Nesterov Presenter: - PowerPoint PPT Presentation

Minimization Problem with Smooth Components Yu. Nesterov Presenter: Lei Tang Department of CSE Arizona State University Dec. 7th, 2008 1 / 39 Outline MiniMax problem Gradient Mapping for MiniMax problem ; The complexity of gradient and


  1. Minimization Problem with Smooth Components Yu. Nesterov Presenter: Lei Tang Department of CSE Arizona State University Dec. 7th, 2008 1 / 39

  2. Outline MiniMax problem Gradient Mapping for MiniMax problem ; The complexity of gradient and optimal method; Optimization with functional constraint (General constrained optimization problem) Constrained Minimization Problem 2 / 39

  3. MiniMax Problem Objective function is composed with several components. The simplest problem of that type is minimax problem. We’ll focus on smooth minimax problem: � � min f ( x ) = max 1 ≤ i ≤ m f i ( x ) x ∈ Q where f i ∈ S 1 , 1 µ, L ( R n ) , i = 1 , · · · , m and Q is a closed convex set. f ( x ): the max-type function composed by the components f i ( x ). In general, f ( x ) is not differentiable. We use f ∈ S 1 , 1 µ, L ( R n ) to denote all the f i ∈ S 1 , 1 µ, L ( R n ). 3 / 39

  4. MiniMax Problem Objective function is composed with several components. The simplest problem of that type is minimax problem. We’ll focus on smooth minimax problem: � � min f ( x ) = max 1 ≤ i ≤ m f i ( x ) x ∈ Q where f i ∈ S 1 , 1 µ, L ( R n ) , i = 1 , · · · , m and Q is a closed convex set. f ( x ): the max-type function composed by the components f i ( x ). In general, f ( x ) is not differentiable. We use f ∈ S 1 , 1 µ, L ( R n ) to denote all the f i ∈ S 1 , 1 µ, L ( R n ). 3 / 39

  5. Connection with General Minimization Problem General Minimization Problem min f 0 ( x ) (1) f i ( x ) ≤ 0 , i = 1 , · · · , m (2) s . t . x ∈ Q (3) parametric max-type function f ( t ; x ) = max { f 0 ( x ) − t ; f i ( x ) } Will be showed later: the optimal value of f 0 ( x ) corresponds to the root t of f ( t ; x ) = 0; minimax problem is used as a subroutine to solve (1); 4 / 39

  6. Linear approximation Linearization max-type function f ( x ) = max 1 ≤ i ≤ m f i ( x ) � � f ′ linearization of f ( x ) f (¯ x ; x ) = max 1 ≤ i ≤ m [ f i (¯ x ) + i (¯ x ) , x − ¯ ] x Essentially, linearization over each component. Properties x || 2 ≤ f ( x ) ≤ f (¯ x ; x ) + µ x ; x ) + L x || 2 ; f (¯ 2 || x − ¯ 2 || x − ¯ x ∗ ∈ Q ⇔ f ( x ∗ ; x ) ≥ f ( x ∗ ; x ∗ ) = f ( x ∗ ). f ( x ) ≥ f ( x ∗ ) + µ 2 || x − x ∗ || 2 the solution x ∗ exists and unique. 5 / 39

  7. Lemma 2.3.1 x || 2 ≤ f ( x ) ≤ f (¯ x ; x ) + µ x ; x ) + L x || 2 f (¯ 2 || x − ¯ 2 || x − ¯ f i ∈ S 1 , 1 µ, L ( R n ) For strongly convex function, we have x ) � + µ x ) + � f ′ x || 2 f i ( x ) ≥ f i (¯ i (¯ x , x − ¯ 2 || x − ¯ x ; x ) + µ x || 2 = f (¯ 2 || x − ¯ x ; x ) + µ x || 2 Take the max on both sides: f ( x ) ≥ f (¯ 2 || x − ¯ For Lipshitz continuous function, it follows x ) � + L x ) + � f ′ x || 2 f i ( x ) ≤ f i (¯ i (¯ x , x − ¯ 2 || x − ¯ x ; x ) + L x || 2 = f (¯ 2 || x − ¯ max operation keeps the property as smooth strongly convex function. 6 / 39

  8. Lemma 2.3.1 x || 2 ≤ f ( x ) ≤ f (¯ x ; x ) + µ x ; x ) + L x || 2 f (¯ 2 || x − ¯ 2 || x − ¯ f i ∈ S 1 , 1 µ, L ( R n ) For strongly convex function, we have x ) � + µ x ) + � f ′ x || 2 f i ( x ) ≥ f i (¯ i (¯ x , x − ¯ 2 || x − ¯ x ; x ) + µ x || 2 = f (¯ 2 || x − ¯ x ; x ) + µ x || 2 Take the max on both sides: f ( x ) ≥ f (¯ 2 || x − ¯ For Lipshitz continuous function, it follows x ) � + L x ) + � f ′ x || 2 f i ( x ) ≤ f i (¯ i (¯ x , x − ¯ 2 || x − ¯ x ; x ) + L x || 2 = f (¯ 2 || x − ¯ max operation keeps the property as smooth strongly convex function. 6 / 39

  9. Lemma 2.3.1 x || 2 ≤ f ( x ) ≤ f (¯ x ; x ) + µ x ; x ) + L x || 2 f (¯ 2 || x − ¯ 2 || x − ¯ f i ∈ S 1 , 1 µ, L ( R n ) For strongly convex function, we have x ) � + µ x ) + � f ′ x || 2 f i ( x ) ≥ f i (¯ i (¯ x , x − ¯ 2 || x − ¯ x ; x ) + µ x || 2 = f (¯ 2 || x − ¯ x ; x ) + µ x || 2 Take the max on both sides: f ( x ) ≥ f (¯ 2 || x − ¯ For Lipshitz continuous function, it follows x ) � + L x ) + � f ′ x || 2 f i ( x ) ≤ f i (¯ i (¯ x , x − ¯ 2 || x − ¯ x ; x ) + L x || 2 = f (¯ 2 || x − ¯ max operation keeps the property as smooth strongly convex function. 6 / 39

  10. Theorem 2.3.1: x ∗ ∈ Q ⇔ f ( x ∗ ; x ) ≥ f ( x ∗ ; x ∗ ) = f ( x ∗ ) x ; x ) + µ x || 2 , we have ⇐ As f ( x ) ≥ f (¯ 2 || x − ¯ f ( x ) ≥ f ( x ∗ ; x ) + µ 2 || x − x ∗ || 2 ≥ f ( x ∗ ; x ∗ ) + 0 = f ( x ∗ ) ⇒ Prove by contradiction: if f ( x ∗ ; x ) < f ( x ∗ ), then for 1 ≤ i ≤ m f i ( x ∗ ) + � f ′ (¯ x ; x ∗ ) , x − x ∗ � < f ( x ∗ ) = max 1 ≤ i ≤ m f i ( x ∗ ) Define φ i ( α ) = f i ( x ∗ + α ( x − x ∗ )) , α ∈ [0 , 1] So either φ i (0) ≡ f i ( x ∗ ) < f ( x ∗ ) or φ i (0) = f ( x ∗ ) , φ ′ i (0) = � f ′ i ( x ∗ ) , x − x ∗ � < 0 So small enough α , f i ( x ∗ + α ( x − x ∗ )) = φ i ( α ) < f ( x ∗ ) ∀ 1 ≤ i ≤ m contradiction! Linearization achieves its minimum at x ∗ 7 / 39

  11. Theorem 2.3.1: x ∗ ∈ Q ⇔ f ( x ∗ ; x ) ≥ f ( x ∗ ; x ∗ ) = f ( x ∗ ) x ; x ) + µ x || 2 , we have ⇐ As f ( x ) ≥ f (¯ 2 || x − ¯ f ( x ) ≥ f ( x ∗ ; x ) + µ 2 || x − x ∗ || 2 ≥ f ( x ∗ ; x ∗ ) + 0 = f ( x ∗ ) ⇒ Prove by contradiction: if f ( x ∗ ; x ) < f ( x ∗ ), then for 1 ≤ i ≤ m f i ( x ∗ ) + � f ′ (¯ x ; x ∗ ) , x − x ∗ � < f ( x ∗ ) = max 1 ≤ i ≤ m f i ( x ∗ ) Define φ i ( α ) = f i ( x ∗ + α ( x − x ∗ )) , α ∈ [0 , 1] So either φ i (0) ≡ f i ( x ∗ ) < f ( x ∗ ) or φ i (0) = f ( x ∗ ) , φ ′ i (0) = � f ′ i ( x ∗ ) , x − x ∗ � < 0 So small enough α , f i ( x ∗ + α ( x − x ∗ )) = φ i ( α ) < f ( x ∗ ) ∀ 1 ≤ i ≤ m contradiction! Linearization achieves its minimum at x ∗ 7 / 39

  12. Theorem 2.3.1: x ∗ ∈ Q ⇔ f ( x ∗ ; x ) ≥ f ( x ∗ ; x ∗ ) = f ( x ∗ ) x ; x ) + µ x || 2 , we have ⇐ As f ( x ) ≥ f (¯ 2 || x − ¯ f ( x ) ≥ f ( x ∗ ; x ) + µ 2 || x − x ∗ || 2 ≥ f ( x ∗ ; x ∗ ) + 0 = f ( x ∗ ) ⇒ Prove by contradiction: if f ( x ∗ ; x ) < f ( x ∗ ), then for 1 ≤ i ≤ m f i ( x ∗ ) + � f ′ (¯ x ; x ∗ ) , x − x ∗ � < f ( x ∗ ) = max 1 ≤ i ≤ m f i ( x ∗ ) Define φ i ( α ) = f i ( x ∗ + α ( x − x ∗ )) , α ∈ [0 , 1] So either φ i (0) ≡ f i ( x ∗ ) < f ( x ∗ ) or φ i (0) = f ( x ∗ ) , φ ′ i (0) = � f ′ i ( x ∗ ) , x − x ∗ � < 0 So small enough α , f i ( x ∗ + α ( x − x ∗ )) = φ i ( α ) < f ( x ∗ ) ∀ 1 ≤ i ≤ m contradiction! Linearization achieves its minimum at x ∗ 7 / 39

  13. Theorem 2.3.1: x ∗ ∈ Q ⇔ f ( x ∗ ; x ) ≥ f ( x ∗ ; x ∗ ) = f ( x ∗ ) x ; x ) + µ x || 2 , we have ⇐ As f ( x ) ≥ f (¯ 2 || x − ¯ f ( x ) ≥ f ( x ∗ ; x ) + µ 2 || x − x ∗ || 2 ≥ f ( x ∗ ; x ∗ ) + 0 = f ( x ∗ ) ⇒ Prove by contradiction: if f ( x ∗ ; x ) < f ( x ∗ ), then for 1 ≤ i ≤ m f i ( x ∗ ) + � f ′ (¯ x ; x ∗ ) , x − x ∗ � < f ( x ∗ ) = max 1 ≤ i ≤ m f i ( x ∗ ) Define φ i ( α ) = f i ( x ∗ + α ( x − x ∗ )) , α ∈ [0 , 1] So either φ i (0) ≡ f i ( x ∗ ) < f ( x ∗ ) or φ i (0) = f ( x ∗ ) , φ ′ i (0) = � f ′ i ( x ∗ ) , x − x ∗ � < 0 So small enough α , f i ( x ∗ + α ( x − x ∗ )) = φ i ( α ) < f ( x ∗ ) ∀ 1 ≤ i ≤ m contradiction! Linearization achieves its minimum at x ∗ 7 / 39

  14. Corollary 2.3.1 f ( x ) ≥ f ( x ∗ ) + µ 2 || x − x ∗ || 2 x ; x ) + µ x || 2 f ( x ) ≥ f (¯ 2 || x − ¯ f ( x ∗ ; x ) + µ 2 || x − x ∗ || 2 ≥ f ( x ∗ ; x ∗ ) + µ 2 || x − x ∗ || 2 ≥ f ( x ∗ ) + µ 2 || x − x ∗ || 2 = So if x ∗ exists, it must be unique. 8 / 39

  15. Corollary 2.3.1 f ( x ) ≥ f ( x ∗ ) + µ 2 || x − x ∗ || 2 x ; x ) + µ x || 2 f ( x ) ≥ f (¯ 2 || x − ¯ f ( x ∗ ; x ) + µ 2 || x − x ∗ || 2 ≥ f ( x ∗ ; x ∗ ) + µ 2 || x − x ∗ || 2 ≥ f ( x ∗ ) + µ 2 || x − x ∗ || 2 = So if x ∗ exists, it must be unique. 8 / 39

  16. Theorem 3.2 Let a max-type function f ( x ) ∈ S 1 µ ( R n ), µ > 0, and Q be a closed convex set. Then the solution x ∗ exists and unique. x ∈ Q , consider the set ¯ Let ¯ Q = { x ∈ Q | f ( x ) ≤ f (¯ x ) } . Transform to a problem as min { f ( x ) | x ∈ ¯ Q } Need to show ¯ Q is bounded. x � + µ x ) + � f ′ (¯ x || 2 f (¯ x ) ≥ f i ( x ) ≥ f i (¯ x ) , x − ¯ 2 || x − ¯ µ x || 2 ≤ || f ′ (¯ = ⇒ 2 || x − ¯ x ) || · || x − ¯ x || + f (¯ x ) − f i (¯ x ) So the solution x ∗ exists and is unique 9 / 39

  17. Quick Summary MiniMax, though generally not smooth, share all the properties as minimizing smooth strongly convex functions over simple convex set. Linearization max-type function f ( x ) = max 1 ≤ i ≤ m f i ( x ) � � f ′ linearization of f ( x ) f (¯ x ; x ) = max 1 ≤ i ≤ m [ f i (¯ x ) + i (¯ x ) , x − ¯ x ] Essentially, linearization over each component. Properties x || 2 ≤ f ( x ) ≤ f (¯ x ; x ) + µ x ; x ) + L x || 2 ; f (¯ 2 || x − ¯ 2 || x − ¯ x ∗ ∈ Q ⇔ f ( x ∗ ; x ) ≥ f ( x ∗ ; x ∗ ) = f ( x ∗ ). f ( x ) ≥ f ( x ∗ ) + µ 2 || x − x ∗ || 2 the solution x ∗ exists and unique. 10 / 39

Recommend


More recommend