Problem Set Two: Some Examples of Norms Definition: Let V be a real vector space. A norm on V is a function : V Reals so that x 0 iff x 0 these four properties hold for all x, z in V and scalars c. (a) (c) cx c x (b) x 0 xz x z . (d, Triangle Inequality) For example, absolute value is a norm on the reals. Theorem: If is a norm on V then d(x, z) x z defines a metric. Definition: Let V be a real vector space. An inner product on V is a function , : V V Reals so that these five properties hold for all x, y, z in V and scalars c. (a) x, x 0 iff x 0 (b) x, x 0 (c) x, z z, x (d) c x, z c x, z (e) x y, z x, z y, z . An inner product must also be linear in its second argument, i.e., (d') x, c z c x, z (e') z, x y z, x z, y also hold. Properties (b) - (e) define a positive, symmetric bilinear form. Theorem: Let , be a positive, symmetric bilinear form on V. Write x (a, Cauchy-Schwartz Inequality) For any x and z in V, x, z x (b) For all x, z in V and scalars c, x 0 , c x c (c) For an inner product x Corollary: Let x x , and x, x . z . xz x z . x, x defines a norm on V. x, x be an inner product norm on V. For any x in V x sup{ x, z : z 1 } . If x 0 the supremum is attained at z x 1 x . Example: R n , the set of all n-tuples of reals, is a vector space under the coordinate-wise operations. The standard inner product on R n is x, z nk 1 x(k) z(k) , with associated Euclidean norm, or 2-norm, x 2 n k 1 x(k) 2 . Example: R[a,b], the set of bounded, Riemann integrable functions on [a,b], is a vector space under the pointwise operations. Defining b f, g a f(x) g(x) d x yields a positive, symmetric bilinear form. The associated expression f 2 b a f (x) 2 dx is usually called the 2-norm of f, even though it's not really a norm on R[a,b]. b Theorem: If f R [a, b] and a f (x) 2 d x 0 then f(z) = 0 at each point of continuity of f. Corollary: C[a,b], the set of continuous real valued function on [a,b], is a vector space under the pointwise operations. The 2-norm is a norm on C[a,b]. PROBLEMS Problem 2-1: For x, y and z points in a metric space, d(x, z) d(y, z) d(x, y) . For (S,d) a metric space and T a subset of S, the restriction of d to T T is clearly a metric on T, called the inherited metric on T. The pair (T,d) is called a subspace of (S,d). For instance { 0 ,1} N is a subspace of [ 0 ,1 ] N . Another way to generate metric spaces is by taking products. Problem 2-2: If (X,d) and ( Y , ) are metric spaces then setting (d ρ) (x, y), (u, v) max { d(x, u), ρ(y, v) } defines a metric on the product X Y . ( X Y, d ρ ) is the product of the metric spaces (X,d) and ( Y , ) . The next problems define two of the p-norms x p on R n . The metric space ( R n , x p ) is denoted by l np . The 2-norm is the Euclidean norm. Problem 2-3: For x ( x(1), x(2), ... , x(n) ) define x (a) x max { x(1) , x(2) , ... , x(n) } . is a norm on R n . (b) The metric determined by this norm is called the box metric. To see why sketch the closed balls CB(0,1) in R 2 and R 3 . (c) l 2 is the product of two copies of R, each with the usual metric. Problem 2-4: For x ( x(1), x(2), ... , x(n) ) define x (a) x 1 is a norm on R n . 1 n k 1 x(k) . (b) Sketch the closed balls CB(0,1) in R 2 and R 3 . (c) On the subset S { 0 , 1 } n the inherited metric is the Hamming Metric H(x, y) card { k : x(k) y(k) } Problem 2-5: (a, Parallelogram Law) In the 2-norm x y 2 xy 2 2 2 x 2 y 2 for any two vectors x and y in R n . (b) Show by example that the parallelogram law fails for both x and x 1 . Problem 2-6: Fix an n-tuple x and consider the minimum of f(t) x (t, t, ... , t) p . (a) For p = 2 the minimum occurs when t 0 n 1 nk 1 x(k) . What is n 1/2 f(t 0 ) called? (b) For p = 1 the minimum occurs at each median of x(1), x(2), ... , x(n). (c) Find the minimum and the point(s) at which it occurs in case p . Problem 2-7: A set E in a vector space V is convex iff E contains the line segment between any two points in E, equivalently, iff t x ( 1 t) y E for all x and y in E and scalars t [0,1] . (a) If the metric is determined by a norm then all open and closed balls are convex. (b) Give an example of a metric on a vector space and an open ball that is not convex. OTHER p-NORMS Lemma: Fix p > 1 and let q satisfy 1 / p 1 / q 1 (q is the conjugate of p). For a and b nonnegative, a b inf { (t a) p /p ( t 1 b) q /q : 0 t } . Sketch Proof: Write f(t) (t a) p /p ( t 1 b) q /q. If a = 0 then inf f lim t ( t 1 b) q /q 0 and similarly when b = 0. For positive a and b, f has a minimum when f (t) 0 . The derivative is zero for t 0 b 1/p a 1/q and f(t 0 ) a b . For p > 1 and x ( x(1), x(2), ... , x(n) ) in R n the p-norm of x is x p n k 1 p 1/p x(k) . The sign of a real number t is sgn(t) = 1 if t 0 , and sgn(t) 1 otherwise. Theorem: Fix p > 1 and let q satisfy 1 / p 1 / q 1 . (a, Holder's Inequality) For x and z in R n , x, z x (b) For each x in R n , x p sup{ x, z : z p/q the vector with coordinates z(k) x z p q . 1 } . If x 0 the supremum is attained at q sgn x(k) p 1 x(k) . (c) The p-norm is a norm on R n . The p-norm’s triangle inequality is the Minkowski Inequality. Sketch Proof: First use the lemma in each coordinate. For t > 0 and any k x(k) z(k) (1/p) t p x(k) p (1/q) t q z(k) q . Summing the last inequalities, x, z nk 1 x(k) z(k) p (1/p) t nk 1 x(k) (1/p) t p x p p (1/q) t (1/q) t q z q n k 1 z(k) q q Using the lemma again to minimize the last expression establishes (a). Parts (b) and (c) follow by arguing as with the 2-norm.