Convexity, Inequalities, and Norms



Similar documents
In nite Sequences. Dr. Philippe B. Laval Kennesaw State University. October 9, 2008

Lecture 13. Lecturer: Jonathan Kelner Scribe: Jonathan Pines (2009)

SAMPLE QUESTIONS FOR FINAL EXAM. (1) (2) (3) (4) Find the following using the definition of the Riemann integral: (2x + 1)dx

Properties of MLE: consistency, asymptotic normality. Fisher information.

Lecture 5: Span, linear independence, bases, and dimension

Trigonometric Form of a Complex Number. The Complex Plane. axis. ( 2, 1) or 2 i FIGURE The absolute value of the complex number z a bi is

Sequences and Series

Chapter 5: Inner Product Spaces

Lecture 4: Cauchy sequences, Bolzano-Weierstrass, and the Squeeze theorem

Asymptotic Growth of Functions

FIBONACCI NUMBERS: AN APPLICATION OF LINEAR ALGEBRA. 1. Powers of a matrix

1. MATHEMATICAL INDUCTION

Theorems About Power Series

Section 11.3: The Integral Test

Metric, Normed, and Topological Spaces

A probabilistic proof of a binomial identity

The following example will help us understand The Sampling Distribution of the Mean. C1 C2 C3 C4 C5 50 miles 84 miles 38 miles 120 miles 48 miles

4.3. The Integral and Comparison Tests

1. C. The formula for the confidence interval for a population mean is: x t, which was

Infinite Sequences and Series

University of California, Los Angeles Department of Statistics. Distributions related to the normal distribution

Class Meeting # 16: The Fourier Transform on R n

Example 2 Find the square root of 0. The only square root of 0 is 0 (since 0 is not positive or negative, so those choices don t exist here).

Chapter 6: Variance, the law of large numbers and the Monte-Carlo method

Chapter 7 Methods of Finding Estimators

FOUNDATIONS OF MATHEMATICS AND PRE-CALCULUS GRADE 10

Our aim is to show that under reasonable assumptions a given 2π-periodic function f can be represented as convergent series

Soving Recurrence Relations

I. Chi-squared Distributions

Lecture 4: Cheeger s Inequality

Approximating Area under a curve with rectangles. To find the area under a curve we approximate the area using rectangles and then use limits to find

3. Greatest Common Divisor - Least Common Multiple

Overview of some probability distributions.

Sampling Distribution And Central Limit Theorem

CHAPTER 7: Central Limit Theorem: CLT for Averages (Means)

INFINITE SERIES KEITH CONRAD

CME 302: NUMERICAL LINEAR ALGEBRA FALL 2005/06 LECTURE 8

Rényi Divergence and L p -affine surface area for convex bodies

Department of Computer Science, University of Otago

WHEN IS THE (CO)SINE OF A RATIONAL ANGLE EQUAL TO A RATIONAL NUMBER?

Convex Bodies of Minimal Volume, Surface Area and Mean Width with Respect to Thin Shells

THE ARITHMETIC OF INTEGERS. - multiplication, exponentiation, division, addition, and subtraction

MARTINGALES AND A BASIC APPLICATION

Basic Elements of Arithmetic Sequences and Series

1 Correlation and Regression Analysis

Normal Distribution.

Math C067 Sampling Distributions

5 Boolean Decision Trees (February 11)

AP Calculus BC 2003 Scoring Guidelines Form B

Building Blocks Problem Related to Harmonic Series

1 The Gaussian channel

CS103A Handout 23 Winter 2002 February 22, 2002 Solving Recurrence Relations

Maximum Likelihood Estimators.

Lecture 3. denote the orthogonal complement of S k. Then. 1 x S k. n. 2 x T Ax = ( ) λ x. with x = 1, we have. i = λ k x 2 = λ k.

Biology 171L Environment and Ecology Lab Lab 2: Descriptive Statistics, Presenting Data and Graphing Relationships

Measures of Spread and Boxplots Discrete Math, Section 9.4

THE REGRESSION MODEL IN MATRIX FORM. For simple linear regression, meaning one predictor, the model is. for i = 1, 2, 3,, n

CS103X: Discrete Structures Homework 4 Solutions

How To Solve An Old Japanese Geometry Problem

3 Basic Definitions of Probability Theory

Chapter 14 Nonparametric Statistics

Case Study. Normal and t Distributions. Density Plot. Normal Distributions


Definition. A variable X that takes on values X 1, X 2, X 3,...X k with respective frequencies f 1, f 2, f 3,...f k has mean

Discrete Mathematics and Probability Theory Spring 2014 Anant Sahai Note 13

Confidence Intervals. CI for a population mean (σ is known and n > 30 or the variable is normally distributed in the.

AP Calculus AB 2006 Scoring Guidelines Form B

Irreducible polynomials with consecutive zero coefficients

Running Time ( 3.1) Analysis of Algorithms. Experimental Studies ( 3.1.1) Limitations of Experiments. Pseudocode ( 3.1.2) Theoretical Analysis

Here are a couple of warnings to my students who may be here to get a copy of what happened on a day that you missed.

GCSE STATISTICS. 4) How to calculate the range: The difference between the biggest number and the smallest number.

A Note on Sums of Greatest (Least) Prime Factors

2-3 The Remainder and Factor Theorems

UC Berkeley Department of Electrical Engineering and Computer Science. EE 126: Probablity and Random Processes. Solutions 9 Spring 2006

Confidence Intervals for One Mean

Notes on exponential generating functions and structures.

S. Tanny MAT 344 Spring be the minimum number of moves required.

Incremental calculation of weighted mean and variance

Week 3 Conditional probabilities, Bayes formula, WEEK 3 page 1 Expected value of a random variable

THE HEIGHT OF q-binary SEARCH TREES

SECTION 1.5 : SUMMATION NOTATION + WORK WITH SEQUENCES

Entropy of bi-capacities

THIN SEQUENCES AND THE GRAM MATRIX PAMELA GORKIN, JOHN E. MCCARTHY, SANDRA POTT, AND BRETT D. WICK

Analysis Notes (only a draft, and the first one!)

PROCEEDINGS OF THE YEREVAN STATE UNIVERSITY AN ALTERNATIVE MODEL FOR BONUS-MALUS SYSTEM

This document contains a collection of formulas and constants useful for SPC chart construction. It assumes you are already familiar with SPC.

Vladimir N. Burkov, Dmitri A. Novikov MODELS AND METHODS OF MULTIPROJECTS MANAGEMENT

.04. This means $1000 is multiplied by 1.02 five times, once for each of the remaining sixmonth

PART TWO. Measure, Integration, and Differentiation

Chapter 7 - Sampling Distributions. 1 Introduction. What is statistics? It consist of three major areas:

5: Introduction to Estimation

Project Deliverables. CS 361, Lecture 28. Outline. Project Deliverables. Administrative. Project Comments

Non-life insurance mathematics. Nils F. Haavardsson, University of Oslo and DNB Skadeforsikring

Permutations, the Parity Theorem, and Determinants

Some Inequalities for p-geominimal Surface Area and Related Results

ON THE DENSE TRAJECTORY OF LASOTA EQUATION

Math 113 HW #11 Solutions

Betting on Football Pools

Virtile Reguli And Radiational Optaprints

Elementary Theory of Russian Roulette

Transcription:

Covexity, Iequalities, ad Norms Covex Fuctios You are probably familiar with the otio of cocavity of fuctios. Give a twicedifferetiable fuctio ϕ: R R, We say that ϕ is covex (or cocave up) if ϕ (x) 0 for all x R. We say that ϕ is cocave (or cocave dow) if ϕ (x) 0 for all x R. For example, a quadratic fuctio ϕ(x) = ax 2 + bx + c is covex if a 0, ad is cocave if a 0. Ufortuately, the defiitios above are ot sufficietly geeral, sice they require ϕ to be twice differetiable. Istead, we will use the followig defiitios: Defiitio: Covex ad Cocave Fuctios Let a < b, ad let ϕ: (a, b) R be a fuctio. 1. We say that ϕ is covex if for all x, y (a, b) ad λ [0, 1]. 2. We say that ϕ is cocave if for all x, y (a, b) ad λ [0, 1]. ϕ ( (1 λ)x + λy ) (1 λ)ϕ(x) + λϕ(y) ϕ ( (1 λ)x + λy ) (1 λ)ϕ(x) + λϕ(y)

Covexity, Iequalities, ad Norms 2 y φ x y, φ y x, φ x Figure 1: For a covex fuctio, every chord lies above the graph. Geometrically, the fuctio λ ( (1 λ)x + λy, (1 λ)ϕ(x) + λϕ(y) ), 0 λ 1 is a parametrizatio of a lie segmet i R 2. This lie segmet has edpoits ( x, ϕ(x) ) ad ( y, ϕ(y) ), ad is therefore a chord of the graph of ϕ (see figure 1). Thus our defiitios of cocave ad covex ca be iterpreted as follows: A fuctio ϕ is covex if every chord lies above the graph of ϕ. A fuctio ϕ is cocave if every chord lies below the graph of ϕ. Aother fudametal geometric property of covex fuctios is that each taget lie lies etirely below the graph of the fuctio. This statemet ca be made precise eve for fuctios that are ot differetiable: Theorem 1 Taget Lies for Covex Fuctios Let ϕ: (a, b) R be a covex fuctio. The for every poit c (a, b), there exists a lie L i R 2 with the followig properties: 1. L passes through the poit ( c, ϕ(c) ). 2. The graph of ϕ lies etirely above L. PROOF See exercise 1.

Covexity, Iequalities, ad Norms 3 Figure 2: A taget lie to y = x at the poit (0, 0). We will refer to ay lie satisfyig the coclusios of the above theorem as a taget lie for ϕ at c. If ϕ is ot differetiable, the the slope of a taget lie may ot be uiquely determied. For example, if ϕ(x) = x, the a taget lie for ϕ at 0 may have ay slope betwee 1 ad 1 (see figure 2). We shall use the existece of taget lies to provide a geometric proof of the cotiuity of covex fuctios: Theorem 2 Cotiuity of Covex Fuctios Every covex fuctio is cotiuous. PROOF Let ϕ: (a, b) R be a covex fuctio, ad let c (a, b). Let L be a liear fuctio whose graph is a taget lie for ϕ at c, ad let P be a piecewiseliear fuctio cosistig of two chords to the graph of ϕ meetig at c (see figure 3). The L ϕ P i a eighborhood of c, ad L(c) = ϕ(c) = P (c). Sice L ad P are cotiuous at c, it follows from the Squeeze Theorem that ϕ is also cotiuous at c. We ow come to oe of the most importat iequalities i aalysis:

Covexity, Iequalities, ad Norms 4 P x φ x L x c Figure 3: Usig the Squeeze Theorem to prove that ϕ(x) is cotiuous at c Theorem 3 Jese s Iequality (Fiite Versio) Let ϕ: (a, b) R be a covex fuctio, where a < b, ad let x 1,..., x (a, b). The ϕ(λ 1 x 1 + + λ x ) λ 1 ϕ(x 1 ) + + λ ϕ(x ) for ay λ 1,..., λ [0, 1] satisfyig λ 1 + + λ = 1. PROOF Let c = λ 1 x 1 + + λ x, ad let L be a liear fuctio whose graph is a taget lie for ϕ at c. Sice λ 1 + + λ = 1, we kow that L(λ 1 x 1 + + λ x ) = λ 1 L(x 1 ) + + λ L(x ). Sice L ϕ ad L(c) = ϕ(c), we coclude that ϕ(c) = L(c) = L(λ 1 x 1 + + λ x ) = λ 1 L(x 1 ) + + λ L(x ) λ 1 ϕ(x 1 ) + + λ ϕ(x ). This statemet ca be geeralized from fiite sums to itegrals. Specifically, we ca replace the poits x 1,..., x by a fuctio f : R, ad we ca replace the weights λ 1,..., λ by a measure µ o for which µ() = 1.

Covexity, Iequalities, ad Norms 5 Theorem 4 Jese s Iequality (Itegral Versio) Let (, µ) be a measure space with µ() = 1. Let ϕ: (a, b) R be a covex fuctio, where a < b, ad let f : (a, b) be a L 1 fuctio. The ( ) ϕ f dµ (ϕ f) dµ PROOF Let c = f dµ, ad let L be a liear fuctio whose graph is a taget lie for ϕ at c. Sice µ() = 1, we kow that L( f dµ) = (L f) dµ. Sice L(c) = ϕ(c) ad L ϕ, this gives ( ) ϕ(c) = L(c) = L f dµ = (L f) dµ (ϕ f) dµ. Meas You are probably aware of the arithmetic mea ad geometric mea of positive umbers: x 1 + + x ad x 1 x. More geerally, we ca defie weighted versios of these meas. Give positive weights λ 1,..., λ satisfyig λ 1 + + λ = 1, the correspodig weighted arithmetic ad geometric meas are λ 1 x 1 + + λ x ad x λ 1 1 x λ. These reduce to the uweighted meas i the case where λ 1 = = λ = 1/. Arithmetic ad geometric meas satisfy a famous iequality, amely that the geometric mea is always less tha or equal to the arithmetic mea. This turs out to be a simple applicatio of Jese s iequality: Theorem 5 AM GM Iequality Let x 1,..., x > 0, ad let λ 1,..., λ [0, 1] so that λ 1 + + λ = 1. The x λ 1 1 x λ λ 1 x 1 + + λ x.

Covexity, Iequalities, ad Norms 6 b a b 2 y e x ab a log a 1 log a log b log b 2 Figure 4: A visual proof that ab < (a + b)/2. PROOF This theorem is equivalet to the covexity of the expoetial fuctio (see figure 4). Specifically, we kow that e λ 1t 1 + λ t λ 1 e t 1 + λ e t for all t 1,..., t R. Substitutig x i = e t i gives the desired result. The followig theorem geeralizes this iequality to arbitrary measure spaces. The proof is essetially the same as the proof of the previous theorem. Theorem 6 Itegral AM GM Iequality Let (, µ) be a measure space with µ() = 1, ad let f : (0, ) be a measurable fuctio. The ( ) exp log f dµ f dµ PROOF Sice the expoetial fuctio is covex, Jese s iequality gives ( ) exp log f dµ exp(log f) dµ = f dµ. By the way, we ca rescale to get a versio of this iequality that applies wheever

Covexity, Iequalities, ad Norms 7 µ() is fiite ad ozero: ( 1 exp µ() ) log f dµ 1 f dµ. µ() Note that the quatity o the right is simply the average value of f o. The quatity o the left ca be thought of as the (cotiuous) geometric mea of f. p-meas There are may importat meas i mathematics ad sciece, beyod just the arithmetic ad geometric meas. For example, the harmoic mea of positive umbers x 1,..., x is 1/x 1 + + 1/x. This mea is used, for example, i calculatig the average resistace of resistors i parallel. For aother example, the Euclidea mea of x 1,..., x is x 2 1 + + x 2. This mea is used to average measuremets take for the stadard deviatio of a radom variable. The AM GM iequality ca be exteded to cover both of these meas. I particular, the iequality 1/x 1 + + 1/x x 1 x x 1 + + x holds for all x 1,..., x (0, ). Both of these meas are examples of p-meas: x 2 1 + + x 2. Defiitio: p-meas Let x 1,..., x > 0. If p R {0}, the p-mea of x 1,..., x is ( x p ) 1 + + x p 1/p. For example: The 2-mea is the same as the Euclidea mea. The 1-mea is the same as the arithmetic mea.

Covexity, Iequalities, ad Norms 8 The ( 1)-mea is the same as the harmoic mea. Though it may ot be obvious, the geometric mea also fits ito the family of p-meas. I particular, it is possible to show that ( x p ) 1 + + x p 1/p lim = p 0 x 1 x for ay x 1,..., x (0, ). Thus, we may thik of the geometric mea as the 0-mea. It is also possible to use limits to defie meas for ad. It turs out the the -mea of x 1,..., x is simply max(x 1,..., x ), while the ( )-mea is mi(x 1,..., x ). As with the arithmetic ad geometric meas, we ca also defie weighted versios of p-meas. Give positive weights λ 1,..., λ satisfyig λ 1 + + λ = 1, the correspodig weighted p-mea is ( ) λ1 x p 1/p. 1 + + λ x p As you may have guessed, the p-meas satisfy a geeralizatio of the AM-GM iequality: Theorem 7 Geeralized Mea Iequality Let x 1,..., x > 0, ad let λ 1,..., λ [0, 1] so that λ 1 + + λ = 1. The p q ( ) λ1 x p 1 + + λ x p 1/p ( λ1 x q 1 + + λ x) q 1/q for all p, q R {0}. PROOF If p = 1 ad q > 1, this iequality takes the form ( λ1 x 1 + + λ x ) q λ1 x q 1 + + λ x q which follows immediately from the covexity of the fuctio ϕ(x) = x q. The case where 0 < p < q follows from this. Specifically, sice q/p > 1, we have ( λ1 (x p 1) + + λ (x p ) ) q/p λ1 (x p 1) q/p + + λ (x p ) q/p ad the desired iequality follows. Cases ivolvig egative values of p or q are left as a exercise to the reader.

Covexity, Iequalities, ad Norms 9 Applyig the same reasoig usig the itegral versio of Jese s iequality gives p q ( f p dµ) 1/p ( ) 1/q f q dµ for ay L 1 fuctio f : (0, ), where (, µ) is a measure space with a total measure of oe. Norms A orm is a fuctio that measures the legths of vectors i a vector space. The most familiar orm is the Euclidea orm o R, which is defied by the formula (x 1,..., x ) = x 2 1 + + x 2. Defiitio: Norm o a Vector Space Let V be a vector space over R. A orm o V is a fuctio : V R, deoted v v, with the followig properties: 1. v 0 for all v V, ad v = 0 if ad oly if v = 0. 2. λv = λ v for all λ R ad v V. 3. v + w v + w for all v, w V. For those familiar with topology, ay orm o a vector space gives a metric d o the vector space defied by the formula d(v, w) = v w. Thus ay vector space with a orm ca be thought of as a topological space. The Euclidea orm o R ca be geeralized to the family of p-orms. ay p 1, the p-orm o R is defied by the formula For (x 1,..., x ) p = ( x 1 p + + x p ) 1/p The usual Euclidea orm correspods to the case where p = 2. It is easy to see that the defiitio of the p-orm satisfies axioms (1) ad (2) for a orm, but third axiom (which is kow as the triagle iequality) is far from clear. The followig theorem establishes the p-orm is i fact a orm.

Covexity, Iequalities, ad Norms 10 Theorem 8 Mikowski s Iequality If u, v R ad p [1, ), the u + v p u p + v p PROOF Sice p 1, the fuctio x x p is covex. It follows that (1 λ)u + λv p p = (1 λ)u i + λv i p i=1 (1 λ) u i p + λ v i p = (1 λ) u p p + λ v p p i=1 for all u ad v ad λ [0, 1]. I particular, this proves that (1 λ)u + λv p 1 wheever u p = v p = 1. From this Mikowski s iequality follows. I particular, we may assume that u ad v are ozero. The u/ u p ad v/ v p are uit vectors, so u + v p u p + v p = u p u p + v p u u p + v p u p + v p v v p 1. p If is a orm o a vector space V, the uit ball i V is the set B V (1) = { v V v 1 }. For example, the uit ball i R 2 with respect to the Euclidea orm is a roud disk of radius 2 cetered at the origi. Figure 5 shows the uit ball i R 2 with respect to various p-orms. Their shapes are all fairly similar, with the the uit ball beig a diagoal square whe p = 1, ad the uit ball approachig a horizotal square as p. All of these uit balls have a certai importat geometric property. Defiitio: Covex Set Let V is a vector space over R. If v ad w are poits i V, the lie segmet from v to w is the set L(v, w) = {λv + (1 λ)w 0 λ 1}. A subset S V is covex if L(v, w) S for all v, w S. The followig theorem gives a geometric iterpretatio of Mikowski s iequality.

Covexity, Iequalities, ad Norms 11 p = 1 p = 1.5 p = 2 p = 3 p = 10 Figure 5: The shape of the uit ball i R 2 for various p-orms. Theorem 9 Shapes of Uit Balls Let V be a vector space over R, ad let : V R be a fuctio satisfyig coditios (1) ad (2) for a orm. The satisfies the triagle iequality if ad oly if the uit ball { v V v 1 } is covex. PROOF Suppose first that satisfies the triagle iequality. Let v ad w be poits i the uit ball, ad let p = λv + (1 λ)w be ay poit o the lie segmet from v to w. The p = λv + (1 λ)w λv + (1 λ)w = λ v + (1 λ) w λ(1) + (1 λ)(1) = 1 so p lies i the uit ball as well. For the coverse, suppose that the uit ball is covex, ad let v, w V. If v or w is 0, the clearly v + w v + w, so suppose they are both ozero. Let ˆv = v/ v ad ŵ = w/ w. The ˆv = 1 v v = 1 v v = 1 ad similarly ŵ = 1. Thus ˆv ad ŵ both lie i the uit ball. Sice the poit v v + w + w v + w = 1, v + w v + w = v v + w ˆv + w v + w ŵ must lie i the uit ball as well, ad it follows that v + w v + w.

Covexity, Iequalities, ad Norms 12 Fially, we should metio that there is a itegral versio of Mikowski s iequality. This ivolves the otio of a p-orm o a measure space. Defiitio: p-norm Let (, µ) be a measure space, let f be a measurable fuctio o, ad let p [1, ). The p-orm of f is the quatity f p = ( ) 1/p f p dµ Note that f p is always defied, sice it ivolves the itegral of a o-egative fuctio, but it may be ifiite. We ca ow state Mikowski s iequality for arbitrary measure spaces. Theorem 10 Mikowski s Iequality (Itegral Versio) Let (, µ) be a measure space, let f, g : R be measurable fuctios, ad let p [1, ). The f + g p f p + g p. PROOF Sice p 1, the fuctio x x p is covex. It follows that (1 λ)f + λg p p = (1 λ)f + λg p dµ ( (1 λ) f p + λ g p) dµ = (1 λ) f p p + λ g p p for ay measurable fuctios f ad g ad ay λ [0, 1]. I particular, this proves that (1 λ)f + λg p 1 wheever f p = g p = 1. From this Mikowski s iequality follows. First, observe that if f p = 0, the f = 0 almost everywhere, so Mikowski s iequality follows i this case. A similar argumet holds if g p = 0, so suppose that f p > 0 ad g p > 0. Let ˆf = f/ f p ad ĝ = g/ g p, ad ote that ˆf p = ĝ p = 1. The f + g p f p + g p = f p g p ˆf + ĝ f p + g p f p + g p 1. p

Covexity, Iequalities, ad Norms 13 Hölder s Iequality Recall that the Euclidea orm o R satisfies the Cauchy-Schwarz Iequality u v u 2 v 2 for all u, v R. Our ext task is to prove a geeralizatio of this kow as Hölder s iequality. Lemma 11 Youg s Iequality If x, y [0, ) ad p, q (1, ) so that 1/p + 1/q = 1, the xy xp p + yq q PROOF This ca be writte (x p ) 1/p (y q ) 1/q 1 p xp + 1 q yq which is a istace of the weighted AM GM iequality. Theorem 12 Hölder s Iequality Let u, v R, ad let p, q (1, ) so that 1/p + 1/q = 1. The u v u p v q. PROOF By Youg s iequality, u v u 1 v 1 + + u v u 1 p + + u p p + v 1 q + + v q q u p p p + v q q q I particular, if u p = v q = 1, the u v 1/p + 1/q = 1, which proves Hölder s iequality i this case. For the geeral case, we may assume that u ad v are ozero. The u/ u p ad v/ v q are uit vectors for their respective orms, ad therefore u v u p v q 1.

Covexity, Iequalities, ad Norms 14 Multiplyig through by u p v q gives the desired result. This iequality is ot hard to geeralize to itegrals. We begi with the followig defiitio. Defiitio: Ier Product of Fuctios Let (, µ) be a measure space, ad let f ad g be measurable fuctios o. The L 2 ier product of f ad g is defied as follows: f, g = fg dµ. Note that f, g may be udefied if fg is ot Lebesgue itegrable o. Note also that f, g = fg 1 if f ad g are oegative, but that i geeral f, g fg 1. Theorem 13 Hölder s Iequality (Itegral Versio) Let (, µ) be a measure space, let f ad g be measurable fuctios o, ad let p, q (1, ) so that 1/p + 1/q = 1. If f p < ad g p <, the f, g is defied, ad f, g f p g q PROOF Suppose first that f p = g q = 1. By Youg s iequality, fg 1 = fg dµ ( f p p ) + g q dµ f p p + g q q q p q = 1. Sice fg is L 1, it follows that f, g is defied. The f, g fg 1 1, which proves Hölder s iequality i this case. For the geeral case, ote first that if either f p = 0 or g q = 0, the either f = 0 almost everywhere or g = 0 almost everywhere, so Hölder s iequality holds i that case. Otherwise, let ˆf = f/ f p ad ĝ = g/ g q. The ˆf p = ĝ q = 1, so f, g = f p g p ˆf, ĝ f p g q

Covexity, Iequalities, ad Norms 15 I the case where p = q = 2, this gives the followig. Corollary 14 Cauchy-Schwarz Iequality (Itegral Versio) Let (, µ) be a measure space, ad let f ad g be measurable fuctios o. If f 2 < ad g 2 <, the f, g is defied, ad f, g f 2 g 2. Exercises 1. a) Prove that a fuctio ϕ: R R is covex if ad oly if ϕ(y) ϕ(x) y x for all x, y, z R with x < y < z. ϕ(z) ϕ(x) z x ϕ(z) ϕ(y) z y b) Use this characterizatio of covex fuctios to prove Theorem 1 o the existece of taget lies. 2. Let ϕ: (a, b) R be a differetiable fuctio. Prove that ϕ is covex if ad oly if ϕ is o-decreasig. 3. Let ϕ: R R be a covex fuctio. Prove that ϕ ( (1 λ)x + λy ) (1 λ)ϕ(x) + λϕ(y) for λ R [0, 1]. Use this to provide a alterative proof that ϕ is cotiuous. 4. If x, y 0, prove that ( x p + y p lim p 0 What is 2 lim p ) 1/p = xy ad lim p ( ) x p + y p 1/p? 2 ( ) x p + y p 1/p = max(x, y). 5. a) If x 1,..., x, y 1,..., y (0, ), prove that x1 y 1 + x y x1 + + x y1 + + y. That is, the arithmetic mea of geometric meas is less tha or equal to the correspodig geometric mea of arithmetic meas. 2

Covexity, Iequalities, ad Norms 16 b) If λ, µ [0, 1] ad λ + µ = 1, prove that x λ 1y µ 1 + x λ y µ ( ) λ ( ) µ x1 + + x y1 + + y. 6. Prove the Geeralized Mea Iequality (Theorem 6) i the case where p or q is egative. 7. Let f : [0, 1] R be a bouded measurable fuctio, ad defie ϕ: [1, ) R by ϕ(p) = f p dµ. Prove that log ϕ is covex o [1, ). [0,1] 8. Prove that ( 1 + x 2 y + x 4 y 2) 3 ( 1 + x 3 + x 6) 2( 1 + y 3 + y 6) for all x, y (0, ). 9. Let (, µ) be a measure space, let f, g : [0, ) be measurable fuctios, ad let p, q, r (1, ) so that 1/p + 1/q = 1/r. Prove that fg r f p g q.