Jensen’s inequality for non-convex functions Holger STEPHAN 1 Jensen’s inequality is well known: Let P f (x) be a convex function on x ∈ I ⊂ R, x1 , ..., xn ∈ I and qi ≥ 0 are weights with i qi = 1. Then, Jensen’s inequality n X i=1 qi f (xi ) ≥ f n ³X qi xi i=1 ´ holds. This inequality is in some sense equivalent to the definition of convexity. That’s why, it is not well known that Jensen’s inequality holds even for non-convex functions (moreover, this is believed to be false). √ Let’s start with an easy contest problem (from the journal Die wurzel, Jena, 05/2005, problem µ19): Let x1 ,. . . , xn be positive real numbers with 0.075 h(x) 0.05 x1 +· · ·+xn = n. Prove the following inequality. n X i=1 0.025 n X 2 1 ≥ 1 + xi 3 + xi i=1 (1) We use Jensen’s inequality with the function h(x) = x 10 20 30 40 -0.025 -0.05 -0.075 2 1−x 1 − = . 1+x 3+x (1 + x)(3 + x) and the weights qi = n1 . We get ¶ X n µ n n ³1 X ´ 2 1 1 1X − xi = h(1) = 0 = h(xi ) ≥ h n i=1 1 + xi 3 + xi n n i=1 i=1 and therefore √ (1). The √ result is true but the proof is false, because h(x) is not convex for x > 1 + 3 2 + 3 4 = 6.69464 . . . (see the picture). Why does the inequality holds, nevertheless? To investigate this question, we try to prove Jensen’s inequality n P deriving an identity for it. We fix x0 = qi xi and define a function i=1 g(x) = f (x) − f (x0 ) x − x0 (for x = x0 we define g(x0 ) = f 0 (x0 )). Now, we set n n n ³X ´ X X X= qi f (xi ) − f qi xi = qi f (xi ) − f (x0 ) i=1 i=1 i=1 and check, when does X ≥ 0 hold. Using easy calculations, we obtain 1 Cercetător dr., Institutul Weierstrass, Berlin (e-mail: [email protected]) 109 (2) n X n n X ¢ X ¡ f (xi ) − f (x0 ) X= qi f (xi ) − f (x0 ) = qi f (xi ) − f (x0 ) = qi (xi − x0 ) = xi − x0 i=1 i=1 i=1 = = n X qi g(xi ) (xi − x0 ) = i=1 n X i,j=1 = qi qj xi g(xi ) − X 1≤j<i≤n n X n X i=1 n n ³ X ´ X g(xi )qi xi qj − qj xj = qi qj xj g(xi ) = i,j=1 ¡ ¢ qi qj (xi − xj ) g(xi ) − g(xj ) . Thus, we get the identity n n ³X ´ X X= qi f (xi ) − f qi xi = i=1 j=1 n X i,j=1 X i=1 j=1 ¢ ¡ qi qj xi g(xi ) − xj g(xi ) = 1≤j<i≤n ¡ ¢ qi qj (xi − xj ) g(xi ) − g(xj ) (3) We see that X ≥ 0 holds, if the function g(x) is monotone increasing. The function g(x) is called slope function. g(x) is the slope of the secant through the points in x0 and x. If f (x) in convex, g(x) is increasing for any point x0 . But this is not the only case. Function g(x) is increasing, if looking from the point (x0 , f (x0 )) on the graph of f (x), no point of f (x) "lies in the shadow" of the graph. This can happen for some points x0 even if the function is not convex. The function h(x) above is an example (here is x0 = 1). An other amazing example are polynomials of fourth order — typical non-convex functions. We e1 f(x) consider such a function f (x) and its inflection points (the x-coordinates are x1 and x2 ). The e2 inflection point tangents e1 and e2 intersect the graph of f (x) in points with x-coordinates x3 and x4 . Now, Jensen’s inequality holds for x0 ≥ x3 or x x3 x0 ≤ x4 . x4 x2 x1 The typical proof of Jensen’s inequality starts from the inequality for convex functions (xi − x0 )f 0 (x0 ) ≤ f (xi ) − f (x0 ) ≤ (xi − x0 )f 0 (xi ) Multiplying by qi and adding up over i we get n n n n X X X X qi (xi − x0 )f 0 (x0 ) ≤ qi f (xi ) − f (x0 ) qi ≤ qi (xi − x0 )f 0 (xi ) i=1 i=1 The left hand side in zero because of i=1 n P qi = 1 and i=1 i=1 n P qi xi = x0 . The middle is X i=1 and transforming the right hand sind in a similar way as above (3) we obtain. n n ³X ´ X X ¡ ¢ 0≤ qi f (xi ) − f qi xi ≤ qi qj (xi − xj ) f 0 (xi ) − f 0 (xj ) i=1 i=1 1≤j<i≤n This is similar to identity (3), but holds only for convex functions. Moreover, identity (3) is very useful, if we are not only interested in Jensen’s inequality, but if we want to estimate the difference on the left hand side of (3). 110
© Copyright 2025 Paperzz