Appendices · Appendices

The inequalities of Hölder and Minkowski

Appendix I 45 min

Appendix I: the theorem of the means with real weights, and the two inequalities that turn (∑ak)1/k\left(\sum a^k\right)^{1/k} into a length.

Why this matters

The ordinary distance in the plane, x2+y2\sqrt{x^2 + y^2}, is one member of a family: for any k≥1k \ge 1 the quantity (∣x∣k+∣y∣k)1/k(|x|^k + |y|^k)^{1/k}, or with nn coordinates (∑∣am∣k)1/k\left(\sum |a_m|^k\right)^{1/k}, also behaves like a length. That it does is Minkowski’s inequality, and the tool that proves it is Hölder’s inequality, a far-reaching generalisation of Cauchy’s inequality (∑ab)2≤∑a2∑b2\left(\sum ab\right)^2 \le \sum a^2 \sum b^2. Both rest on one fact, the theorem of the arithmetic and geometric means with arbitrary positive real weights. Appendix I proves all three in a few pages. This lesson supplies the pictures.

(The appendix has no section numbers. Hardy first proves inequalities of this kind, for rational exponents, in §74, and points forward to Appendix I from there.)

The idea

Everything starts from the two-number case. If a,b>0a, b > 0 and the weights α,β>0\alpha, \beta > 0 add up to 11, then

aαbβ≤αa+βb,a^{\alpha} b^{\beta} \le \alpha a + \beta b,

with equality only when a=ba = b. With α=β=12\alpha = \beta = \tfrac12 this is the familiar ab≤12(a+b)\sqrt{ab} \le \tfrac12(a + b). Here is a way to see the general case, due to W. H. Young. Put α=1/k\alpha = 1/k and β=1/k′\beta = 1/k', where k>1k > 1 and k′=k/(k−1)k' = k/(k-1) is chosen so that 1/k+1/k′=11/k + 1/k' = 1, and write a=xka = x^k, b=yk′b = y^{k'}. The inequality becomes

xy≤xkk+yk′k′.xy \le \frac{x^k}{k} + \frac{y^{k'}}{k'}.

The left side is the area of a rectangle; the two terms on the right are areas under and beside one curve.

Explore Young's inequality as two areas

Try this. Drag the purple corner (x,y)(x, y) of the rectangle. The blue area lies under the curve s=tk−1s = t^{k-1} between t=0t = 0 and t=xt = x; the ochre area lies to the left of the same curve between s=0s = 0 and s=ys = y. Try to make the two areas fit the rectangle exactly, then change kk and try again.

ts
  • curve s = t^(k−1)
  • area under the curve on [0, x] = x^k/k
  • area left of the curve on [0, y] = y^k′/k′
  • rectangle x × y (drag its corner)
k=3,  k′=1.5.k = 3,\; k' = 1.5. Rectangle xy=1.26xy = 1.26; blue xk/k=0.915x^k/k = 0.915; ochre yk′/k′=0.569y^{k'}/k' = 0.569; sum 1.481.48. The two areas exceed the rectangle by 0.224. They fit exactly only when the corner lies on the curve, y = x^(k−1).
What to notice

The two coloured areas always cover the whole rectangle, and usually spill over: the spill-over is the gap in the inequality. They fit exactly only when the corner sits on the curve, y=xk−1y = x^{k-1}, which is the same as yk′=xky^{k'} = x^k, that is, a=ba = b. The curve read sideways is t=sk′−1t = s^{k'-1}, because (k−1)(k′−1)=1(k-1)(k'-1) = 1: this is why the conjugate index k′k' appears.

The areas are honest integrals: ∫0xtk−1 dt=xk/k\int_0^x t^{k-1}\,dt = x^k/k and ∫0ysk′−1 ds=yk′/k′\int_0^y s^{k'-1}\,ds = y^{k'}/k'. So the picture is already a proof of the two-number inequality, once one believes that the two regions together cover the rectangle. Hardy gives two other proofs, by the mean value theorem and by the shape of the graph of log⁡x\log x, and we follow him below.

From the two-number case, three things follow in a chain:

  1. the same inequality for mm numbers with mm weights (by induction);
  2. Hölder’s inequality, by applying the two-number case to each pair am,bma_m, b_m and adding;
  3. Minkowski’s inequality, by applying Hölder twice.

Precise statements

Throughout, as in Hardy, all letters stand for strictly positive numbers.

Theorem The general theorem of the means Hardy: G, Appendix I (2), (7)

If q1,q2,…,qmq_1, q_2, \dots, q_m are positive real numbers with q1+q2+⋯+qm=1q_1 + q_2 + \dots + q_m = 1, then

a1q1a2q2⋯amqm≤q1a1+q2a2+⋯+qmam,a_1^{q_1} a_2^{q_2} \cdots a_m^{q_m} \le q_1 a_1 + q_2 a_2 + \dots + q_m a_m,

with equality only if a1=a2=⋯=ama_1 = a_2 = \dots = a_m. The left side is the weighted geometric mean of the aa‘s, the right side their weighted arithmetic mean.

For m=2m = 2 Hardy writes this as aαb1−α≤αa+(1−α)ba^\alpha b^{1-\alpha} \le \alpha a + (1-\alpha) b (his (5)); he calls the mm-term statement GmG_m. When all qv=1/mq_v = 1/m it is the ordinary theorem that the geometric mean of mm numbers does not exceed their arithmetic mean (Misc. Examples I, 9).

Theorem Hölder's inequality Hardy: H, Appendix I (10)

Let k>1k > 1 and k′=k/(k−1)k' = k/(k-1), so that 1k+1k′=1\dfrac1k + \dfrac1{k'} = 1. For positive numbers a1,…,ana_1, \dots, a_n and b1,…,bnb_1, \dots, b_n,

∑m=1nambm≤(∑m=1namk)1/k(∑m=1nbmk′)1/k′,\sum_{m=1}^n a_m b_m \le \left(\sum_{m=1}^n a_m^k\right)^{1/k} \left(\sum_{m=1}^n b_m^{k'}\right)^{1/k'},

with equality only if amk/bmk′a_m^k / b_m^{k'} is the same for every mm.

Theorem Minkowski's inequality Hardy: M, Appendix I (15)

If k>1k > 1, then for positive numbers a1,…,ana_1, \dots, a_n and b1,…,bnb_1, \dots, b_n,

(∑m=1n(am+bm)k)1/k≤(∑m=1namk)1/k+(∑m=1nbmk)1/k,\left(\sum_{m=1}^n (a_m + b_m)^k\right)^{1/k} \le \left(\sum_{m=1}^n a_m^k\right)^{1/k} + \left(\sum_{m=1}^n b_m^k\right)^{1/k},

with equality only if the sets (a)(a) and (b)(b) are proportional, that is, am/bma_m / b_m is the same for every mm.

Corollary The companion inequality Hardy: Appendix I (17)

If k>1k > 1, then ∑(am+bm)k>∑amk+∑bmk\sum (a_m + b_m)^k > \sum a_m^k + \sum b_m^k.

The corollary goes the other way from Minkowski: raising to the power kk before adding makes the sum of the parts smaller than the whole. Its proof is one line: if a+b=1a + b = 1 then ak<aa^k < a and bk<bb^k < b, so ak+bk<1=(a+b)ka^k + b^k < 1 = (a+b)^k; both sides are homogeneous of degree kk, so the inequality (a+b)k>ak+bk(a + b)^k > a^k + b^k holds for all positive a,ba, b, and we add over mm.

Minkowski’s inequality is the triangle inequality for the “kk-distance”. In three dimensions with k=2k = 2 it says that one side of a triangle is at most the sum of the other two; in general it says that ∥u+v∥k≤∥u∥k+∥v∥k\|u + v\|_k \le \|u\|_k + \|v\|_k, where ∥u∥k=(∑∣um∣k)1/k\|u\|_k = \left(\sum |u_m|^k\right)^{1/k}. (For vectors with coordinates of either sign, apply the theorem to ∣um∣|u_m| and ∣vm∣|v_m|, and use ∣um+vm∣≤∣um∣+∣vm∣|u_m + v_m| \le |u_m| + |v_m|.)

Explore The kk-distance and its triangle inequality

Try this. Drag the tips of uu (blue) and vv (ochre). Then move kk: at k=1k = 1 the unit “circle” is a diamond, at k=2k = 2 a circle, for large kk nearly a square. Finally take kk below 11 and put u=(1,0)u = (1, 0), v=(0,1)v = (0, 1).

  • u (drag its tip)
  • v, placed at the tip of u (drag)
  • u + v
  • unit k-circle
  • k-circle of radius ‖u‖ + ‖v‖
With ∥(x,y)∥k=(∣x∣k+∣y∣k)1/k\|(x,y)\|_k = (|x|^k + |y|^k)^{1/k} and k=2k = 2: ∥u∥=1.53, ∥v∥=1.26, ∥u+v∥=2.42\|u\| = 1.53,\ \|v\| = 1.26,\ \|u+v\| = 2.42. The sum ∥u∥+∥v∥=2.79\|u\| + \|v\| = 2.79 is at least ∥u+v∥\|u+v\|: the triangle inequality holds.
What to notice

For every k≥1k \ge 1 the tip of u+vu + v stays inside the green kk-circle of radius ∥u∥+∥v∥\|u\| + \|v\|: that is Minkowski's inequality. For kk below 11 the unit kk-circle caves in, is no longer convex, and u+vu + v escapes: with u=(1,0)u = (1,0), v=(0,1)v = (0,1) the sum has length 21/k2^{1/k}, which is more than 22.

Proofs

The two-number case

Proof G2G_2: aαb1−α<αa+(1−α)ba^{\alpha} b^{1-\alpha} < \alpha a + (1 - \alpha) b when a≠ba \ne b
Strategy

Divide out by a power of aa to turn the inequality into a statement about the difference b1−α−a1−αb^{1-\alpha} - a^{1-\alpha}, and estimate that difference with the mean value theorem.

Reduce to a<ba < b and rearrange

The statement is symmetric under swapping (a,α)(a, \alpha) with (b,1−α)(b, 1 - \alpha), so we may suppose a<ba < b. Subtract aa from both sides: we need aαb1−α−a<(1−α)(b−a)a^\alpha b^{1-\alpha} - a < (1-\alpha)(b - a). Dividing by aα>0a^\alpha > 0 gives the equivalent form

b1−α−a1−α<(1−α)(b−a) a−α.b^{1-\alpha} - a^{1-\alpha} < (1 - \alpha)(b - a)\,a^{-\alpha}.

Apply the mean value theorem to x1−αx^{1-\alpha}

The function x1−αx^{1-\alpha} has derivative (1−α)x−α(1-\alpha)x^{-\alpha} for x>0x > 0. By the mean value theorem (§126) there is a ξ\xi with a<ξ<ba < \xi < b and

b1−α−a1−α=(1−α)(b−a) ξ−α.b^{1-\alpha} - a^{1-\alpha} = (1-\alpha)(b - a)\,\xi^{-\alpha}.

Compare ξ−α\xi^{-\alpha} with a−αa^{-\alpha}

Since −α<0-\alpha < 0, the power x−αx^{-\alpha} decreases as xx increases, and ξ>a\xi > a gives ξ−α<a−α\xi^{-\alpha} < a^{-\alpha}. With (1−α)(b−a)>0(1-\alpha)(b-a) > 0 this is exactly the required strict inequality.

There is a second proof that is really a picture. The curve y=log⁡xy = \log x is concave: its slope 1/x1/x decreases, so every chord lies below the curve. Take P=(a,log⁡a)P = (a, \log a) and Q=(b,log⁡b)Q = (b, \log b) and the point RR that divides the chord in the ratio 1−α:α1 - \alpha : \alpha. Its coordinates are the weighted averages

R=(αa+(1−α)b, αlog⁡a+(1−α)log⁡b).R = \bigl(\alpha a + (1-\alpha) b,\ \alpha \log a + (1-\alpha)\log b\bigr).

Since RR is below the curve, αlog⁡a+(1−α)log⁡b<log⁡(αa+(1−α)b)\alpha \log a + (1-\alpha)\log b < \log(\alpha a + (1-\alpha) b), and exponentiating gives G2G_2.

Explore The chord of log⁡x\log x lies under the curve

Try this. Move aa, bb and the weight α\alpha. Watch the purple gap between RR on the chord and the point of the curve directly above it. Then make aa and bb equal.

xlog x
  • y = log x
  • chord PQ
  • gap between R and the curve
  • G and A on the x-axis
Weighted geometric mean G=aαb1−α=1.732G = a^{\alpha} b^{1-\alpha} = 1.732; weighted arithmetic mean A=αa+(1−α)b=2.8A = \alpha a + (1-\alpha) b = 2.8. R sits at height log⁡G=0.5493\log G = 0.5493, the curve above it at log⁡A=1.03\log A = 1.03, so G<AG < A.
What to notice

The height of RR is log⁡G\log G and the curve above it is at log⁡A\log A, so the purple gap is log⁡(A/G)\log(A/G). It is never negative, it is large when aa and bb are far apart, and it vanishes only when a=ba = b (or a weight is 00). The two dashed lines on the axis mark GG and AA: GG is always on the left.

From two numbers to mm numbers

Proof GmG_m follows from G2G_2 and Gm−1G_{m-1}
Strategy

Group the first m−1m - 1 numbers together, so that the mm-term mean becomes a two-term mean of “the first group” and ama_m. Apply G2G_2 to that, then Gm−1G_{m-1} inside the group.

Rescale the first m−1m - 1 weights

Let q=q1+⋯+qm−1q = q_1 + \dots + q_{m-1}, so q+qm=1q + q_m = 1, and put qv′=qv/qq_v' = q_v / q for v≤m−1v \le m - 1. Then q1′+⋯+qm−1′=1q_1' + \dots + q_{m-1}' = 1, and

a1q1⋯am−1qm−1amqm=(a1q1′⋯am−1qm−1′)qamqm.a_1^{q_1}\cdots a_{m-1}^{q_{m-1}} a_m^{q_m} = \bigl(a_1^{q_1'} \cdots a_{m-1}^{q_{m-1}'}\bigr)^{q} a_m^{q_m}.

Apply G2G_2 with weights qq and qmq_m

The right side is a two-term geometric mean, so it is at most q (a1q1′⋯am−1qm−1′)+qmamq\,\bigl(a_1^{q_1'} \cdots a_{m-1}^{q_{m-1}'}\bigr) + q_m a_m, with strict inequality unless the bracket equals ama_m.

Apply Gm−1G_{m-1} inside the bracket

By the induction hypothesis the bracket is at most q1′a1+⋯+qm−1′am−1q_1' a_1 + \dots + q_{m-1}' a_{m-1}, strictly unless a1=⋯=am−1a_1 = \dots = a_{m-1}. Multiplying by qq turns q qv′q\,q_v' back into qvq_v, and the total is q1a1+⋯+qmamq_1 a_1 + \dots + q_m a_m.

When is there equality?

Equality throughout needs a1=⋯=am−1a_1 = \dots = a_{m-1} (third step), and then the bracket equals a1a_1, so the second step needs a1=ama_1 = a_m. Hence all the aa‘s are equal.

Why not just take limits? For rational weights, GmG_m follows from the ordinary theorem of the means by repeating each ava_v the right number of times. One could then approximate real weights by rationals and let them tend to their limits. Hardy points out the flaw: strict inequalities survive a limit only as weak ones (xr>yrx_r > y_r for every rr gives only lim⁡xr≥lim⁡yr\lim x_r \ge \lim y_r), so the condition for equality would be lost. That is why the proof above goes direct. It is also why the appendix belongs after §214: for irrational α\alpha, the power aαa^\alpha has no meaning until exp⁡\exp and log⁡\log have been defined.

Hölder

Proof Hölder's inequality
Strategy

Both sides scale in the same way, so it is enough to prove the inequality when ∑ak=∑bk′=1\sum a^k = \sum b^{k'} = 1. Then the right side is 11, and G2G_2 applied to each term ambma_m b_m gives a sum of weighted means that adds up to exactly 11.

Normalise

If every ama_m is multiplied by λ>0\lambda > 0, both sides of the inequality are multiplied by λ\lambda; similarly for the bb‘s. So we may assume ∑amk=1\sum a_m^k = 1 and ∑bmk′=1\sum b_m^{k'} = 1, and must prove ∑ambm≤1\sum a_m b_m \le 1.

Apply G2G_2 to each product

With α=1/k\alpha = 1/k and β=1/k′\beta = 1/k', we have ambm=(amk)α(bmk′)β≤α amk+β bmk′a_m b_m = (a_m^k)^{\alpha} (b_m^{k'})^{\beta} \le \alpha\, a_m^k + \beta\, b_m^{k'}. (This is Young’s inequality from the widget.)

Add

Summing over mm: ∑ambm≤α∑amk+β∑bmk′=α+β=1\sum a_m b_m \le \alpha \sum a_m^k + \beta \sum b_m^{k'} = \alpha + \beta = 1.

Equality

Equality needs equality in every term, that is amk=bmk′a_m^k = b_m^{k'} for each mm. Undoing the normalisation, this says that amk/bmk′a_m^k / b_m^{k'} does not depend on mm.

Minkowski

Proof Minkowski's inequality
Strategy

Write S=∑(a+b)kS = \sum (a + b)^k and split off one factor a+ba + b: S=∑a(a+b)k−1+∑b(a+b)k−1S = \sum a(a+b)^{k-1} + \sum b(a+b)^{k-1}. Each piece is a sum of products, so Hölder applies to it; both estimates contain the same power of SS, which can be divided out.

Split SS

Dropping the suffixes, S=∑(a+b)k=∑a (a+b)k−1+∑b (a+b)k−1S = \sum (a+b)^k = \sum a\,(a+b)^{k-1} + \sum b\,(a+b)^{k-1}.

Apply Hölder to each sum

Since (k−1)k′=k(k-1)k' = k,

∑a (a+b)k−1≤(∑ak)1/k(∑(a+b)(k−1)k′)1/k′=(∑ak)1/kS1/k′,\sum a\,(a+b)^{k-1} \le \left(\sum a^k\right)^{1/k}\left(\sum (a+b)^{(k-1)k'}\right)^{1/k'} = \left(\sum a^k\right)^{1/k} S^{1/k'},

and similarly for the bb-sum. Adding, S≤[(∑ak)1/k+(∑bk)1/k]S1/k′S \le \left[\left(\sum a^k\right)^{1/k} + \left(\sum b^k\right)^{1/k}\right] S^{1/k'}.

Divide by S1/k′S^{1/k'}

S>0S > 0, and 1−1/k′=1/k1 - 1/k' = 1/k, so dividing gives S1/k≤(∑ak)1/k+(∑bk)1/kS^{1/k} \le \left(\sum a^k\right)^{1/k} + \left(\sum b^k\right)^{1/k}, which is the inequality.

Equality

By Hölder’s equality condition, equality needs aka^k proportional to (a+b)(k−1)k′=(a+b)k(a+b)^{(k-1)k'} = (a+b)^k, so aa proportional to a+ba + b; likewise bb proportional to a+ba + b. Then aa and bb are proportional to each other. Conversely, if bm=λamb_m = \lambda a_m both sides equal (1+λ)(∑ak)1/k(1 + \lambda)\left(\sum a^k\right)^{1/k}.

Exercises

These are our own, chosen to test the ideas of the appendix.

Exercise Cauchy is the case k=2k = 2 ●○○

Show that Hölder’s inequality with k=2k = 2 is Cauchy’s inequality (∑ambm)2≤∑am2∑bm2\left(\sum a_m b_m\right)^2 \le \sum a_m^2 \sum b_m^2, and check that the equality conditions agree.

Hint
What is k′k' when k=2k = 2?

Solution

When k=2k = 2, k′=2/(2−1)=2k' = 2/(2-1) = 2. Hölder reads ∑ab≤(∑a2)1/2(∑b2)1/2\sum ab \le \left(\sum a^2\right)^{1/2}\left(\sum b^2\right)^{1/2}; squaring gives Cauchy’s inequality. Hölder’s equality condition, am2/bm2a_m^2/b_m^2 independent of mm, is the same as am/bma_m/b_m independent of mm (all numbers are positive), which is the usual condition.

Exercise A power mean inequality ●○○ Hardy: Appendix I

Use Hölder’s inequality with b1=⋯=bn=1b_1 = \dots = b_n = 1 to show that (∑m=1nam)k≤nk−1∑m=1namk\left(\sum_{m=1}^n a_m\right)^k \le n^{k-1}\sum_{m=1}^n a_m^k for k>1k > 1. When is there equality? Interpret the result as a comparison of two kinds of average.

Hint
∑bk′=n\sum b^{k'} = n, and (n1/k′)k=nk/k′=nk−1(n^{1/k'})^k = n^{k/k'} = n^{k-1}.

Solution

Hölder gives ∑a=∑a⋅1≤(∑ak)1/kn1/k′\sum a = \sum a \cdot 1 \le \left(\sum a^k\right)^{1/k} n^{1/k'}. Raise to the power kk: (∑a)k≤nk/k′∑ak=nk−1∑ak\left(\sum a\right)^k \le n^{k/k'} \sum a^k = n^{k-1}\sum a^k, since k/k′=k−1k/k' = k - 1. Equality needs amk/1a_m^k / 1 independent of mm, so all ama_m equal. Dividing by nkn^k: (1n∑a)k≤1n∑ak\left(\frac1n\sum a\right)^k \le \frac1n \sum a^k, so the ordinary mean is at most the “kk-th power mean” (1n∑ak)1/k\left(\frac1n\sum a^k\right)^{1/k}.

Exercise Weights that are not equal ●○○

Use GmG_m to show that x2y3z≤(2x+3y+z6)6x^2 y^3 z \le \left(\dfrac{2x + 3y + z}{6}\right)^6 for positive x,y,zx, y, z. When is there equality?

Hint
Take weights 26,36,16\tfrac26, \tfrac36, \tfrac16 and the numbers x,y,zx, y, z.

Solution

With q1=13q_1 = \tfrac13, q2=12q_2 = \tfrac12, q3=16q_3 = \tfrac16 (sum 11), G3G_3 gives x1/3y1/2z1/6≤13x+12y+16z=2x+3y+z6x^{1/3} y^{1/2} z^{1/6} \le \tfrac13 x + \tfrac12 y + \tfrac16 z = \dfrac{2x + 3y + z}{6}. Raising both (positive) sides to the sixth power gives the result. Equality holds exactly when x=y=zx = y = z.

Exercise The best box ●●○

Among all positive x,yx, y with x+2y=3x + 2y = 3, find the largest value of xy2x y^2, using only the theorem of the means.

Hint
Write xy2=4⋅x⋅y2⋅y2xy^2 = 4 \cdot x \cdot \tfrac y2 \cdot \tfrac y2… or, more cleverly, split x+2yx + 2y into three equal-weight pieces x,y,yx, y, y.

Solution

By the ordinary theorem of the means with three numbers x,y,yx, y, y: (xyy)1/3≤13(x+y+y)=13(x+2y)=1(x y y)^{1/3} \le \tfrac13(x + y + y) = \tfrac13(x + 2y) = 1. So xy2≤1xy^2 \le 1, with equality exactly when x=yx = y, that is x=y=1x = y = 1. The maximum is 11, attained at (1,1)(1, 1). No calculus was needed: the equality condition locates the maximum.

Exercise Why k>1k > 1 matters ●●○

(a) Show that for 0<k<10 < k < 1 and positive a,ba, b one has (a+b)k<ak+bk(a + b)^k < a^k + b^k, the reverse of the companion inequality. (b) Deduce that Minkowski’s inequality fails for k=12k = \tfrac12, by an example with n=2n = 2.

Hint
Copy Hardy’s proof of (16): if a+b=1a + b = 1, compare aka^k with aa when k<1k < 1.
Hint
Try a=(1,0)a = (1, 0)… but all numbers must be positive here, so try a=(1,ε)a = (1, \varepsilon), b=(ε,1)b = (\varepsilon, 1) with ε\varepsilon small, or argue by continuity from (1,0)(1, 0) and (0,1)(0, 1).

Solution

(a) If a+b=1a + b = 1 then 0<a,b<10 < a, b < 1, and for k<1k < 1 we have ak>aa^k > a, bk>bb^k > b, so ak+bk>a+b=1=(a+b)ka^k + b^k > a + b = 1 = (a + b)^k. Both sides are homogeneous of degree kk, so the inequality holds for all positive a,ba, b.

(b) Take k=12k = \tfrac12, a=(1,ε)a = (1, \varepsilon), b=(ε,1)b = (\varepsilon, 1). The left side of Minkowski is (21+ε)2=4(1+ε)\left(2\sqrt{1 + \varepsilon}\right)^2 = 4(1+\varepsilon). Each term on the right is (1+ε)2(1 + \sqrt\varepsilon)^2, so the right side is 2(1+ε)22(1 + \sqrt\varepsilon)^2. As ε→0\varepsilon \to 0 the left tends to 44 and the right to 22, so for small ε\varepsilon (for example ε=0.01\varepsilon = 0.01: 4.044.04 against 2.422.42) the inequality is reversed. Geometrically: the unit “circle” for k=12k = \tfrac12 is not convex, as the widget shows.

Exercise Hölder for three sets ●●○ Hardy: Appendix I (13)

Show that if α,β,γ>0\alpha, \beta, \gamma > 0 and α+β+γ=1\alpha + \beta + \gamma = 1 then ∑aαbβcγ≤(∑a)α(∑b)β(∑c)γ\sum a^\alpha b^\beta c^\gamma \le \left(\sum a\right)^\alpha \left(\sum b\right)^\beta \left(\sum c\right)^\gamma.

Hint
Normalise so that ∑a=∑b=∑c=1\sum a = \sum b = \sum c = 1 and use G3G_3 on each term.

Solution

Each side is multiplied by λα\lambda^\alpha when every aa is multiplied by λ\lambda, and similarly for bb and cc; so we may suppose ∑a=∑b=∑c=1\sum a = \sum b = \sum c = 1. By G3G_3, aαbβcγ≤αa+βb+γca^\alpha b^\beta c^\gamma \le \alpha a + \beta b + \gamma c for each mm. Adding, ∑aαbβcγ≤α+β+γ=1\sum a^\alpha b^\beta c^\gamma \le \alpha + \beta + \gamma = 1, which is the right side. Equality needs am=bm=cma_m = b_m = c_m for each mm after normalising, that is, the three sets proportional.

Exercise Reading the Young picture ●●●

In the widget, the blue and ochre regions together always cover the rectangle [0,x]×[0,y][0, x] \times [0, y]. Prove this, and deduce Young’s inequality xy≤xk/k+yk′/k′xy \le x^k/k + y^{k'}/k' without using the theorem of the means.

Hint
Suppose y≤xk−1y \le x^{k-1} (the corner is below the curve). Split the rectangle by the vertical line through the point where the curve reaches height yy.

Solution

Let g(t)=tk−1g(t) = t^{k-1}; it is increasing with inverse g−1(s)=s1/(k−1)=sk′−1g^{-1}(s) = s^{1/(k-1)} = s^{k'-1}. Suppose y≤g(x)y \le g(x) and let t0=g−1(y)≤xt_0 = g^{-1}(y) \le x. A point (t,s)(t, s) of the rectangle with s≤g(t)s \le g(t) lies under the curve over [0,x][0, x], so in the blue region. A point with s>g(t)s > g(t) has t<g−1(s)t < g^{-1}(s) and s≤ys \le y, so it lies to the left of the curve over [0,y][0, y], in the ochre region. So the rectangle is covered, and its area is at most the sum: xy≤∫0xg+∫0yg−1=xk/k+yk′/k′xy \le \int_0^x g + \int_0^y g^{-1} = x^k/k + y^{k'}/k'. The excess is the area between the curve and the line s=ys = y for t0≤t≤xt_0 \le t \le x, which is zero only if t0=xt_0 = x, that is y=xk−1y = x^{k-1}. The case y≥g(x)y \ge g(x) is the same with the roles of the axes exchanged.

Check yourself

Check yourself Enter a number

What is the conjugate index k′k' when k=3k = 3?

Why

k′=k/(k−1)=3/2k' = k/(k-1) = 3/2. Check: 13+23=1\tfrac13 + \tfrac23 = 1. Note that kk and k′k' are symmetric: the conjugate of 32\tfrac32 is 33, and k=2k = 2 is the only self-conjugate index.

Check yourself Choose one

For k=3k = 3, when is there equality in Hölder’s inequality ∑ab≤(∑a3)1/3(∑b3/2)2/3\sum ab \le \left(\sum a^3\right)^{1/3}\left(\sum b^{3/2}\right)^{2/3}?

Why

Equality needs amka_m^k proportional to bmk′b_m^{k'}. With k=3k = 3, k′=32k' = \tfrac32: a3∝b3/2a^3 \propto b^{3/2}, i.e. b∝a2b \propto a^2. “Proportional sets” is the right condition only when k=2k = 2, or for Hölder in the form ∑aαbβ≤(∑a)α(∑b)β\sum a^\alpha b^\beta \le (\sum a)^\alpha(\sum b)^\beta, which is the form in which Hardy proves the equality case (his statement of (10) carries the same words, which is a slip). With n=1n = 1 there is always equality, but not only then.

Check yourself True or false?

If xr>yrx_r > y_r for every rr, and xr→xx_r \to x, yr→yy_r \to y, then x>yx > y.

Why

False: xr=1/rx_r = 1/r, yr=0y_r = 0 gives x=y=0x = y = 0. Strict inequalities become weak in the limit. That is exactly why Hardy does not prove GG for real weights by approximating with rational weights: it would prove ≤\le but lose the case of equality.

Check yourself Choose one

Minkowski’s inequality with k=2k = 2 and n=3n = 3 says that…

Why

With k=2k = 2, (∑(a+b)2)1/2\left(\sum (a+b)^2\right)^{1/2} is the length of the vector a+ba + b, and the right side is the sum of the lengths of aa and bb. In the triangle with vertices 00, aa and −b-b these are the three sides. The other options are true or false statements about different things.

Hardy’s language

There are three inequalities which are particularly important in analysis, the theorem of the arithmetic and geometric means and Hölder’s and Minkowski’s inequalities.

— G. H. Hardy, Appendix I

Summary

Summary
  • The weighted theorem of the means a1q1⋯amqm≤q1a1+⋯+qmama_1^{q_1}\cdots a_m^{q_m} \le q_1 a_1 + \dots + q_m a_m (∑q=1\sum q = 1) holds for real weights, with equality only when all aa‘s are equal. The two-term case follows from the mean value theorem, or from the concavity of log⁡x\log x; induction does the rest.
  • Young’s form xy≤xk/k+yk′/k′xy \le x^k/k + y^{k'}/k' is a statement about two areas covering a rectangle.
  • Hölder: normalise, apply the two-term case to each product, add. Equality when ak∝bk′a^k \propto b^{k'}.
  • Minkowski: split ∑(a+b)k\sum(a+b)^k, apply Hölder twice, divide. It is the triangle inequality for the kk-distance, and it fails for k<1k < 1.
  • Approximating real weights by rational ones would lose the equality case: strict inequalities turn weak in the limit.