The ordinary distance in the plane, x2+y2, is one member of a family: for any k≥1 the quantity (∣x∣k+∣y∣k)1/k, or with n coordinates (∑∣am∣k)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. 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>0 and the weights α,β>0 add up to 1, then
aαbβ≤αa+βb,
with equality only when a=b. With α=β=21 this is the familiar ab≤21(a+b). Here is a way to see the general case, due to W. H. Young. Put α=1/k and β=1/k′, where k>1 and k′=k/(k−1) is chosen so that 1/k+1/k′=1, and write a=xk, b=yk′. The inequality becomes
xy≤kxk+k′yk′.
The left side is the area of a rectangle; the two terms on the right are areas under and beside one curve.
ExploreYoung's inequality as two areas
Try this.Drag the purple corner (x,y) of the rectangle. The blue area lies under the curve s=tk−1 between t=0 and t=x; the ochre area lies to the left of the same curve between s=0 and s=y. Try to make the two areas fit the rectangle exactly, then change k 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. Rectangle xy=1.26; blue xk/k=0.915; ochre yk′/k′=0.569; sum 1.48. 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−1, which is the same as yk′=xk, that is, a=b. The curve read sideways is t=sk′−1, because (k−1)(k′−1)=1: this is why the conjugate index k′ appears.
The areas are honest integrals: ∫0xtk−1dt=xk/k and ∫0ysk′−1ds=yk′/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 logx, and we follow him below.
From the two-number case, three things follow in a chain:
the same inequality for m numbers with m weights (by induction);
Hölder’s inequality, by applying the two-number case to each pair am,bm and adding;
Minkowski’s inequality, by applying Hölder twice.
Precise statements
Throughout, as in Hardy, all letters stand for strictly positive numbers.
TheoremThe general theorem of the meansHardy: G, Appendix I (2), (7)
If q1,q2,…,qm are positive real numbers with q1+q2+⋯+qm=1, then
a1q1a2q2⋯amqm≤q1a1+q2a2+⋯+qmam,
with equality only if a1=a2=⋯=am. The left side is the weighted geometric mean of the a‘s, the right side their weighted arithmetic mean.
For m=2 Hardy writes this as aαb1−α≤αa+(1−α)b (his (5)); he calls the m-term statement Gm. When all qv=1/m it is the ordinary theorem that the geometric mean of m numbers does not exceed their arithmetic mean (Misc. Examples I, 9).
TheoremHölder's inequalityHardy: H, Appendix I (10)
Let k>1 and k′=k/(k−1), so that k1+k′1=1. For positive numbers a1,…,an and b1,…,bn,
m=1∑nambm≤(m=1∑namk)1/k(m=1∑nbmk′)1/k′,
with equality only if amk/bmk′ is the same for every m.
TheoremMinkowski's inequalityHardy: M, Appendix I (15)
If k>1, then for positive numbers a1,…,an and b1,…,bn,
with equality only if the sets (a) and (b) are proportional, that is, am/bm is the same for every m.
CorollaryThe companion inequalityHardy: Appendix I (17)
If k>1, then ∑(am+bm)k>∑amk+∑bmk.
The corollary goes the other way from Minkowski: raising to the power kbefore adding makes the sum of the parts smaller than the whole. Its proof is one line: if a+b=1 then ak<a and bk<b, so ak+bk<1=(a+b)k; both sides are homogeneous of degree k, so the inequality (a+b)k>ak+bk holds for all positive a,b, and we add over m.
Minkowski’s inequality is the triangle inequality for the “k-distance”. In three dimensions with k=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, where ∥u∥k=(∑∣um∣k)1/k. (For vectors with coordinates of either sign, apply the theorem to ∣um∣ and ∣vm∣, and use ∣um+vm∣≤∣um∣+∣vm∣.)
ExploreThe k-distance and its triangle inequality
Try this.Drag the tips of u (blue) and v (ochre). Then move k: at k=1 the unit “circle” is a diamond, at k=2 a circle, for large k nearly a square. Finally take k below 1 and put u=(1,0), 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 and k=2: ∥u∥=1.53,∥v∥=1.26,∥u+v∥=2.42. The sum ∥u∥+∥v∥=2.79 is at least ∥u+v∥: the triangle inequality holds.
What to notice
For every k≥1 the tip of u+v stays inside the green k-circle of radius ∥u∥+∥v∥: that is Minkowski's inequality. For k below 1 the unit k-circle caves in, is no longer convex, and u+v escapes: with u=(1,0), v=(0,1) the sum has length 21/k, which is more than 2.
Proofs
The two-number case
ProofG2: aαb1−α<αa+(1−α)b when a=b
Strategy
Divide out by a power of a to turn the inequality into a statement about the difference b1−α−a1−α, and estimate that difference with the mean value theorem.
Reduce to a<b and rearrange
The statement is symmetric under swapping (a,α) with (b,1−α), so we may suppose a<b. Subtract a from both sides: we need aαb1−α−a<(1−α)(b−a). Dividing by aα>0 gives the equivalent form
b1−α−a1−α<(1−α)(b−a)a−α.
Apply the mean value theorem to x1−α
The function x1−α has derivative (1−α)x−α for x>0. By the mean value theorem (§126) there is a ξ with a<ξ<b and
b1−α−a1−α=(1−α)(b−a)ξ−α.
Compare ξ−α with a−α
Since −α<0, the power x−α decreases as x increases, and ξ>a gives ξ−α<a−α. With (1−α)(b−a)>0 this is exactly the required strict inequality.
∎
There is a second proof that is really a picture. The curve y=logx is concave: its slope 1/x decreases, so every chord lies below the curve. Take P=(a,loga) and Q=(b,logb) and the point R that divides the chord in the ratio 1−α:α. Its coordinates are the weighted averages
R=(αa+(1−α)b,αloga+(1−α)logb).
Since R is below the curve, αloga+(1−α)logb<log(αa+(1−α)b), and exponentiating gives G2.
ExploreThe chord of logx lies under the curve
Try this.Move a, b and the weight α. Watch the purple gap between R on the chord and the point of the curve directly above it. Then make a and b 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=1.732; weighted arithmetic mean A=αa+(1−α)b=2.8. R sits at height logG=0.5493, the curve above it at logA=1.03, so G<A.
What to notice
The height of R is logG and the curve above it is at logA, so the purple gap is log(A/G). It is never negative, it is large when a and b are far apart, and it vanishes only when a=b (or a weight is 0). The two dashed lines on the axis mark G and A: G is always on the left.
From two numbers to m numbers
ProofGm follows from G2 and Gm−1
Strategy
Group the first m−1 numbers together, so that the m-term mean becomes a two-term mean of “the first group” and am. Apply G2 to that, then Gm−1 inside the group.
Rescale the first m−1 weights
Let q=q1+⋯+qm−1, so q+qm=1, and put qv′=qv/q for v≤m−1. Then q1′+⋯+qm−1′=1, and
The right side is a two-term geometric mean, so it is at most q(a1q1′⋯am−1qm−1′)+qmam, with strict inequality unless the bracket equals am.
Apply Gm−1 inside the bracket
By the induction hypothesis the bracket is at most q1′a1+⋯+qm−1′am−1, strictly unless a1=⋯=am−1. Multiplying by q turns qqv′ back into qv, and the total is q1a1+⋯+qmam.
When is there equality?
Equality throughout needs a1=⋯=am−1 (third step), and then the bracket equals a1, so the second step needs a1=am. Hence all the a‘s are equal.
∎
Why not just take limits? For rational weights, Gm follows from the ordinary theorem of the means by repeating each av 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>yr for every r gives only limxr≥limyr), 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, the power aα has no meaning until exp and log have been defined.
Hölder
ProofHölder's inequality
Strategy
Both sides scale in the same way, so it is enough to prove the inequality when ∑ak=∑bk′=1. Then the right side is 1, and G2 applied to each term ambm gives a sum of weighted means that adds up to exactly 1.
Normalise
If every am is multiplied by λ>0, both sides of the inequality are multiplied by λ; similarly for the b‘s. So we may assume ∑amk=1 and ∑bmk′=1, and must prove ∑ambm≤1.
Apply G2 to each product
With α=1/k and β=1/k′, we have ambm=(amk)α(bmk′)β≤αamk+βbmk′. (This is Young’s inequality from the widget.)
Add
Summing over m: ∑ambm≤α∑amk+β∑bmk′=α+β=1.
Equality
Equality needs equality in every term, that is amk=bmk′ for each m. Undoing the normalisation, this says that amk/bmk′ does not depend on m.
∎
Minkowski
ProofMinkowski's inequality
Strategy
Write S=∑(a+b)k and split off one factor a+b: S=∑a(a+b)k−1+∑b(a+b)k−1. Each piece is a sum of products, so Hölder applies to it; both estimates contain the same power of S, which can be divided out.
Split S
Dropping the suffixes, S=∑(a+b)k=∑a(a+b)k−1+∑b(a+b)k−1.
and similarly for the b-sum. Adding, S≤[(∑ak)1/k+(∑bk)1/k]S1/k′.
Divide by S1/k′
S>0, and 1−1/k′=1/k, so dividing gives S1/k≤(∑ak)1/k+(∑bk)1/k, which is the inequality.
Equality
By Hölder’s equality condition, equality needs ak proportional to (a+b)(k−1)k′=(a+b)k, so a proportional to a+b; likewise b proportional to a+b. Then a and b are proportional to each other. Conversely, if bm=λam both sides equal (1+λ)(∑ak)1/k.
∎
Exercises
These are our own, chosen to test the ideas of the appendix.
ExerciseCauchy is the case k=2●○○
Show that Hölder’s inequality with k=2 is Cauchy’s inequality (∑ambm)2≤∑am2∑bm2, and check that the equality conditions agree.
Hint
What is k′ when k=2?
Solution
When k=2, k′=2/(2−1)=2. Hölder reads ∑ab≤(∑a2)1/2(∑b2)1/2; squaring gives Cauchy’s inequality. Hölder’s equality condition, am2/bm2 independent of m, is the same as am/bm independent of m (all numbers are positive), which is the usual condition.
ExerciseA power mean inequality●○○Hardy: Appendix I
Use Hölder’s inequality with b1=⋯=bn=1 to show that (∑m=1nam)k≤nk−1∑m=1namk for k>1. When is there equality? Interpret the result as a comparison of two kinds of average.
Hint
∑bk′=n, and (n1/k′)k=nk/k′=nk−1.
Solution
Hölder gives ∑a=∑a⋅1≤(∑ak)1/kn1/k′. Raise to the power k: (∑a)k≤nk/k′∑ak=nk−1∑ak, since k/k′=k−1. Equality needs amk/1 independent of m, so all am equal. Dividing by nk: (n1∑a)k≤n1∑ak, so the ordinary mean is at most the “k-th power mean” (n1∑ak)1/k.
ExerciseWeights that are not equal●○○
Use Gm to show that x2y3z≤(62x+3y+z)6 for positive x,y,z. When is there equality?
Hint
Take weights 62,63,61 and the numbers x,y,z.
Solution
With q1=31, q2=21, q3=61 (sum 1), G3 gives x1/3y1/2z1/6≤31x+21y+61z=62x+3y+z. Raising both (positive) sides to the sixth power gives the result. Equality holds exactly when x=y=z.
ExerciseThe best box●●○
Among all positive x,y with x+2y=3, find the largest value of xy2, using only the theorem of the means.
Hint
Write xy2=4⋅x⋅2y⋅2y… or, more cleverly, split x+2y into three equal-weight pieces x,y,y.
Solution
By the ordinary theorem of the means with three numbers x,y,y: (xyy)1/3≤31(x+y+y)=31(x+2y)=1. So xy2≤1, with equality exactly when x=y, that is x=y=1. The maximum is 1, attained at (1,1). No calculus was needed: the equality condition locates the maximum.
ExerciseWhy k>1 matters●●○
(a) Show that for 0<k<1 and positive a,b one has (a+b)k<ak+bk, the reverse of the companion inequality. (b) Deduce that Minkowski’s inequality fails for k=21, by an example with n=2.
Hint
Copy Hardy’s proof of (16): if a+b=1, compare ak with a when k<1.
Hint
Try a=(1,0)… but all numbers must be positive here, so try a=(1,ε), b=(ε,1) with ε small, or argue by continuity from (1,0) and (0,1).
Solution
(a) If a+b=1 then 0<a,b<1, and for k<1 we have ak>a, bk>b, so ak+bk>a+b=1=(a+b)k. Both sides are homogeneous of degree k, so the inequality holds for all positive a,b.
(b) Take k=21, a=(1,ε), b=(ε,1). The left side of Minkowski is (21+ε)2=4(1+ε). Each term on the right is (1+ε)2, so the right side is 2(1+ε)2. As ε→0 the left tends to 4 and the right to 2, so for small ε (for example ε=0.01: 4.04 against 2.42) the inequality is reversed. Geometrically: the unit “circle” for k=21 is not convex, as the widget shows.
ExerciseHölder for three sets●●○Hardy: Appendix I (13)
Show that if α,β,γ>0 and α+β+γ=1 then ∑aαbβcγ≤(∑a)α(∑b)β(∑c)γ.
Hint
Normalise so that ∑a=∑b=∑c=1 and use G3 on each term.
Solution
Each side is multiplied by λα when every a is multiplied by λ, and similarly for b and c; so we may suppose ∑a=∑b=∑c=1. By G3, aαbβcγ≤αa+βb+γc for each m. Adding, ∑aαbβcγ≤α+β+γ=1, which is the right side. Equality needs am=bm=cm for each m after normalising, that is, the three sets proportional.
ExerciseReading the Young picture●●●
In the widget, the blue and ochre regions together always cover the rectangle [0,x]×[0,y]. Prove this, and deduce Young’s inequality xy≤xk/k+yk′/k′ without using the theorem of the means.
Hint
Suppose y≤xk−1 (the corner is below the curve). Split the rectangle by the vertical line through the point where the curve reaches height y.
Solution
Let g(t)=tk−1; it is increasing with inverse g−1(s)=s1/(k−1)=sk′−1. Suppose y≤g(x) and let t0=g−1(y)≤x. A point (t,s) of the rectangle with s≤g(t) lies under the curve over [0,x], so in the blue region. A point with s>g(t) has t<g−1(s) and s≤y, so it lies to the left of the curve over [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′. The excess is the area between the curve and the line s=y for t0≤t≤x, which is zero only if t0=x, that is y=xk−1. The case y≥g(x) is the same with the roles of the axes exchanged.
Check yourself
Check yourselfEnter a number
What is the conjugate index k′ when k=3?
Why
k′=k/(k−1)=3/2. Check: 31+32=1. Note that k and k′ are symmetric: the conjugate of 23 is 3, and k=2 is the only self-conjugate index.
Check yourselfChoose one
For k=3, when is there equality in Hölder’s inequality ∑ab≤(∑a3)1/3(∑b3/2)2/3?
Why
Equality needs amk proportional to bmk′. With k=3, k′=23: a3∝b3/2, i.e. b∝a2. “Proportional sets” is the right condition only when k=2, or for Hölder in the form ∑aαbβ≤(∑a)α(∑b)β, 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=1 there is always equality, but not only then.
Check yourselfTrue or false?
If xr>yr for every r, and xr→x, yr→y, then x>y.
Why
False: xr=1/r, yr=0 gives x=y=0. Strict inequalities become weak in the limit. That is exactly why Hardy does not prove G for real weights by approximating with rational weights: it would prove ≤ but lose the case of equality.
Check yourselfChoose one
Minkowski’s inequality with k=2 and n=3 says that…
Why
With k=2, (∑(a+b)2)1/2 is the length of the vector a+b, and the right side is the sum of the lengths of a and b. In the triangle with vertices 0, a and −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+⋯+qmam (∑q=1) holds for real weights, with equality only when all a‘s are equal. The two-term case follows from the mean value theorem, or from the concavity of logx; induction does the rest.
Young’s form xy≤xk/k+yk′/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′.
Minkowski: split ∑(a+b)k, apply Hölder twice, divide. It is the triangle inequality for the k-distance, and it fails for k<1.
Approximating real weights by rational ones would lose the equality case: strict inequalities turn weak in the limit.