HelpContactFor teachers and institutes
>_StemPy Lab
Mock TestsClasses

>_StemPy Lab

Learn. Solve. Code the solutions. Test yourself.

Live examsFree mock exams, timed and rankedEveryone starts together, and your score and rank come the moment it ends.See the exams →
Live examsInternational live exams are comingTell us which exam you want first. Timed practice papers are open now.See what is coming →

Learn

  • Physics
  • Chemistry
  • Mathematics
  • Biology
  • Python Programming
  • Statistics & Data Science
  • Computer Science
  • Space Science and Astronomy
  • Earth and Environmental Science
  • Classes
  • Groups

Practice

  • Problems
  • Quizzes
  • Mock Tests
  • Prepare for an Exam
  • Question Sets
  • Exams
  • Certificates
  • Python Labs
  • AI Practice Generator
  • Student-Created Tests

Community

  • Discussions
  • Resources
  • Notes
  • Glossary
  • Blog

Account

  • Log in
© 2026 senLearn. All rights reserved.HelpFor SchoolsAboutContactPrivacy PolicyTerms of UseRefund Policy
The AM-GM-HM Inequality and Power Means
  1. 1Arithmetic, geometric and harmonic means
  2. 2Weighted means, power means and Bernoulli
  3. 3Greatest and least values with the means
  1. Mathematics
  2. /The AM-GM-HM Inequality and Power Means
  3. /Weighted means, power means and Bernoulli

Course 29.1·Lesson 2 of 3

Weighted means, power means and Bernoulli

Weights let one number count more than another: w1a1+w2a2+⋯≥a1w1a2w2⋯w_1 a_1 + w_2 a_2 + \dots \ge a_1^{w_1} a_2^{w_2} \cdotsw1​a1​+w2​a2​+⋯≥a1w1​​a2w2​​⋯

Lesson 2 of 3Show all lessonsHide
The AM-GM-HM Inequality and Power Means
  1. 1Arithmetic, geometric and harmonic means
  2. 2Weighted means, power means and Bernoulli
  3. 3Greatest and least values with the means

Try the code and the simulations beside the lesson.

What you will be able to do

  • use the weighted AM-GM inequality to handle terms with unequal roles;
  • define the power mean MpM_pMp​ and use the chain Mp≤MqM_p \le M_qMp​≤Mq​ for p<qp < qp<q;
  • prove and use Bernoulli's inequality.

When the numbers do not count equally

A student scores 606060 in a test worth one part and 909090 in a test worth two parts. The final mark is not the plain average 757575. It is 13×60+23×90=80\dfrac{1}{3} \times 60 + \dfrac{2}{3} \times 90 = 8031​×60+32​×90=80. The numbers 13\dfrac{1}{3}31​ and 23\dfrac{2}{3}32​ are weights: positive, with sum 111.

Weighted AM-GM

For positive a1,…,ana_1, \dots, a_na1​,…,an​ and positive weights w1,…,wnw_1, \dots, w_nw1​,…,wn​ with w1+⋯+wn=1w_1 + \dots + w_n = 1w1​+⋯+wn​=1,

w1a1+w2a2+⋯+wnan ≥ a1w1a2w2⋯anwn,\displaystyle w_1 a_1 + w_2 a_2 + \dots + w_n a_n \ \ge\ a_1^{w_1} a_2^{w_2} \cdots a_n^{w_n},w1​a1​+w2​a2​+⋯+wn​an​ ≥ a1w1​​a2w2​​⋯anwn​​,

with equality only when all the aia_iai​ are equal.

Why is this true? Take rational weights first, say w1=25w_1 = \dfrac{2}{5}w1​=52​ and w2=35w_2 = \dfrac{3}{5}w2​=53​. Write aaa twice and bbb three times, and use plain AM-GM on these five numbers:

a+a+b+b+b5≥a2b35,that is25a+35b≥a2/5b3/5.\displaystyle \dfrac{a + a + b + b + b}{5} \ge \sqrt[5]{a^2 b^3}, \quad\text{that is}\quad \dfrac{2}{5}a + \dfrac{3}{5}b \ge a^{2/5} b^{3/5}.5a+a+b+b+b​≥5a2b3​,that is52​a+53​b≥a2/5b3/5.

Any real weight is a limit of rational ones, so the inequality holds for real weights too. Course 3 gives a second proof with Jensen's inequality.

How weights find a maximum

Find the greatest value of x2yx^2 yx2y when x,y>0x, y > 0x,y>0 and x+y=9x + y = 9x+y=9.

  1. Split xxx into two equal parts. The three numbers x2,x2,y\dfrac{x}{2}, \dfrac{x}{2}, y2x​,2x​,y have sum 999, and their product is x2y4\dfrac{x^2 y}{4}4x2y​.
  2. Use AM-GM. x2y4≤(93)3=27\dfrac{x^2 y}{4} \le \left(\dfrac{9}{3}\right)^3 = 274x2y​≤(39​)3=27, so x2y≤4×27x^2 y \le 4 \times 27x2y≤4×27.
  3. Check equality. The three parts must be equal: x2=y\dfrac{x}{2} = y2x​=y, so x=6x = 6x=6 and y=3y = 3y=3. Then x2y=36×3x^2 y = 36 \times 3x2y=36×3, which reaches the bound.

The split is chosen so that the parts can all be equal at the same time. That is the whole art.

Power means

For positive numbers and p≠0p \ne 0p=0, the power mean with exponent ppp is

Mp=(a1p+a2p+⋯+anpn)1/p.\displaystyle M_p = \left(\dfrac{a_1^p + a_2^p + \dots + a_n^p}{n}\right)^{1/p}.Mp​=(na1p​+a2p​+⋯+anp​​)1/p.

So M1M_1M1​ is the arithmetic mean, M2M_2M2​ the quadratic mean and M−1M_{-1}M−1​ the harmonic mean. As p→0p \to 0p→0 the value tends to the geometric mean, so we set M0=GM_0 = GM0​=G.

The power mean inequality. If p<qp < qp<q, then Mp≤MqM_p \le M_qMp​≤Mq​, with equality only when all the numbers are equal.

For the numbers 111 and 444 the means are M−1=85M_{-1} = \dfrac{8}{5}M−1​=58​, M0=2M_0 = 2M0​=2, M1=52M_1 = \dfrac{5}{2}M1​=25​ and M2=8.5M_2 = \sqrt{8.5}M2​=8.5​. They climb, as the picture shows. Course 3 proves the inequality with Jensen.

A rising blue curve: the power mean of the numbers 1 and 4 against the exponent p, for p from -5 to 5. Four red dots sit on the curve and are labelled: HM = 1.6 at p = -1, GM = 2 at p = 0, AM = 2.5 at p = 1 and QM = 2.92 at p = 2. Two dashed grey lines show the curve levelling off towards 1 on the left and towards 4 on the right.

A useful case is p=1p = 1p=1 and q=3q = 3q=3: if positive numbers have a fixed sum SSS, then a13+⋯+an3≥S3n2a_1^3 + \dots + a_n^3 \ge \dfrac{S^3}{n^2}a13​+⋯+an3​≥n2S3​.

Bernoulli's inequality

For a whole number n≥1n \ge 1n≥1 and x≥−1x \ge -1x≥−1,

(1+x)n ≥ 1+nx.\displaystyle (1 + x)^n \ \ge\ 1 + nx.(1+x)n ≥ 1+nx.

Proof by induction. For n=1n = 1n=1 both sides are 1+x1 + x1+x. Suppose (1+x)k≥1+kx(1 + x)^k \ge 1 + kx(1+x)k≥1+kx. Multiply by 1+x1 + x1+x, which is not negative:

(1+x)k+1≥(1+kx)(1+x)=1+(k+1)x+kx2≥1+(k+1)x.\displaystyle (1 + x)^{k+1} \ge (1 + kx)(1 + x) = 1 + (k + 1)x + kx^2 \ge 1 + (k + 1)x.(1+x)k+1≥(1+kx)(1+x)=1+(k+1)x+kx2≥1+(k+1)x.

Proof by AM-GM (for x>−1x > -1x>−1): the nnn numbers 1+nx,1,1,…,11 + nx, 1, 1, \dots, 11+nx,1,1,…,1 have arithmetic mean 1+x1 + x1+x when 1+nx>01 + nx > 01+nx>0. So 1+x≥1+nxn1 + x \ge \sqrt[n]{1 + nx}1+x≥n1+nx​, and the power nnn gives Bernoulli.

The same holds for any real exponent r≥1r \ge 1r≥1. For 0≤r≤10 \le r \le 10≤r≤1 the sign turns round: (1+x)r≤1+rx(1 + x)^r \le 1 + rx(1+x)r≤1+rx. A-Level books often meet this as the first two terms of the binomial series.

Practise with new numbers

Try it in Python

The program prints the power means of a list for several exponents, so that you can see them climb. It then tests Bernoulli's inequality for many values of xxx and nnn.

Read it first. Then run it. Do the power means rise as p rises?

1. Make all the numbers in data equal. What happens to the column of power means?

2. Change x = -0.5 to x = -4. For which values of n does Bernoulli's inequality fail now? Which condition of the inequality is broken?

Common mistakes

  • Weights that do not add up to 1. 12a+13b≥a1/2b1/3\dfrac{1}{2}a + \dfrac{1}{3}b \ge a^{1/2}b^{1/3}21​a+31​b≥a1/2b1/3 is false in general. Scale the weights first.
  • A split that cannot be equal. Splitting x2yx^2 yx2y as x⋅x⋅yx \cdot x \cdot yx⋅x⋅y with x+x+yx + x + yx+x+y fixed is not the same problem as x+yx + yx+y fixed.
  • Forgetting x≥−1x \ge -1x≥−1 in Bernoulli. For x=−4x = -4x=−4 and n=3n = 3n=3, (1+x)3=−27(1 + x)^3 = -27(1+x)3=−27 but 1+3x=−111 + 3x = -111+3x=−11. The inequality fails, because the proof multiplied by 1+x1 + x1+x, which is negative here.
  • Taking M0M_0M0​ from the formula. The formula has 1p\dfrac{1}{p}p1​, so p=0p = 0p=0 is defined by the limit: M0M_0M0​ is the geometric mean.

Remember these

  • Weighted AM-GM: ∑wiai≥∏aiwi\sum w_i a_i \ge \prod a_i^{w_i}∑wi​ai​≥∏aiwi​​ when ∑wi=1\sum w_i = 1∑wi​=1.
  • M−1≤M0≤M1≤M2M_{-1} \le M_0 \le M_1 \le M_2M−1​≤M0​≤M1​≤M2​, that is H≤G≤A≤QH \le G \le A \le QH≤G≤A≤Q.
  • Bernoulli: (1+x)n≥1+nx(1 + x)^n \ge 1 + nx(1+x)n≥1+nx for x≥−1x \ge -1x≥−1.
  • Split a term so that all parts can be equal at once.

Check yourself

Try each one on your own first. If you are stuck, open a hint. Then check with Show the answer.

  1. Find the weighted mean of 505050 and 808080 with weights 35\dfrac{3}{5}53​ and 25\dfrac{2}{5}52​.
  1. Find the greatest value of xy2xy^2xy2 when x,y>0x, y > 0x,y>0 and x+y=6x + y = 6x+y=6.
  1. Find M−1M_{-1}M−1​, M1M_1M1​ and M2M_2M2​ for the numbers 333 and 666.
  1. Use Bernoulli's inequality to give a lower bound for 1.02301.02^{30}1.0230.
  1. Four positive numbers add up to 888. What is the least value of the sum of their cubes?

References

  • Hardy, Littlewood and Pólya, Inequalities, Chapter 2, on weighted and power means.
  • Steele, The Cauchy-Schwarz Master Class, Chapter 2, on the AM-GM inequality and its weighted form.

Enroll free in The AM-GM-HM Inequality and Power Means

Every lesson is open to read. Enrolling also lets you:

  • Mark lessons done and carry on where you stopped.
  • Take the tests of this lesson and its course.
  • Count the course towards the certificate for its chapter.
No payment. You can leave at any time.

Problems (3)

A company has a budget that it splits between advertising and research…Hard · High School (11–12) · by StemPy Lab
A bank pays 3% interest each year, added once a year, so an amount grows by the factor 1.03²⁰ in 20…Hard · High School (11–12) · by StemPy Lab
Three positive numbers a, b and c add up to 6…Advanced · High School (11–12) · by StemPy Lab
Next: lesson 3Greatest and least values with the means
← The AM-GM-HM Inequality and Power Means
← Arithmetic, geometric and harmonic meansGreatest and least values with the means →

Discussion (0)

Log in to join the discussion.

Loading…

when the weights are positive and add up to
111
. The power mean with exponent
ppp
grows as
ppp
grows, so the harmonic, geometric, arithmetic and quadratic means sit on one rising curve. Bernoulli's inequality
(1+x)n≥1+nx(1 + x)^n \ge 1 + nx(1+x)n≥1+nx
gives a quick lower bound for a power.
about 25 min1 simulation1 practice question1 Python program5 questions to check yourself

This lesson is free to read. Enroll free to keep your place and take the tests.