uniolympiads Ελληνικά

Classical inequalities

AM-GM, power means, rearrangement, Chebyshev, Cauchy–Schwarz, Hölder, Minkowski, Jensen and the tangent line, extrema of convex functions at the endpoints, Karamata, smoothing.

What to know by heart

Throughout, xi,ai,bi≥0x_i, a_i, b_i \ge 0 unless stated otherwise.

  • AM-GM. x1+⋯+xnn≥x1⋯xnn\dfrac{x_1 + \dots + x_n}{n} \ge \sqrt[n]{x_1 \cdots x_n}, with equality only when x1=⋯=xnx_1 = \dots = x_n. With weights wi≥0w_i \ge 0, ∑wi=1\sum w_i = 1:  ∑wixi≥∏xiwi\ \sum w_i x_i \ge \prod x_i^{w_i}.
  • Power means. For xi>0x_i > 0 and p≠0p \ne 0 put Mp=(1n∑xip)1/pM_p = \big(\frac1n \sum x_i^p\big)^{1/p} and M0=x1⋯xnnM_0 = \sqrt[n]{x_1 \cdots x_n}. The map p↦Mpp \mapsto M_p is increasing. In particular  min⁡xi≤M−1≤M0≤M1≤M2≤max⁡xi\ \min x_i \le M_{-1} \le M_0 \le M_1 \le M_2 \le \max x_i (harmonic ≤\le geometric ≤\le arithmetic ≤\le quadratic).
  • Rearrangement. If a1≤⋯≤ana_1 \le \dots \le a_n and b1≤⋯≤bnb_1 \le \dots \le b_n (signs do not matter here), then for every permutation σ\sigma
    ∑iaibn+1−i≤∑iaibσ(i)≤∑iaibi.\sum_{i} a_i b_{n+1-i} \le \sum_i a_i b_{\sigma(i)} \le \sum_i a_i b_i .
  • Chebyshev. If (ai)(a_i), (bi)(b_i) are similarly ordered,  n∑aibi≥(∑ai)(∑bi)\ n \sum a_i b_i \ge \big(\sum a_i\big)\big(\sum b_i\big). For integrals: if f,gf, g are both increasing on [a,b][a, b],  (b−a)∫abfg≥∫abf∫abg\ (b - a)\int_a^b fg \ge \int_a^b f \int_a^b g.
  • Cauchy–Schwarz. (∑aibi)2≤∑ai2∑bi2\big(\sum a_i b_i\big)^2 \le \sum a_i^2 \sum b_i^2 and (∫fg)2≤∫f2∫g2\big(\int fg\big)^2 \le \int f^2 \int g^2 (for real ai,bia_i, b_i and functions, with no sign restriction).
  • Hölder. For p,q>1p, q > 1 with 1p+1q=1\frac1p + \frac1q = 1:  ∑aibi≤(∑aip)1/p(∑biq)1/q\ \sum a_i b_i \le \big(\sum a_i^p\big)^{1/p} \big(\sum b_i^q\big)^{1/q}, and ∫∣fg∣≤∥f∥p∥g∥q\int |fg| \le \|f\|_p \|g\|_q.
  • Minkowski. For p≥1p \ge 1:  (∑(ai+bi)p)1/p≤(∑aip)1/p+(∑bip)1/p\ \big(\sum (a_i + b_i)^p\big)^{1/p} \le \big(\sum a_i^p\big)^{1/p} + \big(\sum b_i^p\big)^{1/p}, and ∥f+g∥p≤∥f∥p+∥g∥p\|f + g\|_p \le \|f\|_p + \|g\|_p.
  • Jensen. If φ\varphi is convex on an interval II, xi∈Ix_i \in I and wi≥0w_i \ge 0 with ∑wi=1\sum w_i = 1, then φ(∑wixi)≤∑wiφ(xi)\varphi\big(\sum w_i x_i\big) \le \sum w_i \varphi(x_i). For integrals: if f ⁣:[0,1]→If\colon [0, 1] \to I is continuous,  φ(∫01f)≤∫01φ(f)\ \varphi\big(\int_0^1 f\big) \le \int_0^1 \varphi(f). For concave φ\varphi the direction is reversed.
  • The one-line inequalities. ex≥1+xe^x \ge 1 + x,  ln⁡x≤x−1\ \ln x \le x - 1,  sin⁡x≤x\ \sin x \le x for x≥0x \ge 0,  2ab≤a2+b2\ 2ab \le a^2 + b^2. They all say that a convex (concave) function lies above (below) its tangent line.

The basic lemmas, with short proofs

AM-GM from et≥1+te^t \ge 1 + t. Let A=1n∑xi>0A = \frac1n \sum x_i > 0. From et≥1+te^{t} \ge 1 + t with t=xiA−1t = \frac{x_i}{A} - 1 we get xiA≤exi/A−1\frac{x_i}{A} \le e^{x_i/A - 1}. Multiplying,

x1⋯xnAn≤e∑(xi/A−1)=e0=1.\frac{x_1 \cdots x_n}{A^n} \le e^{\sum (x_i/A - 1)} = e^{0} = 1 .

Since et=1+te^t = 1 + t only for t=0t = 0, equality holds only if xi=Ax_i = A for every ii.

Rearrangement by swapping. If a permutation has i<ji < j with σ(i)>σ(j)\sigma(i) > \sigma(j), swapping bσ(i)b_{\sigma(i)} and bσ(j)b_{\sigma(j)} changes the sum by

(aibσ(j)+ajbσ(i))−(aibσ(i)+ajbσ(j))=(aj−ai)(bσ(i)−bσ(j))≥0.\big(a_i b_{\sigma(j)} + a_j b_{\sigma(i)}\big) - \big(a_i b_{\sigma(i)} + a_j b_{\sigma(j)}\big) = (a_j - a_i)\big(b_{\sigma(i)} - b_{\sigma(j)}\big) \ge 0 .

Each swap reduces the number of inversions, so after finitely many steps we reach the identity without decreasing the sum. The lower bound follows by applying the same to −bn+1−i-b_{n+1-i}.

Chebyshev by a double sum. Since (ai)(a_i), (bi)(b_i) are similarly ordered, every term of ∑i,j(ai−aj)(bi−bj)\sum_{i, j} (a_i - a_j)(b_i - b_j) is ≥0\ge 0. Expanding, the sum equals 2n∑aibi−2∑ai∑bi2n \sum a_i b_i - 2 \sum a_i \sum b_i. For integrals, the same with ∬(f(x)−f(y))(g(x)−g(y)) dx dy≥0\iint (f(x) - f(y))(g(x) - g(y))\,dx\,dy \ge 0.

Cauchy–Schwarz by the discriminant. The quantity ∑(ait+bi)2=t2∑ai2+2t∑aibi+∑bi2\sum (a_i t + b_i)^2 = t^2 \sum a_i^2 + 2t \sum a_i b_i + \sum b_i^2 is ≥0\ge 0 for every tt, so the discriminant is ≤0\le 0. For integrals see also the integrals sheet.

Hölder from Young. For u,v≥0u, v \ge 0 we have uv≤upp+vqquv \le \frac{u^p}{p} + \frac{v^q}{q} (weighted AM-GM with weights 1p,1q\frac1p, \frac1q on up,vqu^p, v^q). If A=(∑aip)1/pA = \big(\sum a_i^p\big)^{1/p} and B=(∑biq)1/qB = \big(\sum b_i^q\big)^{1/q} are positive, put u=ai/Au = a_i/A, v=bi/Bv = b_i/B and sum:  ∑aibiAB≤1p+1q=1\ \frac{\sum a_i b_i}{AB} \le \frac1p + \frac1q = 1.

Minkowski from Hölder. For p>1p > 1 write ∑(ai+bi)p=∑ai(ai+bi)p−1+∑bi(ai+bi)p−1\sum (a_i + b_i)^p = \sum a_i (a_i + b_i)^{p-1} + \sum b_i (a_i + b_i)^{p-1} and apply Hölder to each sum. Since (p−1)q=p(p - 1)q = p, you get S≤(∥a∥p+∥b∥p)S1/qS \le \big(\|a\|_p + \|b\|_p\big) S^{1/q} with S=∑(ai+bi)pS = \sum (a_i + b_i)^p, and you divide by S1/qS^{1/q}.

Jensen from a supporting line. A convex φ\varphi has a supporting line at every interior point mm of II: there is kk with φ(x)≥φ(m)+k(x−m)\varphi(x) \ge \varphi(m) + k(x - m) for every x∈Ix \in I (if φ\varphi is differentiable, k=φ′(m)k = \varphi'(m)). With m=∑wixim = \sum w_i x_i, multiply by wiw_i and sum: the terms k(xi−m)k(x_i - m) add up to 00. For integrals put x=f(t)x = f(t) and integrate.

The tangent line trick. If you want ∑φ(xi)≥nφ(m)\sum \varphi(x_i) \ge n \varphi(m) under ∑xi=nm\sum x_i = nm, but φ\varphi is not convex everywhere, it is enough to show φ(x)≥φ(m)+φ′(m)(x−m)\varphi(x) \ge \varphi(m) + \varphi'(m)(x - m) only for the xx in the domain. Usually the difference is a fraction with a factor (x−m)2(x - m)^2. Summing, the linear term disappears.

Convex function: maximum at the endpoints. Every x∈[a,b]x \in [a, b] can be written x=(1−λ)a+λbx = (1 - \lambda) a + \lambda b with λ∈[0,1]\lambda \in [0, 1], so

φ(x)≤(1−λ)φ(a)+λφ(b)≤max⁡{φ(a),φ(b)}.\varphi(x) \le (1 - \lambda)\varphi(a) + \lambda \varphi(b) \le \max\{\varphi(a), \varphi(b)\} .

Likewise a concave function attains its minimum at an endpoint. In several variables: if the expression is convex (or linear) in each variable separately on a box ∏[ai,bi]\prod [a_i, b_i], the maximum is attained at a vertex of the box. Keep the others fixed and push one variable at a time to an endpoint.

Karamata, briefly. We say that x=(x1≥⋯≥xn)x = (x_1 \ge \dots \ge x_n) majorises y=(y1≥⋯≥yn)y = (y_1 \ge \dots \ge y_n) if x1+⋯+xk≥y1+⋯+ykx_1 + \dots + x_k \ge y_1 + \dots + y_k for every kk, with equality for k=nk = n. Then for convex φ\varphi we have ∑φ(xi)≥∑φ(yi)\sum \varphi(x_i) \ge \sum \varphi(y_i). Example: (2,0)(2, 0) majorises (1,1)(1, 1), and two-point Jensen is a special case. The practical consequence: with a fixed sum and xi∈[a,b]x_i \in [a, b], ∑φ(xi)\sum \varphi(x_i) for convex φ\varphi is maximised when as many xix_i as possible sit at the endpoints aa, bb (at most one in between).

Smoothing. If replacing two variables xi,xjx_i, x_j by their mean xi+xj2\frac{x_i + x_j}{2} keeps the constraint and does not increase the expression, the minimum is at the point with all variables equal. Mind the logic: first make sure that a minimum point exists (continuity on a compact set), then show that if they are not all equal, a smoothing step gives a strictly smaller value.

Typical tricks

  • Find the equality case first. It tells you how to split the terms in AM-GM. E.g. for x>0x > 0, write x2+16xx^2 + \frac{16}{x} as x2+8x+8xx^2 + \frac{8}{x} + \frac{8}{x}, so that the product is constant and equality (x=2x = 2) is attainable.
  • Homogenise and normalise. With a constraint such as abc=1abc = 1 or a+b+c=3a + b + c = 3 you make both sides of the same degree, or conversely, in a homogeneous inequality you choose the scale that suits you.
  • Squares (SOS). Write the difference as a sum of squares, e.g. 3(a2+b2+c2)−(a+b+c)2=(a−b)2+(b−c)2+(c−a)23(a^2 + b^2 + c^2) - (a + b + c)^2 = (a - b)^2 + (b - c)^2 + (c - a)^2.
  • Cauchy–Schwarz with the constant 11. (∫abf)2≤(b−a)∫abf2\big(\int_a^b f\big)^2 \le (b - a)\int_a^b f^2. For sums, the Engel form: ∑ui2vi≥(∑ui)2∑vi\sum \frac{u_i^2}{v_i} \ge \frac{(\sum u_i)^2}{\sum v_i} for vi>0v_i > 0.
  • Convexity without a second derivative. Sums of convex functions, a convex function composed with a linear one, and the max⁡\max of convex functions are convex. t\sqrt{t} is concave, tln⁡tt \ln t is convex.
  • One-variable minimisation: if you end with αc+βc\alpha c + \frac{\beta}{c} with α,β,c>0\alpha, \beta, c > 0, AM-GM gives ≥2αβ\ge 2\sqrt{\alpha\beta} with no derivative.

How it appears in the IMC

The IMC rarely asks for a "pure" three-variable inequality. Usually the inequality is one step inside an analysis or linear algebra problem:

  • IMC 2026, Day 2, Problem 7: an area is bounded below by an integral depending on one parameter, and the minimisation is two-term AM-GM.
  • IMC 2026, Day 2, Problem 8: the sum of the ∣λi∣|\lambda_i| is written as a strictly concave function of one parameter on a closed interval. The minimum is at an endpoint, and it remains to compare two values.
  • IMC 2026, Day 1, Problem 1: the crucial bound is ∣sin⁡x±cos⁡x∣≤2|\sin x \pm \cos x| \le \sqrt{2}, together with 2<π2\sqrt{2} < \frac{\pi}{2}.
  • IMC 2025, Day 1, Problem 2: Cauchy–Schwarz for integrals gives the minimum of some ∫(f′′)2\int (f'')^2 with boundary conditions.

Bridges to these problems are the exercises of the AM-GM and convexity ladders.

Pitfalls

  • AM-GM needs non-negative numbers. For x=−1x = -1, x+1x≥2x + \frac1x \ge 2 is false.
  • Equality must be attainable. x2+16x≥216xx^2 + \frac{16}{x} \ge 2\sqrt{16x} is true but useless, since the right-hand side depends on xx.
  • Direction of Jensen. Convex: the value at the mean is below the mean of the values. Check with two points before using it.
  • Smoothing without a minimum proves nothing. Infinitely many steps that decrease a quantity need not reach the point you want.
  • In Hölder the exponents are conjugate. With other exponents the inequality fails in general: for p=q=4p = q = 4 and ai=bi=1a_i = b_i = 1 (n≥2n \ge 2 terms) it would give n≤nn \le \sqrt{n}.

Further reading

  • Gelca & Andreescu, Putnam and Beyond, §2.1.3 (The Cauchy–Schwarz Inequality)
  • Gelca & Andreescu, Putnam and Beyond, §3.2.10 (Inequalities for Integrals)

Contest problems using these tools

No problems with these tools yet.