The Isoperimetric Inequality: Why the Circle Encloses the Most Area

What the inequality says, and what it needs

Take a closed curve in the plane that never crosses itself — a simple closed curve, also called a Jordan curve. Write L for its length and A for the area of the region it encloses. Then

L² ≥ 4πA,

with equality if and only if the curve is a circle. The word isoperimetric is Greek for "same perimeter": among all loops of one fixed length, the circle traps the most area. The statement is symmetric in a useful way — maximising A with L held fixed and minimising L with A held fixed are the same problem, because the inequality couples them.

Three hypotheses are carrying weight. Closed, or there is no interior to measure. Simple, so that the Jordan curve theorem hands you a well-defined inside with a well-defined area. And rectifiable — the curve must have finite length in the first place. Drop rectifiability and the inequality does not become false, it becomes empty: the Koch snowflake has infinite perimeter and finite area, so L² ≥ 4πA holds trivially and says nothing at all. The modern statement, due to Ennio De Giorgi in 1958, replaces "curve" with a set of finite perimeter and "length" with perimeter in the sense of Caccioppoli, which covers fractal-ish boundaries, sets with holes, and the limits of minimising sequences all at once.

The quantity Q = 4πA / L² is the isoperimetric quotient. It is dimensionless and scale-invariant: double a shape and A quadruples while L doubles, so Q does not move. Q therefore measures pure shape, with Q = 1 for the circle and Q < 1 for everything else. A square gives Q = π/4 = 0.7854. Concretely, 100 m of fence bent into a circle encloses 100²/(4π) = 795.77 m²; the same fence bent into a square encloses 25² = 625 m². The circle wins by 170.8 m², a factor of exactly 4/π = 1.2732.

Dido's rope, and two thousand years of almost-proofs

The problem has a founding myth. In Virgil's Aeneid, Dido flees Tyre, lands in North Africa and is offered as much land as she can enclose with a single oxhide. She cuts the hide into a long thin strip and lays it out to enclose the ground that became Carthage. The mathematical version — Dido's problem — is the isoperimetric problem with a twist: because the sea provides a free straight boundary, the optimal curve is a semicircle with its diameter on the coast. With 100 m of rope, the free-standing circle encloses 795.8 m² but the semicircle against a coastline encloses L²/(2π) = 1591.5 m² — exactly double, because you get a full circle for free by reflecting in the shoreline.

The first serious mathematics is Zenodorus (c. 200–140 BC), whose lost treatise On Isometric Figures survives through quotations in Theon of Alexandria and Pappus. He proved two genuine theorems: among polygons with a given number of sides and a given perimeter, the regular one has the largest area; and a circle beats any polygon of the same perimeter. What he could not do — what nobody could do for another two millennia — was handle all curves at once.

Jakob Steiner published five arguments in 1838, including the symmetrisation and hinge constructions below. Each one shows that a non-circular shape can be improved. Each one therefore proves "if a best shape exists, it is the circle" — and none of them proves that a best shape exists. Weierstrass pressed exactly this point in his Berlin lectures in the 1870s, with the standard cautionary analogue: the same style of argument "proves" that 1 is the largest positive integer, since for any n > 1 the number n² is larger. The existence gap is not pedantry; it is the whole difficulty.

Step one: the best shape has no dents

This is the step the animation performs, and it is exact rather than suggestive. Suppose the boundary contains an arc γ running between two boundary points p and q, and suppose γ caves inward: it lies on the interior side of the chord pq. Let D be the region trapped between the chord and the arc — the bite taken out of the shape.

Now reflect γ across the line pq, leaving the rest of the boundary alone. Reflection is an isometry, so the reflected arc has exactly the same length as γ and the endpoints p and q do not move: the perimeter is unchanged, to the last decimal. But the bite D is no longer removed, and its mirror image is now added. The area goes up by exactly 2·area(D) — strictly up, since D has positive area whenever the dent is real. In the video the dented loop measures 2.372 square units and the un-dented one 2.903, on a string of length 6.283 that never changes.

The conclusion is that no shape with a dent can be optimal, so the maximiser must be convex. One caveat keeps this honest: reflecting a dent outward can, in awkward cases, make the new arc collide with a distant part of the boundary. The airtight version of the same idea avoids that entirely — replace the region K by its convex hull. The convex hull has area at least as large (strictly larger if K was not convex) and perimeter no larger (strictly smaller if K was not convex), so scaling the hull back up to perimeter L increases the area a second time. Either way, convexity is forced.

Step two: Steiner's four-hinge argument

Assume for the moment that a maximiser K exists; by the previous step it is convex. Draw a chord PQ that cuts the boundary into two arcs of equal length L/2. That chord must also cut the area in half: if one side held more area, you could throw the smaller side away, reflect the larger side across the chord, and get a region with the same perimeter and strictly more area.

Now work with one half — the arc from P to Q, with the chord closing it off. Pick any point X on the arc. The half splits into the triangle PXQ plus two circular caps sitting on the segments XP and XQ. Hinge the figure at X: keep the two caps rigid and change the angle θ = ∠PXQ. The two arc lengths do not change, and neither do the cap areas, and since K is rebuilt as the half plus its mirror image, K's perimeter is twice the arc length and does not change either. Only the triangle changes, and its area is

½ · |XP| · |XQ| · sin θ,

which is largest exactly when θ = 90°. So in the maximiser, every point X on the arc sees the chord PQ at a right angle — and by the converse of Thales' theorem, the locus of such points is the circle with diameter PQ. The arc is a semicircle, and K is a disc.

That is a complete and correct argument for a conditional statement: if a maximiser exists, it is the disc. The existence half was supplied later by compactness. Wilhelm Blaschke's selection theorem (1916) says that any sequence of convex bodies inside a fixed bounded region has a subsequence converging in the Hausdorff metric; since area is continuous and perimeter is upper semicontinuous on that space, a maximising sequence of convex sets with perimeter L has a limit that attains the maximum. Only then does Steiner's argument become a proof.

Hurwitz's proof: Fourier series settle it in one page

The first fully rigorous proof that does not have to argue about existence at all is Adolf Hurwitz's of 1901, and it is short enough to give in full. Parametrise the curve by arc length s, then rescale to t = 2πs/L so that t runs over [0, 2π]. Constant speed means

ẋ² + ẏ² = (L / 2π)², so   ∫₀^2π (ẋ² + ẏ²) dt = L² / (2π).

Green's theorem gives the enclosed area as A = ½∫(x ẏ − y ẋ) dt, and integrating by parts over a full period turns that into A = ∫ x ẏ dt. Translate the curve so that x has mean zero — a translation changes neither L nor A. Then

L² − 4πA = 2π ∫ (ẋ² + ẏ² − 2 x ẏ) dt = 2π ∫ (ẏ − x)² dt + 2π ∫ (ẋ² − x²) dt.

The first integral is obviously non-negative. The second is non-negative by Wirtinger's inequality: for a 2π-periodic, absolutely continuous function f with mean zero, ∫ḟ² ≥ ∫f². That is a one-line consequence of Parseval — if f has Fourier coefficients a_n, b_n, then ∫f² = π∑(a_n² + b_n²) while ∫ḟ² = π∑n²(a_n² + b_n²), and n² ≥ 1 for every harmonic that is present. Hence L² ≥ 4πA.

The equality case falls out of the same computation, which is what makes this proof so satisfying. Equality forces both integrals to vanish. The second vanishing means every harmonic above the first is zero, so x = a cos t + b sin t; the first means ẏ = x, so y = a sin t − b cos t + c. That is a circle of radius √(a² + b²) — no case analysis, no appeal to existence, and the only regularity needed is enough to write down the Fourier series, which rectifiability supplies.

Higher dimensions, curved spaces, and discrete cousins

In n dimensions the inequality relates the perimeter (surface measure) P of a set to its volume V:

Pⁿ ≥ nⁿ ω_n Vⁿ⁻¹,   where ω_n = π^n/2 / Γ(n/2 + 1) is the volume of the unit ball,

with equality only for balls. For n = 2 this is ω₂ = π and L² ≥ 4πA. For n = 3 it is A³ ≥ 36πV²: a sphere of radius r has A = 4πr² and V = (4/3)πr³, and both sides come out to 64π³r⁶. Hermann Schwarz proved the three-dimensional case in 1884; Erhard Schmidt gave a proof valid in every dimension in 1938–39, and the slickest modern route is two lines from the Brunn–Minkowski inequality, |X + Y|^1/n ≥ |X|^1/n + |Y|^1/n, applied to X = the set and Y = a small ball.

Curvature changes the constant in a precise way. On the unit sphere, a curve of length L bounding a region of area A satisfies L² ≥ A(4π − A), with equality for spherical caps — note that a great circle gives L = 2π and A = 2π, and both sides equal 4π². In the hyperbolic plane of curvature −1 the inequality becomes L² ≥ 4πA + A², which is why hyperbolic discs have perimeter growing like their area. The general pattern, due to Weil (1926), Bol (1941) and Alexandrov, is that a simply connected surface with Gaussian curvature K ≤ κ obeys L² ≥ 4πA − κA².

Discretely, the regular n-gon of perimeter L has quotient Q = π / (n tan(π/n)): 0.6046 for a triangle, 0.7854 for a square, 0.9069 for a hexagon, 0.9770 for a 12-gon, 0.9997 for a 100-gon — climbing to 1 as the polygon rounds off. The hexagon's 0.9069 is the reason honeycombs are hexagonal, a statement made precise by Thomas Hales' 2001 proof of the honeycomb conjecture: among all ways of partitioning the plane into regions of equal area, the regular hexagonal tiling has the least total perimeter. On graphs the same instinct becomes the Cheeger constant, and isoperimetric-style bounds there control mixing times and spectral gaps.

How close is close? Stability, and where the inequality bites

Knowing that L² − 4πA ≥ 0 is only half a theorem; the useful question is what a small deficit tells you. Bonnesen's inequality (Tommy Bonnesen, 1921) answers it for convex plane regions:

L² − 4πA ≥ π²(R − r)²,

where R and r are the circumradius and inradius. Since R = r exactly for a disc, a near-zero deficit forces the region to be sandwiched between two nearly equal concentric circles. For the 100 m square: R = 17.68 m, r = 12.50 m, so the bound reads 2146.0 ≥ 264.6 — true, and comfortably so, which is the usual state of affairs for a shape that is not close to round.

The n-dimensional sharp version took until 2008. Define the Fraenkel asymmetry α(E) as the smallest relative volume of the symmetric difference between E and any ball of the same volume. Fusco, Maggi and Pratelli proved in the Annals of Mathematics that the isoperimetric deficit dominates α(E)² up to a dimensional constant, and that the exponent 2 cannot be improved. In words: a set whose surface area is within 1 % of optimal must genuinely look like a ball, not merely have the right numbers.

The inequality is also a law of physics in disguise. Surface tension is an energy proportional to surface area, so a free droplet or a soap bubble minimises area at fixed volume and therefore becomes a sphere; the double bubble theorem (Hutchings, Morgan, Ritoré and Ros, 2002) settles the corresponding two-volume problem. The same arithmetic explains why pipes and arteries are round — a circular bore carries the most cross-section per unit of wall material — why cells and vesicles pull toward spheres unless a cytoskeleton resists, and why Dido cut her oxhide into a strip rather than laying it out flat.

One last honest note about the video. Bending a fixed loop through an oval, a square, a triangle and a star and watching the area meter fall is an illustration: it verifies the inequality on five shapes, not on all of them. The dent-reflection beat is different — that one is a genuine step of the real proof, exact to the last decimal, and it is the step that forces the answer to be convex. Everything after that is Steiner's hinge argument plus a compactness theorem, or Hurwitz's page of Fourier series.

Every shape below is built from the same 100 metres of fence, so only the shape changes. Q = 4πA/L² is the isoperimetric quotient, which equals 1 only for the circle.

Shape (same 100 m of fence)	Isoperimetric quotient Q	Area enclosed	Area lost vs. the circle
Circle	1 exactly	795.77 m²	0 %
Regular hexagon	0.9069 = π / (6 tan 30°)	721.69 m²	9.3 %
Ellipse, axes in 2:1 ratio	0.8412	669.38 m²	15.9 %
Square	0.7854 = π/4	625.00 m²	21.5 %
Equilateral triangle	0.6046 = π / (3√3)	481.13 m²	39.5 %