One rule: a set that holds its own chords
Pick two points anywhere inside a set and draw the segment between them. If that whole segment stays inside the set — no matter which two points you chose — the set is convex. That is the entire definition, and it is worth letting it sink in before piling anything on top. A filled disk is convex; so is a filled triangle, a regular hexagon, a solid cube, a half-plane. A crescent moon is not, because a chord across the two horns slips out through the bite. A five-pointed star is not, for the same reason: the segment joining two adjacent points cuts across the empty notch between them.
The earlier rungs already handed you the word for the polygon case: a polygon is convex or concave exactly according to whether every diagonal stays inside. Convexity simply lifts that test off polygons and applies it to any shape at all, in any number of dimensions, with no need for straight edges or corners. And it behaves beautifully under intersection: if you overlap two convex sets, the common part is convex again, because a chord trapped inside both is trapped inside the overlap. A half-plane is convex; intersect a stack of them and you carve out exactly the convex polygons and polyhedra — the shapes that are 'cut from flat slabs'.
The convex hull: the shrink-wrap of a point set
Scatter a handful of nails on a board and stretch a rubber band around them all; let go, and it snaps taut against the outermost nails, enclosing every point in the smallest convex region possible. That taut outline is the convex hull of the point set. Made precise, the convex hull is the intersection of every convex set that contains your points — and since intersections of convex sets are convex, the result is itself the smallest convex set wrapping them. In the plane it is a convex polygon; in space, a convex polyhedron; the rubber band becomes a shrink-wrapped membrane.
There is a second, dual way to see the hull that pays off enormously later. Every point of the hull can be written as a weighted average — a convex combination — of the original points, where the weights are non-negative and sum to 1. Two points give you the whole segment between them (weights t and 1 - t); three non-collinear points give the filled triangle they span; n points give everything you can reach by mixing them with non-negative weights that total one. So the hull is, at once, the smallest convex container and the set of all honest averages of the data. Both descriptions matter, and one striking fact reconciles them: in the plane, every point of the hull is already an average of at most three of the original points, never needing more.
Supporting hyperplanes: leaning a flat wall against a convex body
Convexity has a second face, just as fundamental as the chord rule, and it is the one optimization lives on. Take a convex body and slide a flat wall — a line in the plane, a plane in space, a hyperplane in higher dimensions — straight up against it until the wall just touches and goes no further. The whole body then sits entirely on one side of that wall. A wall that touches a convex set while leaving all of it on one side is a supporting hyperplane. Every boundary point of a convex body has at least one such wall leaning on it; at a smooth point the wall is the unique tangent, and at a sharp corner a whole fan of walls can rest there.
From supporting walls springs the single most useful theorem about convex sets: the separating-hyperplane theorem. If two convex sets do not overlap, you can always slide a single flat wall between them so that one set lies entirely on one side and the other set entirely on the opposite side. Picture two disjoint blobs of dough on a table; convexity guarantees a clean knife-cut — a straight line — separating them with no overlap. This sounds obvious for two circles, but it holds in every dimension and for every pair of disjoint convex sets, and its consequences run deep: a yes-or-no question about overlap becomes the existence of one separating wall, the very idea that powers linear classifiers, support-vector machines, and the duality at the heart of linear programming.
These two faces — the inside view (every chord stays in) and the outside view (a wall leans on every boundary point) — are not two coincidences but the same fact seen from within and without. A convex body is exactly the intersection of all the half-spaces cut off by its supporting walls. That is why a convex polyhedron can be described two completely equivalent ways: list its corner points and take their hull, or list its bounding walls and intersect the half-spaces. Translating between those two descriptions, the vertex picture and the face picture, is one of the recurring labours of computational geometry — and, honestly, it can be expensive, because a polytope with few vertices may have astronomically many faces, and the reverse.
Polytopes: polygons and polyhedra, set free into any dimension
A line segment, a polygon, a polyhedron — each is a convex shape whose boundary is built from flat pieces, and each lives one dimension higher than the last. A polytope is what you get when you refuse to stop at three dimensions: the convex hull of finitely many points, or equally the bounded intersection of finitely many half-spaces, in any dimension d you like. In 2D a polytope is just a convex polygon; in 3D, a convex polyhedron; in 4D and beyond it is a genuine object you can compute with even though you cannot fully picture it. The whole machinery — vertices, edges, flat faces, the hull view and the half-space view — carries upward unchanged.
To see higher dimensions without picturing them, count instead. A square has 4 corners and 4 edges; a cube has 8 corners, 12 edges, 6 square faces. Step up one floor to the 4-dimensional cube, the tesseract: it has 16 corners, 32 edges, 24 square faces, and 8 cubical 'cells' as its three-dimensional boundary pieces. You cannot hold a tesseract in your hands, but you can hold its census of parts, and the pattern that generates those numbers (each new dimension doubles the old corners and adds a copy of the old cube) is perfectly rigorous. This is the quiet liberation of the whole subject: dimension stops being a place you must visualize and becomes a number you compute with.
Counting the parts (the f-vector), dimension by dimension: shape dim vertices edges 2D-faces 3D-cells segment 1 2 - - - square 2 4 4 - - cube 3 8 12 6 - tesseract 4 16 32 24 8 Euler check (must equal 2 for a convex polyhedron): cube: V - E + F = 8 - 12 + 6 = 2
Euler's formula and the surprises of high dimension
Counting parts is not idle bookkeeping; the counts obey laws. The oldest is Euler's polyhedron formula, V - E + F = 2 for any convex polyhedron, where V, E, F are the numbers of vertices, edges, and flat faces. Check it on a cube: 8 - 12 + 6 = 2. On a tetrahedron: 4 - 6 + 4 = 2. It never fails for a convex solid, and that stubborn 2 is no coincidence — it is the Euler characteristic chi = V - E + F of the sphere, the surface every convex polyhedron is topologically. The formula is your first bridge from the metric world of lengths and angles to the topological world of the final guides in this rung, where shape matters but size does not.
High dimension does not merely extend our intuitions — it ambushes them, and honesty demands we say so. In three dimensions there are exactly five Platonic solids, the perfectly regular convex polyhedra: tetrahedron, cube, octahedron, dodecahedron, icosahedron. You might guess the count keeps climbing with dimension. It does the opposite: in 4D there are six regular polytopes, but in dimension 5 and every dimension above it there are only ever three — the simplex, the hypercube, and its dual. The rich variety of regular shapes is a low-dimensional luxury that high dimension quietly takes away. Stating that plainly is more useful than any tidy fable in which patterns always grow.
Why convexity is the hinge of the chapter
Convexity earns its central seat because of one practical miracle: on a convex set, local good news is automatically global good news. If you are minimizing a sensible (convex) cost over a convex region and you reach a point where you cannot improve by any small step, you are done — that point is the true minimum, with no hidden lower valley lurking elsewhere. Drop the convexity and that guarantee evaporates: a non-convex landscape can trap you in a shallow dip while a deeper one waits over the ridge. This is the precise reason convex problems are the ones we can solve reliably and at scale, and why so much of applied mathematics works hard to phrase a problem convexly in the first place.
Look back at the trail. The chord rule gave us convex sets; the rubber band gave us the convex hull and its dual life as averages of data; the leaning wall gave us supporting and separating hyperplanes; the refusal to stop at 3D gave us polytopes and the discipline of counting what we cannot see. Each idea is elementary on its own, yet together they are the working language of optimization, of geometric algorithms, and of the data-shaped geometry of our century. The next guide cashes all of this in: it hands the convex hull and its relatives to a computer and asks how fast a machine can actually build them.