Convex analysis is the mathematics of problems in which local structure can control global behaviour. It studies convex sets, convex functions, supporting hyperplanes, subgradients, conjugates and duality—the structures that make a large class of optimisation problems unusually transparent and computationally reliable.
A convex problem does not mean an easy problem. It means the geometry forbids certain kinds of traps. In a convex optimisation problem, every local minimum is global under the standard assumptions. That single structural fact changes how modelling, algorithms, certificates and computation can be organised.
Series route: Mathematics Learning Hub → How Mathematics Works → Convex Analysis. Useful foundations include Linear Algebra, Real Analysis, Mathematical Optimisation and Functional Analysis.
1. Convexity begins with line segments
A set C in a vector space is convex when every line segment joining two points in C remains entirely inside C.
Formally, if x,y∈C and 0≤θ≤1, then θx+(1−θ)y∈C.
This simple geometric rule is the foundation of the subject because it says that interpolation between feasible points stays feasible.
2. Convex combinations generalise two-point averaging
A convex combination of points x₁,…,x_n has the form Σ_i θ_i x_i where θ_i≥0 and Σ_i θ_i=1.
A set is convex exactly when it contains every finite convex combination of its points.
3. Affine sets are flatter than convex sets
An affine set contains the whole line through any two of its points, not merely the segment between them.
Linear equality constraints Ax=b therefore define affine sets and are automatically convex.
4. Halfspaces are basic convex constraints
A halfspace has the form {x:aᵀx≤b}. Intersections of halfspaces create polyhedra.
This makes linear programming a special case of convex optimisation.
5. Norm balls are convex
For any norm ||·||, the ball {x:||x−c||≤r} is convex.
The triangle inequality guarantees that points interpolated between two feasible points remain inside the ball.
6. Positive semidefinite matrices form a convex cone
Symmetric matrices X satisfying vᵀXv≥0 for every v form the positive-semidefinite cone.
This cone is central to semidefinite programming, covariance modelling and control.
7. Convex functions lie below their chords
A function f is convex on a convex domain when f(θx+(1−θ)y)≤θf(x)+(1−θ)f(y).
The graph never rises above the straight chord connecting two graph points.
8. Jensen’s inequality turns convexity into expectation control
For a convex function f and suitable random variable X, f(E[X])≤E[f(X)].
The inequality says that applying a convex penalty after averaging is cheaper than averaging the penalty after variability has acted.
This is why convexity appears naturally in risk, estimation and information theory.
9. Strict convexity strengthens uniqueness
If the convexity inequality is strict for distinct points and 0<θ<1, the function is strictly convex.
A strictly convex objective has at most one global minimiser over a convex feasible set when a minimiser exists.
10. Strong convexity quantifies curvature
Strong convexity requires the function to lie above its tangent approximation by at least a quadratic margin.
This quantitative curvature produces stronger stability and convergence guarantees for optimisation algorithms.
11. Differentiable convex functions have global tangent lower bounds
For differentiable convex f, f(y)≥f(x)+∇f(x)ᵀ(y−x).
The tangent plane is a global under-estimator, not merely a local approximation.
12. Zero gradient certifies global optimality
If f is differentiable and convex and ∇f(x*)=0, then x* is a global minimiser.
This is stronger than ordinary calculus, where zero derivative can indicate minima, maxima or saddle points.
13. Hessians certify twice-differentiable convexity
If f is twice differentiable on a convex domain, a positive-semidefinite Hessian everywhere implies convexity.
Positive definiteness produces strict local curvature and, under suitable global conditions, stronger uniqueness.
14. Subgradients extend tangents to nonsmooth functions
A vector g is a subgradient of convex f at x when f(y)≥f(x)+gᵀ(y−x) for every y.
A nondifferentiable point can have many subgradients. The collection is the subdifferential ∂f(x).
15. The absolute-value function makes subgradients visible
For f(x)=|x|, the derivative is −1 for x<0 and 1 for x>0.
At x=0, every g∈[−1,1] is a valid subgradient. The kink becomes a whole interval of supporting slopes.
16. Subgradient zero is the nonsmooth optimality test
For a proper convex function, x* minimises f exactly when 0∈∂f(x*) under the standard finite-dimensional setting.
Gradient-zero generalises to subgradient containment.
17. Epigraphs convert functions into sets
The epigraph epi f={(x,t):t≥f(x)} is convex exactly when f is convex.
This lets function convexity be studied using the geometry of convex sets.
18. Sublevel sets reveal quasiconvex structure
Every sublevel set {x:f(x)≤α} of a convex function is convex.
A function can have convex sublevel sets without itself being convex; such functions are called quasiconvex.
19. Separation theorems make geometry algebraic
Under suitable closedness and disjointness conditions, two convex sets can be separated by a hyperplane.
A geometric fact about non-overlap becomes a linear inequality certificate.
20. Supporting hyperplanes touch without cutting through
At a boundary point of a suitable convex set, a supporting hyperplane contains the point while the entire set lies on one side.
Subgradients are the function-level analogue of these supporting hyperplanes.
21. Convex cones organise homogeneous constraints
A convex cone is closed under nonnegative linear combinations.
Nonnegative vectors, positive-semidefinite matrices and second-order cones are central examples.
22. Dual cones encode nonnegative certificates
The dual cone C* contains vectors y satisfying yᵀx≥0 for every x∈C.
Dual cones translate primal geometry into families of valid linear inequalities.
23. Convex conjugates exchange primal and dual descriptions
The Fenchel conjugate is f*(y)=sup_x(yᵀx−f(x)).
It records the tightest affine lower-support information of f in dual coordinates.
24. Fenchel–Young inequality is automatic from the definition
For all x,y, f(x)+f*(y)≥xᵀy.
Equality occurs precisely when y is a subgradient of f at x under standard proper closed convex assumptions.
25. Biconjugation reconstructs closed convex functions
The Fenchel–Moreau theorem says a proper lower-semicontinuous convex function equals its biconjugate.
A function can therefore be rebuilt from all of its affine supports.
26. Lagrange duality creates lower bounds
For a constrained minimisation problem, the Lagrangian combines the objective with weighted constraints.
Minimising the Lagrangian over primal variables gives a dual function that always provides a lower bound on the primal optimum under the standard minimisation convention.
27. Strong duality requires conditions
Convexity alone does not guarantee zero duality gap in every formulation. Constraint qualifications such as Slater’s condition provide strong duality for important classes.
The hypothesis is part of the certificate.
28. KKT conditions unify stationarity, feasibility and complementary slackness
Karush–Kuhn–Tucker conditions combine primal feasibility, dual feasibility, stationarity and complementary slackness.
For convex problems with suitable constraint qualifications, KKT conditions are sufficient for global optimality and often necessary.
29. Proximal operators solve regularised local subproblems
The proximal operator of f maps v to the minimiser of f(x)+(1/2λ)||x−v||² for λ>0 under standard convex assumptions.
It balances reducing f against staying near the current point.
30. Proximal methods handle nonsmooth structure cleanly
When an objective splits into a smooth term plus a nonsmooth convex regulariser, proximal-gradient methods take a gradient step on the smooth term and a proximal step on the regulariser.
L1 regularisation and constrained projections fit naturally into this framework.
31. Projection is itself a proximal operator
The indicator function of a closed convex set C is zero on C and +∞ outside.
Its proximal operator is Euclidean projection onto C.
32. Convex analysis supports modern applications
Signal processing uses convex penalties to recover sparse or structured signals. Machine learning uses convex loss functions and regularisers. Control uses linear matrix inequalities and semidefinite constraints. Finance uses convex risk and portfolio formulations. Network optimisation uses convex flow and allocation models.
Stanford’s long-running convex-optimisation materials emphasise this same bridge from geometric theory to computation and applications.
33. A worked mechanism: projection onto an interval
Minimise (1/2)(x−5)² subject to 0≤x≤3.
- The unconstrained minimiser is x=5.
- It violates x≤3.
- The nearest feasible point is x=3.
- Because the objective is strongly convex, this feasible minimiser is unique.
The solution is the projection of 5 onto the convex set [0,3].
34. Common convex-analysis failure modes
- Convex-set/function confusion: applying a property of convex sets directly to arbitrary functions.
- Local=global overreach: using the local-global guarantee without checking convexity of both objective and feasible set.
- Differentiability assumption: using gradients where only subgradients exist.
- Strong-duality assumption: ignoring constraint qualifications.
- Strict/strong convexity collapse: treating uniqueness and quantitative curvature as the same property.
- Coordinate-only reasoning: overlooking geometric separation and cone structure.
- Solver certainty: treating a numerical status message as a proof that the original modelling assumptions were correct.
35. Convex analysis as a mathematical machine
Convex Feasible Geometry → Convex Objective → Supporting/Subgradient Structure → Separation and Conjugacy → Duality/KKT → Algorithm → Global Certificate → Application.
36. What mastery looks like
- recognise convex sets and convex functions from definitions and composition rules;
- use Jensen’s inequality structurally rather than mechanically;
- replace gradients with subgradients when nonsmoothness appears;
- interpret support hyperplanes and separation geometrically;
- use conjugates and Fenchel–Young duality;
- distinguish weak and strong duality;
- apply KKT conditions with their hypotheses;
- recognise proximal and projection operators;
- connect convex structure to optimisation, control, finance and computation.
37. Further reading
- Stanford: Convex Optimization — Boyd & Vandenberghe
- Stanford EE364A: Convex Optimization I
- AMS Mathematics Subject Classification
38. Conclusion
Convex analysis works because geometry, optimisation and proof reinforce one another. Convex sets preserve interpolation. Convex functions admit global supporting structures. Subgradients extend calculus through kinks. Separation produces linear certificates. Conjugates and duality expose hidden alternative formulations.
Convexity is powerful not because every convex problem is simple, but because it removes entire categories of ambiguity between local calculation and global truth.
How Mathematics Works | Batch 13
- Convex Analysis — this article
- How Mathematics Works | Nonlinear Analysis
- How Mathematics Works | Distribution Theory
- How Mathematics Works | Sobolev Spaces
Return to the Mathematics Learning Hub or the How Mathematics Works root.