Epigraph¶
The upward-closed set of real-height pairs above an extended-real-valued function, preserving each function value as a fiber infimum.
Core Idea¶
For a function \(f:X\to\mathbb{R}\cup\{+\infty\}\), the epigraph is the set \(\operatorname{epi}f=\{(x,t)\in X\times\mathbb{R}: f(x)\leq t\}\). It replaces a value at each input with the whole real vertical ray on or above that value. If \(f(x)=+\infty\), the fiber at \(x\) is empty. Conversely, \(f(x)=\inf\{t:(x,t)\in\operatorname{epi}f\}\) when \(\inf\varnothing=+\infty\). The set therefore preserves the function, not just a rough picture of its graph.[1][2]
The construction does not require \(f\) to be convex or continuous, nor does it by itself optimize anything. Under the cited Euclidean conventions, convexity of \(f\) corresponds to convexity of its epigraph; lower semicontinuity corresponds to closedness. Optimization can use an auxiliary height \(t\) so that minimizing \(f(x)\) becomes minimizing \(t\) over epigraph membership, but linearity or tractability of the resulting constraints depends on the particular \(f\).[1][2]
Structural Signature¶
Sig role-phrases: function value → real comparison height → upward inequality → recoverable vertical fiber.
- Function and domain. An input \(x\in X\) has an extended-real value \(f(x)\). Without that assignment there is no lower boundary for a vertical fiber.[1][2]
- Ordered real-height extension. The pair \((x,t)\) adds a real comparison coordinate. A subset of inputs alone cannot encode the height threshold.[1]
- Above-value inequality. Membership holds exactly when \(t\geq f(x)\). Equality alone gives the graph; reversing the inequality gives a hypograph.[1]
- Upward fiber and recovery. Each nonempty section at fixed \(x\) is \([f(x),+\infty)\), and its infimum recovers the value. For \(f(x)=+\infty\) there is no real section and the empty-set infimum convention recovers \(+\infty\).[2]
Set convexity, closedness, finite linear constraint representation and an optimizer are useful conditional properties, not constitutive roles of every epigraph.[1][2]
What It Is Not¶
- Not the graph of a function. The live Graph of a Function contains equality pairs \((x,f(x))\) only; an epigraph includes all real heights above the value.[1]
- Not a hypograph. The live Hypograph (mathematics) points downward under \(t\leq f(x)\), reversing the defining inequality.
- Not automatically convex or closed. Those set properties correspond to additional function properties under stated domain and topology assumptions.[1][2]
- Not a guarantee of a linear program. Lifting an objective to \(t\) makes that objective linear; arbitrary epigraph-membership constraints may remain nonlinear or hard to describe.[1]
- Not every upper-looking region. Its fixed-\(x\) sections must be complete upper rays with lower endpoints determined by one function. A section with gaps above an included height fails the test.[2]
Scope of Application¶
The baseline definition works with a declared domain \(X\) and extended-real values that may be \(+\infty\) but not \(-\infty\). The empty fiber at an infeasible \(x\) lets \(+\infty\) encode exclusion from an effective domain without putting infinite heights into \(X\times\mathbb R\). An epigraph still exists for a nonconvex or discontinuous function; no theorem about convexity or closedness should be inferred without its hypotheses.[2]
In Euclidean convex analysis, Boyd and Vandenberghe use the epigraph to characterize a convex function by a convex set. Nemirovski's Proposition 2.6.1 states that for \(f:\mathbb R^n\to\mathbb R\cup\{+\infty\}\), lower semicontinuity is equivalent to closed epigraph. Both are geometric tests of extra properties, not definitions of “epigraph.”[1][2]
In optimization, Boyd and Vandenberghe present \(f(x)=\max_i(a_i^Tx+b_i)\). The lift \(\min t\) subject to \(a_i^Tx+b_i\leq t\) for all \(i\) describes its epigraph and is a linear program because these particular branch constraints are affine. Replacing this \(f\) with an arbitrary nonlinear function would not retain that LP conclusion.[1]
Clarity¶
An epigraph is not just “points above a curve” when \(f\) may take \(+\infty\). State the domain and codomain, use a real height variable, and say what an empty vertical fiber means. With \(f(x)=+\infty\), there is no real \(t\geq f(x)\); taking the infimum of that empty fiber to be \(+\infty\) makes recovery consistent.[2]
Distinguish the exact representation from its applications. Convexity asks whether line segments between epigraph points stay in the set; lower semicontinuity asks whether the set is closed; optimization asks how to express membership as constraints. They are different questions asked after constructing the same set.[1][2]
Manages Complexity¶
The epigraph transforms a function-value question into a set-membership question without losing the original values. Geometric properties and set operations can then organize functional reasoning. For a nonempty family on a common domain, \(\operatorname{epi}(\sup_i f_i)=\bigcap_i\operatorname{epi}(f_i)\) because \(t\) dominates every \(f_i(x)\) exactly when it dominates their supremum. Nemirovski uses this to reason about lower semicontinuity of a supremum via intersections of closed sets.[2]
The translation helps only if one keeps the hypotheses visible. An epigraph can be easy to define but hard to describe computationally. The compact lift of a max-of-affine objective works because each branch inequality is explicit and affine; the set representation alone does not make every objective solver-ready.[1]
Abstract Reasoning¶
For a proposed set \(E\subseteq X\times\mathbb R\), inspect each fixed-\(x\) section. If it is empty, assign \(f(x)=+\infty\); if it is a closed upper ray \([a,+\infty)\), assign \(f(x)=a\). If any section contains a height but omits a larger one, or is an open ray when only finite-valued \(f\) with inclusive inequality is allowed, it is not the epigraph of that claimed function. This is the set-side recognition test behind the recovery formula.[2]
Then assess optional geometry. For \(f(x)=x^2\), \(\operatorname{epi}f=\{(x,t):t\geq x^2\}\) is convex and closed; those conclusions use facts about \(x^2\) and the Euclidean topology. For \(f(x)=\max_i(a_i^Tx+b_i)\), membership becomes simultaneous affine inequalities, licensing the specific LP lift.[1][2]
Knowledge Transfer¶
The same upward-fiber representation serves both convex analysis and optimization modeling. In the first setting, set convexity and closedness report properties of \(f\); in the second, membership inequalities become constraints on a lifted height variable. Those uses share the exact set construction, not an unqualified claim that every epigraph is convex or every lift is linear.[1][2]
Live Function (Mapping) supplies the necessary input-to-value skeleton. Live Graph of a Function and Hypograph (mathematics) are comparison neighbors: equality pairs and downward rays respectively. Neither supplies the upward-region identity as a strict genus.[1]
Examples¶
Quadratic function in convex analysis. Let \(f(x)=x^2\) for \(x\in\mathbb R\). Mapped back: function/domain = \(x\mapsto x^2\) on \(\mathbb R\); real-height extension = \((x,t)\in\mathbb R^2\); above-value inequality = \(t\geq x^2\); upward fiber/recovery = \([x^2,+\infty)\) whose infimum is \(x^2\). Its convex closed geometry illustrates optional property correspondences, not extra defining roles.[1][2]
Maximum-of-affine objective. For \(f(x)=\max_{1\leq i\leq m}(a_i^Tx+b_i)\), Boyd and Vandenberghe minimize \(t\) subject to \(a_i^Tx+b_i\leq t\) for each \(i\). Mapped back: function/domain = the pointwise maximum on \(\mathbb R^n\); real-height extension = decision pair \((x,t)\); above-value inequality = every branch lies below \(t\); upward fiber/recovery = the smallest admissible \(t\) at fixed \(x\) is \(f(x)\). The resulting LP is due to the affine branch form.[1]
Boundary: equality graph. Keeping only \((x,f(x))\) omits every higher height at that input. It is the function graph, not its epigraph.[1]
Structural Tensions¶
Exact representation versus tractable description. Any allowed \(f\) has an epigraph, preserving its values, but an arbitrary membership test may be nonlinear or impractical. Demanding explicit linear constraints eases optimization while narrowing to special functions such as max-of-affine. Diagnostic: can the actual epigraph be expressed in the solver's permitted constraint class?[1]
Geometric compression versus assumption leakage. Convexity and closedness can be checked on one set, simplifying function proofs. But applying a Euclidean or lower-semicontinuity theorem without its domain/topology assumptions gives false confidence. Diagnostic: which exact domain, topology and \(+\infty\) convention underwrite the claimed correspondence?[1][2]
Upward-set fidelity versus graph economy. The graph uses one pair per input and is compact, but discards the immediate set geometry of all dominating heights. The epigraph includes an entire ray at each finite input, enabling convex-set reasoning at the cost of a larger representation. Diagnostic: does the proof need equality values only, or the order region above them?[1]
Structural–Framed Character¶
Evaluative weight: Membership is a formal inequality relative to \(f(x)\), not a human assessment of whether a solution is desirable. Optimization preferences enter only when a separate objective is imposed.[1]
Human-practice dependence: A mathematician chooses \(X\), \(f\) and a height convention, yet the resulting set and recovery rule follow from those choices. No observer, institution or negotiation is needed for a pair to meet \(t\geq f(x)\).[2]
Institutional origin: Convex-analysis textbooks made the representation familiar, but neither an institutional label nor a solver convention creates its defining upward fibers. Different disciplines may use the same formal set once the function and order are specified.[1][2]
Vocabulary travel: “Epigraph” travels literally between function geometry and optimization modeling because both use the same set of above-value pairs. It should not be stretched to any “upper bound” in a nonfunctional or unordered domain without reconstructing those roles.[1]
Import versus recognition: To recognize an instance, verify a function boundary and complete upper vertical sections; a shaded picture alone is insufficient. A hypograph or equality graph remains a different typed object even if visually adjacent.[1]
Its character: strongly structural but domain-specific to ordered function geometry. The live Function Mapping prime is portable; the epigraph adds a real ordered height, Cartesian-product set and upward inequality not present in every mapping.[1][2]
Structural Core vs. Domain Accent¶
Skeletal relation: Live Function (Mapping) gives the necessary input-to-output assignment. The staged composition/presupposes edge is warranted because removing \(f(x)\) removes the lower boundary of each ray; it is not subsumption, since an epigraph is a set rather than a function.[1]
Domain-bound residual: \(X\times\mathbb R\), the inclusive above-value inequality, upward sections and infimum recovery distinguish this object. Convexity, closedness and solver-friendly constraints are optional accents tied to further hypotheses and applications.[1][2]
Why not prime: The exact identity is stated in function/set/order mathematics. Its two mapped settings reuse that same mathematical representation, not independently grounded nonmathematical substrates. Removing the real-height inequality to claim broad portability would collapse epigraph into a generic “things above other things” metaphor and erase its recognition test.[1]
Instantiates / Related Primes¶
This entry presupposes Function (Mapping).
The proposed typed relation is composition/presupposes Function (Mapping). A function is required to determine every fiber, but the epigraph itself is not a function subtype. Convexity is not a parent: nonconvex functions still have epigraphs, and convexity is a conditional property of some of those sets.[1]
Live Graph of a Function represents equality pairs and live Hypograph (mathematics) represents below-value pairs. These are neighboring constructions with opposite or stricter membership conditions, not exact aliases or necessary parents.[1]
Relationships to Other Abstractions¶
Current abstraction Epigraph Domain-specific
Parents (1) — more general patterns this builds on
-
Epigraph presupposes Function (Mapping) Prime
Each upward fiber is bounded by a value f(x) assigned by a function.The live Function Mapping prime supplies the input-to-output assignment necessary to define f(x). Removing it leaves no function boundary for the epigraph's vertical rays. The epigraph is not a subtype of a function or its equality graph, and the relation does not assume convexity.
Hierarchy path (1) — routes to 1 parentless root
- Epigraph → Function (Mapping)
Neighborhood in Abstraction Space¶
Epigraph sits in a sparse region of the domain-specific corpus (72nd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Graph Structures & Combinatorial Objects (44 abstractions)
Nearest neighbors
- Slice Sampling — 0.85
- Norm — 0.84
- Euclidean Space — 0.83
- Normal Order of an Arithmetic Function — 0.83
- Invex Function — 0.83
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
Graph of a Function: \(\{(x,f(x))\}\) has one pair at each finite input; \(\operatorname{epi}f\) has all real heights at or above the value.[1]
Hypograph: The lower region \(t\leq f(x)\) reverses the inequality and supports concavity-oriented analogues, but it is not this object.
Epigraph reformulation: The optimization technique uses an epigraph; it is not the definition of one. It linearizes the lifted objective \(t\), not every accompanying constraint.[1]
References¶
[1] Stephen Boyd and Lieven Vandenberghe, Convex Optimization lecture slides, author-hosted PDF with revised slides credited to Boyd, Vandenberghe and Parth Nobel, slide 3.12/PDF p. 56 for epigraph and convexity, and slide 4.15/PDF p. 102 for max-of-affine epigraph LP. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z ↩27 ↩28 ↩29 ↩30 ↩31 ↩32 ↩33 ↩34
[2] Aharon Ben-Tal and Arkadi Nemirovski, Optimization III: Convex Analysis, Nonlinear Programming Theory, Nonlinear Programming Algorithms, original Georgia Tech-hosted lecture notes (Fall 2020), Lecture 2 §2.6.1, PDF pp. 72–74, especially Proposition 2.6.1 and Corollary 2.6.1. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u