Invex Function¶

A differentiable function admitting a comparison map that bounds every value gap below by a gradient pairing at the base point.

Version: v2 · 2026-10-03 · History
Domain-specific #: 13348
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomains: Generalized Convexity, Mathematical Optimization → Mathematics
Aliases: Invex function, Eta-invex function

Core Idea¶

A differentiable real-valued function $f$ on a Euclidean domain is invex when there exists a vector-valued map $\eta(y,x)$ such that, for every ordered pair of domain points $x,y$,

\[ f(y)-f(x)\;\geq\;\eta(y,x)\cdot\nabla f(x). \]

The gradient is read at the base point $x$; the actual value gap is measured from $x$ to $y$. The map $\eta$ supplies the comparison direction and must work for all pairs, not just one local calculation. In ordinary differentiable convexity, $\eta(y,x)=y-x$. Invexity permits another map, so a function can fail the straight-line convexity test while retaining a first-order global bound.^[1]

At a stationary point $x^*$, the gradient vanishes and the inequality becomes $f(y)\geq f(x^*)$ for every $y$. Thus every stationary point of an invex function is a global minimum. In the Euclidean differentiable setting the converse also holds: if every stationary point is globally minimizing, an $\eta$ can be constructed, although that existence result says little about whether the map is convenient for computation.^[1] This property is about the function's global first-order geometry. It is not a promise that a search algorithm will find a stationary point.

Structural Signature¶

Sig role-phrases: differentiable scalar function → ordered comparison pair → direction map $\eta$ → universal value-gap bound.

Differentiable scalar function: $f$ supplies a real value and a gradient at each base point. Without that gradient the defining pairing is not the invex test given here.^[1]
Ordered comparison pair: $x$ is where the gradient is evaluated and $y$ is the competing point. The requirement ranges over every admissible $x,y$; checking a neighborhood or one candidate cannot establish the class.^[1]
Comparison map $\eta$: One vector map sends the pair $(y,x)$ to a direction compatible with the gradient at $x$. It need not equal the straight difference $y-x$; fixing it to that difference would recover only the convex special case.^[1]
Global first-order bound: The actual value difference must be at least $\eta(y,x)\cdot\nabla f(x)$ throughout the domain. If it fails for one pair, the proposed map is not a certificate; if no map works, the function is not invex.^[1]

The zero-gradient consequence is derived from these roles. It is not an additional free-standing role that can substitute for the all-pairs inequality.

What It Is Not¶

Invexity is not convexity under another spelling. A differentiable convex function supplies the easy witness $\eta(y,x)=y-x$, but the class also contains nonconvex members. The nonconvex example below has negative second derivative at $\pi/2$ yet satisfies invexity with a different, less useful map.^[1]

It is not quasiconvexity. The live Quasiconvex Function entry tests whether every sublevel set is convex; that condition neither states the invex gradient inequality nor names a common $\eta$ witness. Nor is an invex function the same object as a constrained invex program, a Type I invex objective/constraint pair, or a numerical method. Each adds its own domain, constraint or algorithmic assumptions.^[1]

It is not enough to observe that a solver found a good minimum, or to verify the inequality at a few sampled points. A nonglobal stationary point rules the class out: at that base point the gradient term is zero for every proposed $\eta$, while another point has a smaller value.

Scope of Application¶

The literal home is generalized convexity and differentiable optimization. Convex objectives give one family of invex functions; explicitly nonconvex objectives can also qualify. The point of the broader class is to distinguish nonconvex functions with the stationary-point/global-minimum property from nonconvex functions with genuinely bad stationary traps.^[1]

Invex functions can be studied as objectives in unconstrained minimization and within specially formulated constrained programs. Those are uses of the definition, not identical identities. Barik, Sra and Honorio define a constrained invex program separately and develop first-order methods under additional assumptions about smoothness, geometry and update feasibility. A KKT statement for a constrained problem requires the hypotheses of its particular theorem; the single-function inequality alone does not certify arbitrary constrained points or supply a convergence rate.^[1]

Clarity¶

The $\eta$ map clarifies which part of convexity is retained. Straight-line chords may misbehave, yet every comparison still has a first-order lower bound based at $x$. This separates a visual impression of nonconvex curvature from the more specific question: can any stationary point have a lower-valued competitor? An invex function answers no. A generic nonconvex function need not.^[1]

The definition also makes the quantifiers visible. A proposed witness has to work for all ordered pairs. If someone gives a different direction for each favored comparison without defining one map on the full pair domain, or proves a bound only near the candidate optimum, they have not established the property. Conversely, a witness may be mathematically valid yet inconvenient for an algorithm; certification and efficient search are different claims.

Manages Complexity¶

Invexity collapses a global optimality question at a stationary point to a local gradient check once the global inequality has been established. The many competing values $f(y)$ are controlled by one reusable comparison rule $\eta$ and the gradient at the point in question. The bound does not require computing each competitor value again at the moment of certification.^[1]

That compression has a cost: finding or verifying a suitable $\eta$ over the whole domain may be harder than locating one stationary point. The broad characterization “every stationary point is global” can establish membership in principle but may yield a direction map with discontinuities or a difficult inverse. The abstraction manages the logical relation between local and global optimality; it does not by itself make large nonconvex optimization instances easy.^[1]

Abstract Reasoning¶

To assess a proposed invex function, first specify the differentiable domain and the comparison map. Then evaluate $f(y)-f(x)-\eta(y,x)\cdot\nabla f(x)$ and prove it is nonnegative for every pair. At a stationary point, set $\nabla f(x)=0$ in the verified inequality. The conclusion $f(y)\geq f(x)$ for all $y$ is then immediate, without a Hessian positivity requirement.^[1]

A useful necessary rejection test runs in reverse: find a stationary $x$ and a competing $y$ with $f(y)<f(x)$. No choice of $\eta(y,x)$ can repair that failure because the gradient at $x$ is zero. For a possible nonconvex positive case, the stationary-point characterization supplies an existence check, but a constructive $\eta$ and its regularity should be examined separately if it will drive an update rule.^[1]

Knowledge Transfer¶

The same named structure applies literally to a convex quadratic and to a nonconvex differentiable objective: both retain the all-pairs inequality, while their comparison maps differ. It also remains meaningful in the manifold formulation used by Barik and colleagues, where the map takes values in a tangent space and the dot product becomes a tangent-space inner product.^[1]

An application to machine-learning loss minimization is literal only after that loss and its domain meet the differentiability and invexity conditions. The portable parent Gradient explains the local directional comparison, but calling a managerial or biological problem “invex” because every observed local improvement looks good would be analogy without a differentiable function and all-pairs certificate.

Examples¶

Convex quadratic. Let $f(t)=t^2$ on $\mathbb R$ and choose $\eta(y,x)=y-x$. Since $f'(x)=2x$, $f(y)-f(x)-\eta(y,x)f'(x)=(y-x)^2\geq0$ for every $x,y$. Mapped back: differentiable scalar function = the quadratic with derivative $2x$; ordered comparison pair = arbitrary real $x,y$; comparison map $\eta$ = $y-x$; global first-order bound = the nonnegative residual square. This is the convex specialization of the paper's definition.^[1]

Nonconvex invex function. Barik, Sra and Honorio give $f(t)=t^2+3\sin^2t$ as an invex but nonconvex example. Here $f'(x)=2x+3\sin(2x)$ has the sign of $x$: for $0<x\leq\pi/2$ both terms are nonnegative, and for $x>\pi/2$ the $2x$ term exceeds $3$; odd symmetry handles negative $x$. Thus $0$ is its only stationary point and $f(0)=0$ is global minimum, while $f''(\pi/2)=2+6\cos\pi=-4$ rules out convexity. One explicit witness is $\eta(y,x)=[f(y)-f(x)]/f'(x)$ when $x\ne0$, and $\eta(y,0)=0$. Mapped back: differentiable scalar function = the printed smooth nonconvex objective; ordered comparison pair = arbitrary real $x,y$, split by $x=0$; comparison map $\eta$ = the piecewise witness; global first-order bound = equality for $x\ne0$, and $f(y)\geq f(0)$ for $x=0$. This author-derived witness proves membership but does not promise a practical gradient update.^[1]

Structural Tensions¶

Broad membership versus constructive geometry. Allowing a freely chosen $\eta$ includes nonconvex functions with sound stationary-point certificates. But a witness manufactured by dividing a value gap by a derivative can be difficult to compute, invert or use in a stable step. Restricting $\eta$ to a convenient family improves algorithm design but may exclude valid invex members. Diagnostic: Is the displayed map only an existence certificate, or can it support the update a proposed method requires?^[1]

Stationary-point guarantee versus reaching stationarity. The inequality turns any stationary point reached into a global minimizer. It does not force an arbitrary algorithm to converge to such a point. Stronger smoothness and geometric assumptions can enable convergence-rate theorems; without them, claiming a fast global solver spends a guarantee the definition never supplied. Diagnostic: Which stated algorithm assumptions establish the path to a stationary point, separately from the invex certificate?^[1]

Structural–Framed Character¶

Invexity lies near the structural end of the spectrum: the inequality and its quantifiers determine membership, independently of whether the function models a loss, cost or energy. Its evaluative weight is neutral; mathematical membership does not say the modeled objective is desirable. Its human-practice dependence concerns the choice of objective, domain and comparison map, but the inequality's truth is then fixed. Its institutional origin in generalized-convexity research supplied the name, not a discretionary classification rule.^[1]

Its vocabulary travels literally between Euclidean and compatible manifold optimization when gradients, tangent directions and global pair comparisons are preserved. Applying “invex” to a merely favorable search story would import the word by analogy; a calculus certificate must be recognized for the exact name to apply. Its character: a sharply structural, mathematically framed function class whose broad use remains inside differentiable analysis and optimization, not a demonstrated substrate-independent prime.

Structural Core vs. Domain Accent¶

The skeletal relation is a global value comparison bounded by local directional information at a base point. That is the bridge to live Gradient, whose directional pairing is a prerequisite for the inequality. The domain accent is indispensable: a differentiable real objective, a vector or tangent comparison map, and universal first-order inequalities. The presence or absence of stationary traps follows from precisely that calculus structure.^[1]

Why not prime: a general pattern of turning local evidence into a global conclusion may travel widely, but the invex function name does not apply to those other cases without a gradient and an $\eta$ certificate. Any wider portable skeleton belongs to an already supported prime such as Gradient, or to a separately argued future abstraction; it is not inferred merely from multiple optimization applications.

This entry presupposes Gradient.

This is not a claim that Gradient itself is an invex function. Convexity is related but not a strict parent: choosing $\eta=y-x$ yields convex functions as a subclass of invex functions, while the nonconvex example lies outside that subclass. Optimization is the activity in which this function class can be useful, not the genus of a differentiable function property.

The live Quasiconvex Function is a separate generalized-convexity neighbor with a sublevel-set test. Biconvex optimization is a two-block program class. Semantic proximity to either entry does not establish identity or hierarchy.

Relationships to Other Abstractions¶

Current abstraction Invex Function Domain-specific

Parents (1) — more general patterns this builds on

Invex Function presupposes Gradient Prime

Invexity compares each value gap with a direction paired against the gradient at a base point.

Hierarchy path (1) — routes to 1 parentless root

Invex Function → Gradient

Neighborhood in Abstraction Space¶

Invex Function sits in a sparse region of the domain-specific corpus (82^nd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2551 abstractions)

Nearest neighbors

Jensen's Inequality — 0.84
Epigraph — 0.83
Multilinear form — 0.82
Ridders' Method — 0.82
Bundle metric — 0.81

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Differentiable convex function: it uses the fixed straight displacement $y-x$; invexity can use another map and include nonconvex functions.^[1]
Quasiconvex function: convexity of sublevel sets is a different recognition criterion from the global gradient pairing.
Constrained invex program or Type I variant: these add objective-and-constraint structure and theorem-specific conditions; the present node classifies one differentiable function.^[1]
A successful gradient method: convergence to a stationary point requires algorithm and geometry assumptions beyond invexity. Once stationarity is actually reached, invexity supplies the global-minimum conclusion.^[1]

References¶

[1] Adarsh Barik, Suvrit Sra and Jean Honorio, “Invex Programs: First Order Algorithms and Their Convergence”, 2023, Fig. 1©, §2 Definition 2.1 and equation (1), §2 Definition 2.4, and §3.1–3.2. Original research paper; the example's derivative-sign argument and displayed piecewise $\eta$ are explicit derivations here from its printed function and definition. 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