--- title: 'Star-Convex Functions: Theory & Applications' url: https://www.emergentmind.com/topics/star-convex-functions type: topic --- # Star-Convex Functions: Theory & Applications Star-convexity is a generalized convexity condition organized around a distinguished center. In optimization, the center is typically a global minimizer \(x^*\), and the defining requirement is that every segment from \(x^*\) to another point lies below the affine interpolation of function values. In one-dimensional geometric formulations, a point \(p\) is a star-center if the segment joining \((p,f(p))\) to \((x,f(x))\) lies entirely in the epigraph or entirely in the hypograph of \(f\). In geometric function theory, the term also appears in the closely related class of uniformly starlike functions. Across these settings, star-convexity strictly generalizes ordinary convexity while retaining strong radial or centered structure that is sufficient for geometric representation theorems, majorization inequalities, and nontrivial optimization guarantees [1511.04466] [1906.11985] [2308.00704] [1106.4377]. ## 1. Definitions and core equivalences For a function \(f:\mathbb R^n\to\mathbb R\) and a fixed point \(x_0\in\mathbb R^n\), star-convexity about \(x_0\) means that for all \(x\in\mathbb R^n\) and all \(\alpha\in[0,1]\), \[ f\!\bigl(\alpha x_0+(1-\alpha)x\bigr)\le \alpha f(x_0)+(1-\alpha)f(x). \] Equivalently, each ray from \(x_0\) is a convex function of radius. In constrained form, if \(\mathcal C\subset\mathbb R^n\) is convex and \(X^*=\arg\min_{x\in\mathcal C} f(x)\neq\emptyset\), then \(f\) is star-convex on \(\mathcal C\) if for every \(x^*\in X^*\), every \(x\in\mathcal C\), and every \(\lambda\in[0,1]\), \[ f\bigl(\lambda x^*+(1-\lambda)x\bigr)\le \lambda f(x^*)+(1-\lambda)f(x). \] For differentiable \(f\), this is equivalent to the first-order inequality \[ f(x^*)-f(x)\ge \nabla f(x)^\top(x^*-x), \] which is the form used repeatedly in optimization analyses [2507.17272] [1511.04466]. A broader differentiable variant is \(\tau\)-star-convexity: \(f\) is \(\tau\)-star-convex about \(x^\star\) if \[ \tau \langle \nabla f(x),x-x^\star\rangle \ge f(x)-f(x^\star), \qquad \forall x. \] The case \(\tau=1\) recovers convexity, whereas \(\tau>1\) weakens it. In the same source, the condition \[ f(x^\star+t(x-x^\star))\le (1-t)f(x^\star)+t f(x), \qquad t\in[0,1], \] is identified with unimodality along every ray through the minimizer [2405.18976]. In the one-dimensional formulation on an interval \(I\subseteq \mathbb R\), the geometric objects attached to \(f:I\to\mathbb R\) are \[ \operatorname{graph} f=\{(x,t)\in I\times \mathbb R:t=f(x)\}, \] \[ \operatorname{epi} f=\{(x,t)\in I\times \mathbb R:t\ge f(x)\}, \] \[ \operatorname{hypo} f=\{(x,t)\in I\times \mathbb R:t\le f(x)\}. \] The function is star-convex if there exists \(p\in I\) such that for every \(x\in I\) and every \(t\in[0,1]\), the point \[ P(t)=t(x,f(x))+(1-t)(p,f(p)) \] lies entirely either in \(\operatorname{epi} f\) or in \(\operatorname{hypo} f\). Equivalently, for all \(x\in I\) and \(t\in[0,1]\), one of the inequalities \[ f(tx+(1-t)p)\le t f(x)+(1-t)f(p) \] or \[ f(tx+(1-t)p)\ge t f(x)+(1-t)f(p) \] holds. The set of all such centers is the central set \(C_f(I)\) [2308.00704]. These definitions immediately separate star-convexity from ordinary convexity. Every convex function is star-convex, but the converse need not hold. A second common misconception is that centered convexity along rays should imply regularity comparable to convex analysis; the optimization literature explicitly shows that this is false in the measurable setting [1511.04466]. ## 2. Algebraic structure and one-dimensional geometry In the interval setting, several structural properties are immediate. If \(f\) is convex, then every line segment joining two points of \(\operatorname{graph} f\) lies in \(\operatorname{epi} f\), so \(C_f(I)=I\). If \(f\) is concave, then \(C_f(I)=I\) as well, with the segment lying in \(\operatorname{hypo} f\). More generally, the class of star-convex functions on \(I\) is a convex cone, meaning it is closed under non-negative linear combinations, and it is symmetric in the sense that \(f\) is star-convex about \(p\) if and only if \(-f\) is star-convex about \(p\) [2308.00704]. A sharp sufficient condition is piecewise convexity or concavity around a selected point \(p\). If \(f:I\to\mathbb R\) is continuous and one of the following four configurations holds, then \(f\) is star-convex with center \(p\): 1. \(f|_{(-\infty,p]}\) is convex and \(f|_{[p,\infty)}\) is convex. 2. \(f|_{(-\infty,p]}\) is concave and \(f|_{[p,\infty)}\) is concave. 3. \(f|_{(-\infty,p]}\) is convex and \(f|_{[p,\infty)}\) is concave. 4. \(f|_{(-\infty,p]}\) is concave and \(f|_{[p,\infty)}\) is convex. In the first two cases, every segment from \((p,f(p))\) to \((x,f(x))\) lies wholly in \(\operatorname{epi} f\) or wholly in \(\operatorname{hypo} f\). In the mixed cases, one half-segment lies in \(\operatorname{epi} f\) and the other in \(\operatorname{hypo} f\), but \(p\) still remains a valid star-center [2308.00704]. This formulation makes the centered nature of the concept explicit. Convexity imposes inequalities between all pairs of points, whereas star-convexity only constrains segments attached to a distinguished center. A plausible implication is that star-convexity is best understood ոչ as a small perturbation of convexity, but as a radial geometry with a privileged anchor point. ## 3. Star-convex bodies and generalized variants The one-dimensional theory in turn induces planar star-convex sets. A subset \(X\subset\mathbb R^2\) is star-convex, or star-shaped, if there exists \(q\in X\) such that for every \(x\in X\), the segment from \(q\) to \(x\) lies entirely in \(X\). If \(f:I\to\mathbb R\) is continuous and star-convex about \(p\), then in the four piecewise convexity/concavity configurations above one obtains corresponding star-convex subsets of \(\mathbb R^2\) centered at \((p,f(p))\): unions of two half-epigraphs, two half-hypographs, or a mixed epi/hypo union, depending on the left and right behavior of \(f\) [2308.00704]. This realizes star-convexity of a scalar function as a star-shaped planar body whose boundary is the graph \(y=f(x)\). A distinct line of generalization is \(w\)-\(m\)-star-convexity on convex subsets \(C\) of Banach spaces. For \(m\in(0,1]\) and a modulus \(w:[0,\infty)\to\mathbb R\), a function \(\Phi:C\to\mathbb R\) is \(w\)-\(m\)-star-convex if for every \(x,y\in C\) and every \(\lambda\in(0,1)\), \[ \Phi\bigl((1-\lambda)x+m\lambda y\bigr) \le (1-\lambda)\Phi(x)+m\lambda \Phi(y)-m\lambda(1-\lambda)w(\|x-y\|). \] If \(w\equiv 0\), this reduces to \(m\)-star-convexity; if in addition \(m=1\), one recovers the usual convexity condition. The parameter \(m\) dilutes the weight placed on \(y\), and \(w(\|x-y\|)\) contributes a perturbation term depending on distance [2207.08058]. Within ordered Banach spaces, this generalized notion supports an extension of the Hardy–Littlewood–Pólya inequality of majorization. If \(\mu=\sum_{k=1}^N A_k\delta_{x_k}\) and \(\nu=\sum_{k=1}^N A_k\delta_{y_k}\) satisfy the preorder \(\mu\prec_{mL^+}\nu\), and if \(\Phi:C\to F\) is Gâteaux-differentiable, \(w\)-\(m\)-star-convex, and has isotone differential \(d\Phi(x)\), then \[ \sum_{k=1}^N A_k\Phi(y_k) \le \sum_{k=1}^N A_k\Phi(x_k) + m\sum_{k=1}^N A_k w(\|x_k-y_k\|). \] The same conclusion also holds under the weaker preorder \(\mu\prec_{wmL^+}\nu\) provided \(\Phi\) itself is isotone [2207.08058]. The same paper records a perspective construction: if \(f\) is \(w\)-\(m\)-star-convex on a cone \(C\subseteq E\), then \[ \tilde f(x,t)=t\,f(x/t), \qquad (x,t)\in C\times(0,\infty), \] is again \(w\)-\(m\)-star-convex. Concrete examples include \(f(x)=x^4-5x^3+9x^2-5x\) on \(x\ge 0\), which is \((m=16/17)\)-star-convex but not convex, and \(y(x)=-2x^3+5x^2+6x\) on \((-\infty,1]\), which is \(27/28\)-star-convex; its perspective \[ Y(x,t)=x^4/t-5x^3+9x^2 t-5x t^2+3t^3 \] has isotone differential on \(I\times [1,\infty)\) [2207.08058]. ## 4. Examples, nonconvexity, and pathological behavior The literature supplies a large family of star-convex functions that are not convex. Examples in one and several variables include \(f(t)=|t|(1-e^{-|t|})\) on \(\mathbb R\), \(f(s,t)=s^2t^2+s^2+t^2\) on \(\mathbb R^2\), and \(f(x,y)=x^2y^2\) on \(\mathbb R^2\), each star-convex about the origin or \((0,0)\) but nonconvex [2507.17272] [1906.11985]. Any continuous, nonnegative, positively homogeneous function of degree \(r\ge 1\) is star-convex about the origin, since \(f(\lambda x)=\lambda^r f(x)\le \lambda f(x)\) for \(\lambda\in[0,1]\) [2507.17272]. The class also includes generalized \(\ell_p\) distances with subunit exponents. For any real \(p\), \[ f(x,y)=\bigl(|x|^p+|y|^p\bigr)^{1/p} \] is star-convex with star center \((0,0)\), while it is nonconvex when \(p<1\). More generally, if \(f\) and \(g\) are star-convex with a common star center \(x_0\) and \(f(x_0)=g(x_0)=0\), then for any real \(p\), \[ h(x)=\Bigl(\frac{f(x)^p+g(x)^p}{2}\Bigr)^{1/p} \] is star-convex about \(x_0\); for \(p<1\) convexity fails but star-convexity is preserved. Another flexible construction is the arbitrary radial extension \[ f(x)=\|x\|\cdot g\!\Bigl(\frac{x}{\|x\|}\Bigr), \] with \(g\) any bounded measurable function on the unit sphere and \(f(0)=0\), yielding star-convexity about the origin even when transverse behavior is discontinuous or highly oscillatory [1511.04466]. These examples are matched by strong negative results. A measurable star-convex function can be discontinuous almost everywhere, can lack subgradients, and can oscillate arbitrarily fast transverse to the optimal ray. One can encode an exponentially large unknown parameter into behavior on an exponentially small set of rational directions, so deterministic gradient or cutting-plane methods fail. The class of Lebesgue-measurable star-convex functions on \(\mathbb R^2\) has cardinality \(2^{|\mathbb R|}\), whereas the class of continuous functions has cardinality \(|\mathbb R|\). There also exist smooth star-convex functions for which any deterministic gradient oracle returns information independent of the true optimum in one coordinate [1511.04466]. At the same time, several sources describe star-convexity as a model for global radial unimodality. One paper states that star-convex functions are strictly unimodal on all lines through a minimizer, and another notes that certain neural-network empirical losses appear star-convex in neighborhoods of their global minima, while many machine-learning losses appear empirically to be star-convex in large basins around global minimizers [2405.18976] [1906.11985]. This suggests that the concept occupies an intermediate position: it is substantially weaker than convexity, yet still strong enough to exclude spurious local minima along rays emanating from a minimizer. ## 5. Optimization algorithms and complexity theory The optimization literature treats star-convexity as a structure that permits global or accelerated methods beyond the convex setting. | Method | Setting | Guarantee | |---|---|---| | Randomized cutting-plane / ellipsoid method | Lebesgue-measurable star-convex \(f:\mathbb R^n\to\mathbb R\) with weak sampling-oracle access | Time \(\mathrm{poly}(n,\log(1/\epsilon),\log R,\log B,\log(1/F))\); returns either a Gaussian with \(\Pr[f(x)\le f^*+\epsilon]\ge 1-\epsilon\) or an ellipsoid of radius \(\tau\ll 1\) around \(x^*\) [1511.04466] | | Accelerated first-order method with binary line search | \(L\)-smooth star-convex \(f\) | \(O(\epsilon^{-1/2}\log(\epsilon^{-1}))\) total function and gradient evaluations; lower bound \(\Omega(\epsilon^{-1/2})\) for deterministic first-order methods [1906.11985] | | Accelerated mirror descent + binary search | \(\tau\)-star-convex, \((L,\kappa)\)-weakly smooth \(f\) in an arbitrary norm, with \((\mu,q)\)-uniformly convex mirror map \(\psi\) | \(f(x_{T+1}^{ag})-f^\star = O_{q,\kappa}\!\left[L\tau^q B^{\kappa/q}(\log T)/T^{(\kappa q+\kappa-q)/q}\right]\); total oracle complexity \(O(T\log T)\) [2405.18976] | | Frank–Wolfe | Differentiable star-convex \(f\) with \(L\)-Lipschitz gradient over compact convex \(\mathcal C\) | \(\mathcal O(1/k)\) for objective values and Frank–Wolfe duality gap under diminishing, Armijo-type, and Lipschitz-based stepsizes [2507.17272] | The 2015 global algorithm addresses the maximally unrestricted setting of measurable star-convexity. Its core devices are a blurred logarithm \[ L_z(x)= \begin{cases} \log \epsilon' & f(x)-z\le \epsilon',\\ \log(f(x)-z) & \epsilon'1\) [2405.18976]. The 2025 Frank–Wolfe analysis shows that classical \(\mathcal O(1/k)\) complexity extends from convex to star-convex objectives over compact convex sets. Both the objective residual and the duality gap decay at this rate under diminishing, Armijo-type, and Lipschitz-based stepsize rules, and the diminishing and Armijo strategies do not require prior knowledge of Lipschitz or curvature constants. The same source contrasts this with general nonconvex Frank–Wolfe, where only \(\mathcal O(1/\sqrt{k})\) is generally available for \(\min_{i\le k} |\omega_i|\) in the absence of star-convex structure [2507.17272]. ## 6. Uniformly starlike functions and open directions In geometric function theory, uniformly starlike functions provide a distinct but related use of the term. Let \(A\) be the class of analytic functions in the unit disk \(D=\{z\in\mathbb C:|z|<1\}\) normalized by \(f(0)=0\) and \(f'(0)=1\), and let \(S\subset A\) be the univalent subclass. A function \(f\in S\) is uniformly starlike, abbreviated UST, if it maps every circular arc \(\gamma\subset D\) whose center \(\zeta\) also lies in \(D\) onto a curve which is starlike with respect to \(f(\zeta)\). Goodman’s two-point criterion gives \[ f\in UST \iff \Re\Bigl\{\frac{(z-\zeta)f'(z)}{f(z)-f(\zeta)}\Bigr\}\ge 0 \qquad \forall z,\zeta\in D. \] Taking \(\zeta=-z\) yields \(\Re\{z f'(z)/f(z)\}\ge 0\), so \(UST\subset S^*\), but no single-variable equivalent of the full two-point condition is known [1106.4377]. The survey records several quantitative results. If \(f(z)=z+\sum_{n\ge 2} a_n z^n\in UST\), then \(|a_n|<2/n\). For \(|z|=r\), if \[ m(r)=\min_{|z|=r}|f(z)|,\qquad M(r)=\max_{|z|=r}|f(z)|, \] then \[ \frac{r}{1+2r}\le m(r)