---
title: Convex Overlap-Based Fidelity
url: https://www.emergentmind.com/topics/convex-overlap-based-fidelity
type: topic
---

# Convex Overlap-Based Fidelity

Searching arXiv for the cited papers and closely related overlap/fidelity work.
arxiv_search(query="2504.18352 OR 1503.00304 OR 1902.00133 OR 2105.07336 OR 1801.00816 OR 2307.00409 OR 2301.02949", max_results=10, sort_by="relevance")
Convex Overlap-Based Fidelity denotes a class of fidelity or similarity constructions in which closeness is measured by overlap, while convexity enters through the objects being compared, the admissible mixtures or transformations, or the optimization landscape itself. In computational geometry, the canonical example is the overlap area
\[
A(t):=\textnormal{Area}((P+t)\cap Q)
\]
of two convex polygons under translation, with fidelity maximized by
\[
\max_{t\in\mathbb{R}^2}\;\textnormal{Area}((P+t)\cap Q),
\]
and with the key structural property that \(\sqrt{A(t)}\) is downward concave on the region where \(A(t)>0\) [2504.18352]. In quantum information, fidelity is an overlap quantity on density matrices,
\[
F(\rho,\sigma)=\operatorname{Tr}\bigl(\sqrt{\sqrt{\rho}\,\sigma\,\sqrt{\rho}}\bigr),
\]
and its geometry is analyzed on the closed, convex set of states and on convex sets of channels [1503.00304]. Related formulations appear in convex polygon matching [2307.00409], convex-hull overlap in causal inference [1801.00816], lattice-valued overlap aggregation [1902.00133], optimal convex approximation of quantum states [2105.07336], overlap functionals turned into convex programs [2201.05704], and smooth convex data fidelities built from linearly involved overlap operators in inverse problems [2509.03258]. Taken together, these works define a coherent research perspective rather than a single universal formula.

## 1. Core formulations and recurring mathematical structure

The cited literature uses “overlap” in several distinct senses, but the underlying pattern is stable: a fidelity is large when two objects align well, when a target lies near a convex set of feasible representations, or when a design remains in a shared support region. In convex shape matching, overlap is literal intersection area or volume under translation [2504.18352]. In quantum theory, fidelity is an overlap-like trace functional on density matrices and reduces to the modulus of an inner product for pure states [1503.00304]. In lattice theory, overlap is a bivariate aggregation operator on a bounded lattice, with symmetry, boundary conditions, and monotonicity built into the definition [1902.00133].

| Setting | Overlap object | Convex structure |
|---|---|---|
| Convex polygons | \(\textnormal{Area}((P+t)\cap Q)\) | Convex shapes; concavity of \(\sqrt{A(t)}\) [2504.18352] |
| Quantum states | \(F(\rho,\sigma)\) | Convex set of density matrices and channels [1503.00304] |
| Convex approximation | \(\max_{\{p_i\}}F(\rho,\sum_i p_i\rho_i)\) | Convex hull of available states [2105.07336] |
| Causal inference | \(\operatorname{co}(\mathbf{Z}_0)\cap \operatorname{co}(\mathbf{Z}_1)\) | Convex hull overlap of treatment groups [1801.00816] |
| Lattice aggregation | \(O:L^2\to L\) | Bounded lattices; convex sum via t-norm/t-conorm [1902.00133] |

A common structural feature is that overlap is often easier to optimize or characterize because convexity suppresses pathological combinatorics. In the geometric setting, convexity yields a well-behaved configuration-space objective. In the quantum setting, concavity of fidelity in each argument ties extremal behavior to pure states or extreme channels. In causal inference, convex hull intersection converts positivity into a geometric separability question. This suggests that convex overlap-based fidelity is less a domain-specific definition than a transferable analytic template.

## 2. Convex polygons under translation as the canonical geometric model

The most explicit formulation appears in Chan and Hair’s treatment of the maximum overlap problem for convex polygons. Given convex polygons \(P,Q\subset\mathbb{R}^2\) with \(n\) and \(m\) edges, the objective is
\[
\max_{t\in\mathbb{R}^2}\;\textnormal{Area}((P+t)\cap Q),
\]
with \(N:=n+m\) and
\[
A(t):=\textnormal{Area}((P+t)\cap Q).
\]
The central imported lemma from de Berg et al. is
\[
\boxed{\sqrt{A(t)}\text{ is downward concave over all }t\text{ for which }A(t)>0,}
\]
which places the problem in the regime of fixed-dimensional concave maximization [2504.18352].

The same work explicitly interprets overlap as a fidelity criterion. It states that the maximum overlap problem may be viewed as shape matching, and that overlap area is “another very natural measure” because “the larger it is, the closer the two shapes are” [2504.18352]. Under this interpretation,
\[
\text{fidelity}(P,Q;t):=A(t)=\textnormal{Area}((P+t)\cap Q)
\]
is a positive similarity measure: it vanishes when the translated shapes do not intersect and increases with alignment quality.

The technical structure of \(A(t)\) is unusually favorable. For each edge pair \(e\subset A\), \(e'\subset B\), the set of translations for which \(e+t\) intersects \(e'\) is a parallelogram \(\pi(e,e')\) in translation space. The union of their boundaries,
\[
\Pi(A,B):=\bigcup_{\text{edges }e\subset A,\,e'\subset B}\partial\pi(e,e'),
\]
partitions configuration space into cells. On any triangle or segment \(\mathcal{T}'\) contained in one cell of \(\Pi(A,B)\), the overlap area becomes a constant-complexity quadratic:
\[
f(t)=\textnormal{Area}((A+t)\cap B)\quad\text{for all }t\in\mathcal{T}'.
\]
This piecewise-quadratic structure is the local algebraic form of geometric overlap fidelity [2504.18352].

Chan and Hair’s principal algorithmic result is the first randomized exact algorithm with expected linear running time:
\[
t^*\in\arg\max_{t\in\mathbb{R}^2}A(t)\quad\text{in expected time }O(N).
\]
The method combines configuration-space geometry, local quadratic formulas, block partitioning by edge angle, \((\mu,b,\mathcal{T})\)-block structures, randomized cuttings, cascade refinements, and multidimensional parametric search. Two key oracle costs are
\[
T_{0\text{-Max}}(N,b,\mu)=O\!\left(N\cdot \frac{\log^2 b}{b}+\mu b\right)
\]
and
\[
T_{\text{Cascade}}(N,b,b',\mu)=O\!\left(N\cdot \frac{\log b'\,\log b}{b}+\mu b^2\right).
\]
The final bound is
\[
T_{2\text{-Max}}(N,2)=O(N),
\]
improving the earlier \(O((n+m)\log(n+m))\) exact algorithm of de Berg, Cheong, Devillers, van Kreveld, and Teillaud [2504.18352].

This geometric model establishes the most concrete meaning of convex overlap-based fidelity: overlap is the fidelity; convexity of the shapes and concavity of \(\sqrt{A(t)}\) make the fidelity landscape tractable; and exact optimization is possible in linear expected time.

## 3. Multi-shape overlap, higher-dimensional configurations, and sets of maximizers

The same overlap perspective extends beyond two polygons. For constant \(k\ge 2\), the overlap area of \(k\) convex polygons under independent translations is
\[
A(v_0;\dots;v_{k-1})=\left|\bigcap_{i=0}^{k-1}(P_i+v_i)\right|,
\]
defined on the quotient configuration space
\[
C\coloneqq
\frac{\{(v_0,\dots,v_{k-1}):v_i\in\mathbb{R}^2\}}
{\{(x,\dots,x):x\in\mathbb{R}^2\}},
\]
which factors out global translation [2307.00409]. The support of \(A\) is compact, \(A\) is continuous, and \(A^{1/2}\) is concave on its support. On regions where no event polytope intersects the interior, the intersection polygon has fixed combinatorics and the shoelace formula shows that \(A\) is a quadratic polynomial in the translation parameters [2307.00409].

For fixed \(k\), the maximization problem admits an
\[
O(n\log^{2k-3}n)
\]
algorithm, and once one maximizer is known, the full set
\[
M=\{v\in C:A(v)=A^{\max}\}
\]
can be computed in \(O(n)\) time as a convex polytope given by \(O(n)\) linear constraints [2307.00409]. A crucial structural fact is that all maximal overlaps are translates of a canonical maximal overlap polygon \(I_{\max}\), and that
\[
M\cong \prod_{i=0}^{k-1}(P_i-I_{\max})/\text{diagonal translations}.
\]
This gives an explicit convex description of the entire maximizer set.

A related three-dimensional variant considers a convex polyhedron \(P\subset\mathbb{R}^3\) and a convex polygon \(Q\) translated in \(\mathbb{R}^3\), with objective
\[
f(v)=|P\cap(Q+v)|.
\]
The overlap function is continuous and piecewise quadratic on the interior of its support, and \(f(v)^{1/2}\) is concave on its support by a Brunn–Minkowski argument [2301.02949]. The configuration space is partitioned by event polygons of three kinds: \(u-Q\) for a vertex \(u\) of \(P\), \(F-v\) for a face \(F\) of \(P\) and a vertex \(v\) of \(Q\), and \(e-e'\) for edges \(e\) of \(P\) and \(e'\) of \(Q\). A deterministic algorithm finds a maximizing translation in
\[
O(n\log^2 n)
\]
time [2301.02949].

The same paper applies the method to the maximum overlap of three convex polygons and to minimization of symmetric difference under homothety. For convex polygons \(P,Q\subset\mathbb{R}^2\), the symmetric difference satisfies
\[
A\triangle B=(A\cup B)\setminus(A\cap B),
\]
and for fixed areas minimizing \(|P\triangle \varphi(Q)|\) is equivalent to maximizing \(|P\cap \varphi(Q)|\). This places overlap-based fidelity and symmetric-difference dissimilarity in direct duality [2301.02949]. A plausible implication is that overlap-based fidelity is especially natural when area or volume is preserved, while symmetric-difference formulations are better adapted to scale changes.

## 4. Quantum fidelity, convex sets of states, and convex approximation

In quantum information, fidelity is introduced algebraically but analyzed geometrically. For density matrices \(\rho\) and \(\sigma\),
\[
F(\rho,\sigma)=\operatorname{Tr}\bigl(\sqrt{\sqrt{\rho}\,\sigma\,\sqrt{\rho}}\bigr)
=\operatorname{Tr}(\rho\#\sigma)
=\operatorname{Tr}(\sqrt{\rho\,\sigma}),
\]
and for pure states it reduces to overlap:
\[
F(|\psi\rangle\langle\psi|,|\phi\rangle\langle\phi|)=|\langle\psi|\phi\rangle|.
\]
The paper on geometric interpretations of quantum fidelity emphasizes that the state space \(S(\mathcal{H})\) is a compact, closed convex set, with pure states as extreme points and \(I/d\) as the maximally mixed center [1503.00304].

Fidelity is concave in each argument:
\[
F\bigl(\rho,\, t\sigma_1+(1-t)\sigma_2\bigr)\ge
tF(\rho,\sigma_1)+(1-t)F(\rho,\sigma_2),
\]
and this concavity drives several convex-geometric interpretations. If \(K\subset S(\mathcal{H})\) contains all pure states, then
\[
\min_{\sigma\in K}F(\rho,\sigma)=\sqrt{\lambda_{\min}(\rho)}.
\]
For quantum channels, minimizing over all CPTP maps gives
\[
\min_{\Phi}F(\rho,\Phi(\sigma))=\sqrt{\lambda_{\min}(\rho)},
\]
while minimization over unital or mixed unitary channels yields
\[
\min_{\Phi\ \text{unital or mixed unitary}}F(\rho,\Phi(\sigma))
=
F\bigl(X^{\downarrow}(\rho),X^{\uparrow}(\sigma)\bigr),
\]
the “anti-aligned eigenvalues” expression [1503.00304]. Here \(\lambda_{\min}(\rho)\) is interpreted as a convex combination coefficient measuring how close \(\rho\) is to the maximally mixed state versus the boundary of singular density matrices.

The same paper gives an explicitly geometric overlap interpretation for subspaces. If
\[
\rho_S=\frac{1}{m}P_S,\qquad \rho_T=\frac{1}{n}P_T
\]
are normalized projections onto subspaces \(S,T\), then in the commuting case
\[
F(\rho_S,\rho_T)=\sqrt{\frac{\dim(S\cap T)}{mn}},
\]
and in general
\[
F(\rho_S,\rho_T)=\frac{1}{\sqrt{mn}}\sum_{k=1}^{\ell}\cos\theta_k,
\]
where \(\{\theta_k\}\) are the canonical angles and \(\ell=\min\{m,n\}\) [1503.00304]. Thus fidelity becomes a scaled overlap of principal directions.

A second quantum line of work studies optimal convex approximations. Given a target state \(\rho\) and available states \(\rho_i\), the fidelity-based approximation problem is
\[
D_B^F(\rho)
=
1-\max_{\{p_i\}}
F\Bigl(\rho,\sum_i p_i\rho_i\Bigr),
\qquad p_i\ge 0,\ \sum_i p_i=1.
\]
For qubits and specific available sets such as the six Pauli eigenstates \(B_3\), complete exact solutions are obtained via Bloch-vector geometry and KKT conditions [2105.07336]. In the region
\[
r_x+r_y+r_z\le 1,
\]
the target state is completely represented by the convex hull of available states and
\[
D_{B_3}^F(\rho)=0.
\]
Outside this region, optimal mixtures live on faces or edges of the octahedral convex hull, so the maximizing mixture typically uses only two or three states [2105.07336]. This is convex overlap-based fidelity in the strict sense of maximizing overlap with a convex set of feasible states.

## 5. Overlap as convex program, design criterion, and aggregation law

The overlap viewpoint also appears where the primary object is not a geometric intersection area. Erdős’ minimum overlap problem studies partitions \(A,B\subset[2n]\) with \(|A|=|B|=n\), difference counts
\[
M_k=\#\{(a,b)\in A\times B:a-b=k\},
\]
and the asymptotic constant
\[
\mu=\lim_{n\to\infty}\frac{M(n)}{n},
\qquad
M(n):=\min_{A\cup B=[2n]}\max_k M_k.
\]
In its continuous form, one considers measurable \(f:[-1,1]\to[0,1]\) with \(\int_{-1}^1 f(x)\,dx=1\), its complement \(g=1-f\), and the overlap functional
\[
M(x)=\int_{-1}^1 f(t)\,g(x+t)\,dt,
\]
so that
\[
\mu=\inf_f\sup_{x\in[-2,2]}M(x).
\]
The paper translates this into a convex optimization program using moment identities, Fourier coefficients, interval averages, and second-order cone constraints, proving
\[
\mu\ge 0.379005
\]
[2201.05704]. Here overlap becomes a cross-correlation-type functional and convexity arises at the level of the feasible program.

In causal inference, Ghosh defines relaxed covariate overlap through convex hulls of treated and control covariates,
\[
\mathcal{C}_1=\operatorname{co}(\mathbf{Z}_1),\qquad
\mathcal{C}_0=\operatorname{co}(\mathbf{Z}_0),
\]
with overlap condition
\[
\mathcal{C}_0\cap \mathcal{C}_1\neq \emptyset.
\]
The distance between hulls is obtained by minimizing
\[
Q(\alpha,\beta)=\frac{1}{2}\big\|\mathbf{Z}_0^\top\alpha-\mathbf{Z}_1^\top\beta\big\|^2
\]
over simplex weights, and failure of overlap is equivalent to the existence of a separating hyperplane, linking convex hull geometry to the margin of a linear SVM [1801.00816]. This is not a fidelity in the trace or area sense, but it is an overlap-based notion of design fidelity: causal comparisons are regarded as credible only in the shared convex support region.

In lattice theory, overlap functions generalize from \([0,1]\) to bounded lattices \(L\). An \(L\)-overlap
\[
O:L^2\to L
\]
satisfies symmetry, the zero boundary
\[
O(x,y)=0_L \iff x=0_L\text{ or }y=0_L,
\]
the one boundary
\[
O(x,y)=1_L \iff x=y=1_L,
\]
and monotonicity; continuity distinguishes overlaps from quasi-overlaps [1902.00133]. A lattice-valued convex combination is built through a t-norm \(\otimes\), a t-conorm \(\oplus\), and weights \((\lambda_i)\) with \(\bigoplus_i\lambda_i=1_L\):
\[
F(x,y)=\bigoplus_{i=1}^n \lambda_i\otimes O_i(x,y).
\]
Under the stated hypotheses, convex sums of overlaps or quasi-overlaps are again overlaps or quasi-overlaps [1902.00133]. This provides an abstract aggregation-theoretic meaning of convex overlap-based fidelity.

## 6. Scope, limitations, and adjacent developments

Several misconceptions are explicitly ruled out by the cited works. First, overlap-based fidelity is not synonymous with distance. In the geometric papers, overlap is a positive similarity, whereas Hausdorff distance and Fréchet distance are negative dissimilarities [2504.18352]. In quantum theory, fidelity and trace distance induce different optimal convex approximations; for many qubit regimes, the fidelity-based optimum is closer to the target state than the trace-norm-based optimum, but the two criteria do not coincide [2105.07336]. In causal inference, convex hull overlap is not the same as scalar propensity-score overlap, because it is multivariate and geometric rather than one-dimensional [1801.00816].

Second, convexity is essential to most exact results. The linear-time algorithm of Chan and Hair is specific to convex polygons under translation and does not directly handle rotation or non-convexity [2504.18352]. For several convex polygons, \(k\) is assumed constant [2307.00409]. For the polyhedron–polygon problem and the homothety problem, the arguments rely on event structures, cuttings, and Brunn–Minkowski-type concavity that are not available in the same form for general non-convex shapes [2301.02949]. Lattice-overlap theory changes substantially when continuity is dropped, which motivates quasi-overlaps [1902.00133].

Third, the overlap perspective extends naturally into inverse problems with smooth convex data fidelities. The LiGME-SCF framework considers objectives of the form
\[
J_{\Psi_B\circ L}(x)=f(Ax)+\mu\,\Psi_B(Lx),
\]
where \(f\) is a smooth convex data fidelity, \(A\) is a linear observation operator, and \(\Psi_B\) is a generalized Moreau enhanced nonconvex regularizer. Overall convexity is guaranteed by
\[
A^*\Lambda A-\mu L^*B^*BL\succeq 0,
\]
equivalently by the requirement that \(f\circ A\) be \(\mu\)-strongly convex relative to \(q_{L^*B^*BL}\) [2509.03258]. The paper treats overlapped measurements as encoded by the linear operator \(A\), so that a convex overlap-based fidelity may take the form \(f(Ax)\) with nonquadratic choices such as Poisson and clipping likelihoods. This suggests a broader interpretation in which “overlap” is represented by linearly shared measurements rather than set intersection.

The main open frontiers named across the literature are consistent. Exact linear-time behavior is currently established in the basic two-dimensional convex-translation case [2504.18352]. Higher dimensions, rigid motions, non-convex shapes, and large-\(k\) multi-object matching remain harder [2307.00409]. In quantum settings, explicit exact formulas are presently strongest for qubits and particular available sets [2105.07336]. In causal inference, the geometric margin identifies a high-fidelity subpopulation but makes the estimand data-adaptive [1801.00816]. These limitations do not weaken the concept; they delimit the settings in which convex overlap-based fidelity is presently mathematically sharp and algorithmically effective.

Source: https://www.emergentmind.com/topics/convex-overlap-based-fidelity