We reduce from LocalMaxCut. Let G = (V, E, w) be a weighted undirected graph.
Reduction Construction. We define an instance of LocalBetweenness-Euclidean in \mathbb{R}^d called \ensuremath{\mathcal{I}}_{btw} as follows. We introduce special vertices X_1,\dots,X_d and choose a hierarchy of weights
M_3 > M_4 > \cdots > M_d > M > \sum_{(u,v) \in E} w_{uv},
such that for each k\ge3,
M_k > \underbrace{\sum_{j>k} \; 3 \tbinom{j-1}{2} M_j}_{\text{lower layer Type C constraints}}
+ \underbrace{2 |V| \tbinom{d}{2} M}_{\text{all Type B constraints}}
+ \underbrace{\sum_{(u,v) \in E} w_{uv}}_{\text{all Type A constraints}}.
Now add constraints as follows:
Type A (Edge Constraints). For each edge (u, v) \in E with weight w_{uv}, add a single constraint (u, X_1, v) with weight w_{uv}.
Type B (Isosceles Constraints). For each v \in V and every pair (X_k, X_\ell) with 1 \leq k < \ell \leq d, add two constraints (X_k, X_\ell, v) and (X_\ell, X_k, v), each with weight M. Intuitively, this will force X_k, X_\ell, v to form an isosceles triangle such that \left\| v - X_k \right\|_2 = \left\| v - X_\ell \right\|_2 \geq \left\| X_k - X_\ell \right\|_2.
Type C (Equilateral Constraints). For every triple (X_j, X_k, X_\ell) with 1 \leq j < k < \ell \leq d, we add three constraints (X_j, X_k, X_\ell), (X_k, X_\ell, X_j), (X_\ell, X_j, X_k), each with weight M_\ell (\ell is the largest index among j,k,\ell). Intuitively, this will force X_j, X_k, X_\ell to form an equilateral triangle.
Lemma. In any local optimum of \ensuremath{\mathcal{I}}_{btw}, all Type C constraints are satisfied, therefore X_1, \dots, X_d form a (d-1)-dimensional regular simplex.(In geometry, a regular simplex is the d-dimensional generalization of an equilateral triangle (2-simplex) or a regular tetrahedron (3-simplex), consisting of d+1 points with all pairwise distances equal.)
Proof (Proof of ?). We prove \{X_1,\dots,X_k\} form a regular simplex for each k=3,\dots,d. We do this by induction on k.
Base Case (k=3). If any Type C constraint over \{X_1,X_2,X_3\} is unsatisfied, moving any of them to form an equilateral triangle with the other two will satisfy at least one new constraint of weight M_3, while losing at most \sum_{j>3} \; 3 \tbinom{j-1}{2} M_j + 2 |V| \tbinom{d}{2} M + \sum_{(u,v) \in E} w_{uv} < M_3, which contradicts local optimality.
Inductive Step. Suppose \{X_1,\dots,X_{k-1}\} form a regular simplex but \{X_1,\dots,X_k\} do not. Similar to the above, moving X_k to form a regular simplex with \{X_1,\dots,X_{k-1}\} will satisfy at least one new constraint of weight M_k, while losing a total strictly less than M_k. Hence, \{X_1,\dots,X_k\} must also form a regular simplex.
Lemma. Let C be the centroid of the simplex X_1,\dots,X_d and \ell = \left\| X_1 - X_2 \right\|_2 be its edge length. Then in any local optimum, all Type B constraints are satisfied, therefore each v \in V must lie on the line through C perpendicular to the hyperplane spanned by the X_i , and moreover at distance at least \ell from every X_i. Formally, each v \in V lies on one of two opposing rays:
R^+ = \bigl\{\, C + t \, \mathbf{u} : t \ge t_0 \bigr\}, \quad
R^- = \bigl\{\, C - t \, \mathbf{u} : t \ge t_0 \bigr\},
where \mathbf{u} is the unit vector along that line and t_0 = \ell \cdot \sqrt{\frac{d+1}{2d}}.
Proof (Proof of ?). Without loss of generality, assume the simplex is in the standard position:
X_1 = (1, 0, \ldots, 0), X_2 = (0, 1, \ldots, 0), \ldots, X_d = (0, 0, \ldots, 1)
with centroid C = \left(\frac{1}{d}, \frac{1}{d}, \ldots, \frac{1}{d}\right). The edge length is \ell = \sqrt{2}.
We denote the i-th coordinate of v as v_i. The equidistant condition \left\| v - X_k \right\|_2 = \left\| v - X_\ell \right\|_2 simplifies to:
v_i^2 - 2v_i = v_j^2 - 2v_j \quad \forall i \neq j.
The only solutions are v = (\alpha, \alpha, \ldots, \alpha) for some \alpha \in \mathbb{R}.
For v = \alpha \mathbf{1}, the distance to any vertex X_k is:
\left\| v - X_k \right\|_2^2 = (\alpha - 1)^2 + (d-1)\alpha^2 = d\alpha^2 - 2\alpha + 1 \geq \ell^2 = 2.
Solving d\alpha^2 - 2\alpha + 1 \geq 2 gives:
\alpha \leq \frac{1 - \sqrt{d+1}}{d} \quad \text{or} \quad \alpha \geq \frac{1 + \sqrt{d+1}}{d}.
Expressing v relative to the centroid C = \frac{1}{d}\mathbf{1}:
v = C + t \, \mathbf{u}, \quad \mathbf{u} = \mathbf{1}/\sqrt{d}, \quad t = \sqrt{d}\left(\alpha - \frac{1}{d}\right).
Substituting the bounds for \alpha yields t \leq -\sqrt{\frac{d+1}{d}} or t \geq \sqrt{\frac{d+1}{d}}.
For arbitrary \ell > 0, scale the canonical simplex by \lambda = \ell/\sqrt{2}. The threshold t_0 scales as:
t_0 = \lambda \sqrt{\frac{d+1}{d}} = \ell \sqrt{\frac{d+1}{2d}}.
The standard basis assumption is justified because any regular simplex can be mapped to this form via a similarity transformation, which preserves the distance relationships.
Since M > \sum_{(u,v) \in E} w_{uv}, any configuration of v that violates the Type B constraints can be improved by moving v to either R^+ or R^-.
Lemma. In any local optimum, the edge constraint (u, X_1, v) is satisfied if and only if u and v lie on opposite rays R^+ and R^-.
Proof (Proof of ?). Let u = C + t_i\mathbf{u} and v = C + t_j\mathbf{u}. Since u, v lie on the line through C perpendicular to the hyperplane spanned by the X_i, we have
\begin{aligned}
\left\| u - X_1 \right\|_2 &= \sqrt{t_i^2 + \left\| X_1 - C \right\|_2^2}, \\
\left\| v - X_1 \right\|_2 &= \sqrt{t_j^2 + \left\| X_1 - C \right\|_2^2}.
\end{aligned}
We can also verify that \left\| X_1 - C \right\|_2 = \ell \sqrt{\frac{d-1}{2d}}.
Case 1: If u and v are on the same ray (e.g., R^+), then:
\left\| u - v \right\|_2 = |t_i - t_j| \leq \max(t_i, t_j) < \max(\left\| u - X_1 \right\|_2, \left\| v - X_1 \right\|_2).
Thus (u, X_1, v) is not satisfied.
Case 2: If u and v are on opposite rays R^+ and R^-, then:
\begin{aligned}
\left\| u - X_1 \right\|_2 &= \sqrt{t_i^2 + \left\| X_1 - C \right\|_2^2} = \sqrt{t_i^2 + \frac{(d-1)\ell^2}{2d}}, \\
\left\| v - X_1 \right\|_2 &= \sqrt{t_j^2 + \left\| X_1 - C \right\|_2^2} = \sqrt{t_j^2 + \frac{(d-1)\ell^2}{2d}}, \\
\left\| u - v \right\|_2 &= |t_i| + |t_j|.
\end{aligned}
We need to show that \left\| u - v \right\|_2 \geq \max(\left\| u - X_1 \right\|_2, \left\| v - X_1 \right\|_2).
To show \left\| u - v \right\|_2 \geq \left\| u - X_1 \right\|_2, we need (|t_i| + |t_j|)^2 \geq t_i^2 + \frac{(d-1)\ell^2}{2d}. Expanding:
t_i^2 + 2 |t_i t_j| + t_j^2 \geq t_i^2 + \frac{(d-1)\ell^2}{2d} \implies 2 |t_i t_j| + t_j^2 \geq \frac{(d-1)\ell^2}{2d}.
Since |t_i|, |t_j| \geq t_0, substitute |t_i| = |t_j| = t_0 (worst case for the inequality):
2t_0^2 + t_0^2 = 3t_0^2 \geq \frac{(d-1)\ell^2}{2d}.
Substitute t_0 = \ell \sqrt{\frac{d+1}{2d}}:
3 \cdot \frac{(d+1)\ell^2}{2d} \geq \frac{(d-1)\ell^2}{2d} \implies 3(d+1) \geq d-1.
This holds for all d \geq 2.
Symmetrically we can show \left\| u - v \right\|_2 \geq \left\| v - X_1 \right\|_2. Thus (u, X_1, v) is satisfied.
Special Case d=1. We note that in the case of d=1, this construction degenerates into only one special vertex X_1, with only edge constraints. We claim that this construction is still correct, as the vertices on different sides of X_1 encode a cut. Since “no local moves can improve the objective function” is a stronger condition than “no local moves of vertices except X_1 can improve the objective function”, any local optimum of \ensuremath{\mathcal{I}}_{btw} is also a local max cut.
This concludes the proof of Theorem Theorem.