--- title: 'Maximum Distance Rule: Theory and Applications' url: https://www.emergentmind.com/topics/maximum-distance-rule type: topic --- # Maximum Distance Rule: Theory and Applications Searching arXiv for recent and directly relevant uses of “maximum distance rule” and closely related terms. Search 1: phrase “maximum distance rule” {"query":"\"maximum distance rule\"","max_results":10,"sort_by":"relevance"} Search 2: geometric “maximum distance problem” {"query":"\"maximum distance problem\" arXiv","max_results":10,"sort_by":"relevance"} Search 3: simplex and pivoting-rule usage {"query":"simplex \"maximum distance rule\"","max_results":10,"sort_by":"relevance"} Search 4: Brownian maximum distance between leader and laggard {"query":"Brownian walkers maximum distance leader laggard","max_results":10,"sort_by":"relevance"} “Maximum Distance Rule” is not a single canonical term across mathematics. In the arXiv literature, it denotes several non-equivalent constructions built around one of three patterns: an exact law for a maximum separation, a rule that chooses an object by a distance-normalized criterion, or an extremal family that attains the largest admissible distance under fixed structural constraints. Representative instances include an exact Brownian first-passage asymptotic for the maximum Leader–Laggard span [1006.5834], a primal simplex pivoting rule \(j_m\in \arg\min\{\bar c_j/\|a_j\|\}\) [2507.04672], and geometric “maximum distance problems” that minimize \(1\)-dimensional Hausdorff measure subject to an \(r\)-neighborhood covering constraint [2004.07323]. Several papers also emphasize that nearby phrases in social choice or connectomics are not names of a literal maximum-distance rule, which makes the term intrinsically field-dependent [1610.01902]. ## 1. Terminological scope and unifying themes Across the literature considered here, the phrase is used in at least four distinct senses. In stochastic processes, it denotes a tail law for the largest separation attained before a stopping event. In continuous and combinatorial optimization, it denotes either a rule that normalizes a gain by a distance-like quantity or a variational problem that controls a maximum covering distance. In coding theory, it denotes extremal families that maximize minimum distance under fixed ambient parameters. In graph theory and discrete geometry, it denotes maximin separation principles such as furthest-neighbour incidences or maximal safety distance [1006.5834]. The common structure is not a single formal definition but a recurring extremal template. One either maximizes a distance quantity directly, minimizes an objective under a maximum-distance constraint, or studies objects whose minimum distance is as large as the relevant Singleton-type or geometric bound permits. This suggests that “Maximum Distance Rule” functions more as a family resemblance term than as a standardized definition. A recurrent source of confusion is terminological overreach. In distance-rationalized social choice, the governing construction is explicitly \[ R(E)=\arg\min_r d(E,K_r), \] so the framework is minimum-distance-to-consensus, not maximum-distance-to-consensus [1610.01902]. In voting under local perturbation, “maximum displacement distance” is a metric constraint on ballot edits, not a winner-selection rule [1901.08711]. In connectomics, the relevant principle is an exponential distance rule, not a hard maximum-distance cutoff [2307.06061]. ## 2. Brownian first-passage usage: maximum Leader–Laggard separation A precise stochastic meaning appears in the three-walker Brownian problem. Three independent walkers \(x_1(t),x_2(t),x_3(t)\) evolve on a line according to \[ \frac{dx_i}{dt}=\eta_i(t), \] with \[ \langle \eta_i(t)\rangle=0,\qquad \langle \eta_i(t)\eta_j(t')\rangle=2D_i\,\delta_{ij}\delta(t-t'). \] Initially \(x_1\le x_2\le x_3\); the left-most particle is the Leader, and the Laggard is the particle currently farthest to the right relative to the Leader. With \[ y_2(t)=x_2(t)-x_1(t),\qquad y_3(t)=x_3(t)-x_1(t), \] the span is \[ z(t)=\max\{y_2(t),y_3(t)\}. \] The process stops at the first time \(t_s\) when either \(y_2(t)\) or \(y_3(t)\) hits zero, and the central observable is \[ m=\max_{0\le t\le t_s} z(t). \] The object of interest is \(P(m\mid y_2,y_3)\) [1006.5834]. The exact large-\(m\) law is \[ P(m\mid y_2,y_3)\sim A(y_2,y_3)\,m^{-\delta}, \] with \[ \delta=\frac{2\pi-\theta}{\pi-\theta}, \qquad \theta=\cos^{-1}\!\left(\frac{D_1}{\sqrt{(D_1+D_2)(D_1+D_3)}}\right). \] This is the paper’s central “maximum distance rule”: the tail exponent is controlled entirely by the angle parameter \(\theta\), itself determined by the diffusion constants. Because \(0\le \theta\le \pi/2\), the exponent satisfies \[ 2\le \delta\le 3. \] The asymptotic amplitude \(A(y_2,y_3)\) is also determined exactly, and is homogeneous in the initial separations: \[ A(\lambda y_2,\lambda y_3)=\lambda^{\pi/(\pi-\theta)}A(y_2,y_3). \] The derivation proceeds by passing from particle coordinates to the correlated diffusion of \((y_2,y_3)\), introducing the cumulative distribution \[ F(y_2,y_3\mid L)=\int_0^L P(m\mid y_2,y_3)\,dm, \] and deriving the backward equation \[ (D_1+D_2)\frac{\partial^2F}{\partial y_2^2} +(D_1+D_3)\frac{\partial^2F}{\partial y_3^2} +2D_1\frac{\partial^2F}{\partial y_2\partial y_3}=0 \] on the square \(0\le y_2,y_3\le L\). A linear change of variables removes the mixed derivative, mapping the problem to Laplace’s equation on a parallelogram whose interior angle is \(\theta\). A Schwarz–Christoffel map sends that parallelogram to the upper half-plane, Poisson’s formula gives the harmonic solution, and the \(L\to\infty\) limit yields the tail exponent [1006.5834]. Two special cases are explicit. For an immobile Leader, \(D_1=0\) and \(D_2=D_3=D\), one gets \(\theta=\pi/2\) and \(\delta=3\), with \[ P(m\mid y_2,y_3)\simeq \frac{1}{4\pi^2}\Gamma^4(1/4)\,\frac{y_2y_3}{m^3}. \] For equal diffusion constants \(D_1=D_2=D_3=D\), one has \(\theta=\pi/3\) and \(\delta=5/2\). The same paper connects this maximum law to first-passage persistence through \[ Q(t)\sim t^{-\theta_1},\qquad \theta_1=\frac{\pi}{2(\pi-\theta)}, \] and the scaling relation \[ \delta=1+\frac{\theta_1}{H}, \] with \(H=1/2\) for Brownian motion, hence \(\delta=1+2\theta_1\) [1006.5834]. ## 3. Linear programming usage: the simplex maximum distance rule In linear programming, the term denotes a primal simplex pivoting rule for the standard-form problem \[ \min\ c^\top x,\qquad Ax=b,\qquad x\ge 0, \] with basis \(B\), nonbasis \(N\), reduced costs \(\bar c_N\), and dictionary \[ x_B=A_B^{-1}b-\bar A_Nx_N,\qquad \bar A_N=A_B^{-1}A_N. \] The maximum distance rule chooses the entering index by \[ j_m \in \arg\min\left\{\frac{\bar c_j}{\|a_j\|}\;\middle|\; j\in N\right\}, \] where \(a_j\) is the \(j\)-th column of the original matrix \(A\) [2507.04672]. Geometrically, the rule selects the nonbasic variable whose corresponding dual constraint has the largest unit violation. Its formal resemblance to the steepest-edge rule is deliberate: steepest edge uses a denominator involving the current transformed column \(\bar a_j\), whereas maximum distance uses the fixed original column norm \(\|a_j\|\). The paper therefore presents it as a lower-cost normalization scheme that is similar in spirit to steepest edge but cheaper to maintain [2507.04672]. The analysis introduces three global parameters: \(\delta\), the minimum positive entry appearing in any BFS; \(\gamma\), the maximum positive entry appearing in any BFS; and a column-norm ratio \(\beta\), described as the ratio of the minimum to the maximum column norm of \(A\), with \(0<\beta\le 1\). The key bridge to Dantzig’s rule is \[ \Delta_{m,t}\ge \beta\,\Delta_{d,t}, \] where \(\Delta_{d,t}\) is the magnitude of the most negative reduced cost at iteration \(t\) and \(\Delta_{m,t}\) is the reduced-cost magnitude selected by the maximum distance rule. This yields the core contraction estimate \[ c^\top x^{t+1}-z^* \le \left(1-\frac{\beta\delta}{m\gamma}\right)\left(c^\top x^t-z^*\right) \] whenever the BFS changes [2507.04672]. From that contraction lemma, the paper proves two deterministic bounds on the number of different BFSs generated. The first is \[ \left\lceil \frac{m\gamma}{\beta\delta} \log\left(\frac{c^\top x^0-z^*}{c^\top \bar x-z^*}\right)\right\rceil, \] where \(\bar x\) is a second-best BFS. The second, independent of objective ratios, is \[ (n-m)\left\lceil \frac{m\gamma}{\beta\delta}\log\left(\frac{m\gamma}{\delta}\right)\right\rceil. \] The paper states these as bounds on the number of different BFSs without requiring nondegeneracy; under nondegeneracy they become iteration bounds. It presents them as the first deterministic theoretical bounds for the maximum distance rule [2507.04672]. ## 4. Geometric optimization: the maximum distance problem A second major usage is geometric. The maximum distance problem asks for the shortest connected one-dimensional set whose \(r\)-neighborhood covers a compact target \(E\). One formulation is \[ \Lambda(E,r):= \inf\left\{ \mathcal H^1(\Gamma): \Gamma\subset \mathbb R^2 \text{ is a curve of finite length and } \mathcal N(\Gamma,r)\supset E \right\}, \] equivalently, \[ \lambda(E,s):=\inf\{\mathcal{H}^1(K):K\ \text{is closed, connected and } E\subset B(K,s)\}. \] Here the “maximum distance” is the allowed covering radius, while the objective is the \(1\)-dimensional Hausdorff measure [2004.07323]. For planar \(E\subset\mathbb R^2\), the problem admits an exact cover-and-connect interpretation. If \(X\) is a finite set of ball centers with \(B(X,s)\supset E\), and \(T_X\) is a minimum spanning tree on \(X\), then minimizing over such coverings recovers the continuous optimum: \[ \sigma'(E,s)=\lambda(E,s). \] The technical core is a lemma asserting that the \(s\)-neighborhood of a Lipschitz curve can be covered by finitely many radius-\(s\) balls whose centers are connectable by another Lipschitz curve with arbitrarily small excess length [2004.07323]. The small-\(r\) asymptotic theory separates rectifiable from fractal regimes. If \(E\) is contained in a rectifiable curve and \(r_j\to 0^+\), there exists a sequence of \(r_j\)-maximum distance minimizers \(\Gamma^*_{r_j}\) converging in Hausdorff distance to a rectifiable curve \(\Gamma^*_0\supset E\), with \[ \lim_{j\to\infty}H^1(\Gamma^*_{r_j})=H^1(\Gamma^*_0) =\inf\{H^1(\Gamma):\Gamma\supset E\}. \] Thus the \(r\to 0\) limit recovers a shortest rectifiable curve containing \(E\), i.e. an Analyst’s Traveling Salesman-type object [2309.08055]. For weak \((\alpha,\beta)\)-bi-Hölder targets, the minimizing length obeys power-law bounds \[ \frac{1}{C} r^{\frac{\beta-1}{\beta}} \le \Lambda(B(\widehat\gamma,r),r) \le C r^{\frac{\alpha-1}{\alpha}}, \] and for the \(\frac13\)-von Koch snowflake \(S\), with \(\alpha=\log_4(3)\), the bounds match: \[ \frac{1}{C}r^{\frac{\alpha-1}{\alpha}} \le \Lambda(B(S,r),r) \le C r^{\frac{\alpha-1}{\alpha}}. \] For path-connected planar \(E\), any minimizer stays within \(O(r)\) of \(E\), yielding Hausdorff convergence \(\Gamma_r^*\to E\) as \(r\to 0^+\) [2309.08055]. A recent finite-scale refinement connects the problem to the Analyst’s Traveling Salesperson Theorem. For compact \(E\subset\mathbb R^2\), \[ |E|-2r+\sum_{Q\in\mathcal D}\max\left\{\beta_E(Q)-\frac r{|Q|},0\right\}^2|Q| \;\lesssim\; \Lambda(E,r) \;\lesssim\; |E|+\sum_{Q\in\mathcal D}\max\left\{\beta_E(Q)-\frac r{|Q|},0\right\}^2|Q|. \] The decisive multiscale term is therefore the truncated square sum \[ \sum_{Q\in\mathcal D}\max\left\{\beta_E(Q)-\frac r{|Q|},0\right\}^2|Q|, \] which measures only the excess flatness above the tolerance \(r\) [2603.14119]. ## 5. Discrete geometry and graph-theoretic max-distance principles In Euclidean discrete geometry, a maximum-distance assignment appears as the furthest-neighbour digraph. For a finite set \(S\subset\mathbb R^d\), \[ D_S(x)=\max_{y\in S}|xy|, \] and the furthest-neighbour count is \[ e_D(S)=\sum_{x\in S}\bigl|\{y\in S:|xy|=D_S(x)\}\bigr|. \] The extremal quantity is \[ g_d(n):=\max\{e_D(S):S\subset\mathbb R^d,\ |S|=n\}. \] For fixed \(d\ge 4\), the asymptotic leading term is \[ g_d(n)\sim \left(1-\frac1{\lfloor d/2\rfloor}\right)n^2, \] and for sufficiently large \(n\), \[ g_d(n)=2M_d(n), \] where \(M_d(n)\) is the maximum number of diameter pairs. The extremal sets are exactly Lenz configurations, and every furthest-neighbour distance equals the global diameter: \[ D_S(x)=\operatorname{diam}(S)\qquad\forall x\in S. \] Thus, in high-dimensional Euclidean space, the maximum-distance rule collapses to the doubled diameter graph of a Lenz construction [1108.4817]. In finite graph theory, the same phrase can denote a maximin safety problem for two moving actors. For admissible walks \(f,g\) on a connected graph \(G\), \[ m_G(f,g)=\min\{d_G(f(i),g(i)): i\in N_l\}, \] and the goal is to maximize this minimum separation while both actors collectively visit all vertices. The paper distinguishes three movement models: traditional, active, and lazy, producing three parameters \(M_s(G)\), \(M_d(G)\), and \(M_c(G)\). It proves exact equivalence with strong, direct, and Cartesian graph spans, and establishes the sharp comparison \[ |\text{direct span}(G)-\text{Cartesian span}(G)|\le 1. \] The difference is realized by explicit examples, while traditional movement always dominates the other two [2208.11366]. The same work computes exact values for several graph classes. For paths \(P_n\), \[ \text{strong span}(P_n)=\left\lfloor \frac{n-1}{2}\right\rfloor,\qquad \text{direct span}(P_n)=1,\qquad \text{Cartesian span}(P_n)=0. \] For hypercubes \(Q_n\), \[ \text{strong span}(Q_n)=n,\qquad \text{direct span}(Q_n)=n,\qquad \text{Cartesian span}(Q_n)=n-1. \] For perfect binary trees \(BT_h\), all three equal \(h-1\). These are distinct graph-theoretic realizations of a maximum-distance principle, but they are unrelated to the stochastic, simplex, or coding-theoretic usages [2208.11366]. ## 6. Coding-theoretic meanings: extremal distance families In coding theory, “maximum distance” is typically an extremal property rather than an algorithmic rule. One instance is the Lee-metric function \[ \Phi(n,K,p^t)=\max\{d_L(C)\mid C\subseteq (\mathbb Z_{p^t})^n \text{ is a linear code of rank }K\}, \] where \[ d_L(C)=\min\{w_L(c)\mid c\in C,\ c\ne 0\}. \] Codes attaining \(\Phi(n,K,p^t)\) are called Maximum Lee Distance codes, or MLDR/MLD in the paper’s terminology. The work derives new Singleton-like upper bounds, including \[ \Phi(n,K,p^t)\le p^{t-1}\lfloor \mu_p(n-K)\rfloor \] in the long-code regimes where MDS socles cannot exist, and a sharper MDS-range bound \[ \Phi(n,K,p^t)\le p^{t-1}\left( \left\lfloor \mu_p(n-K+1)\frac{p-1}{p}\right\rfloor+1 \right) \] when \(K\le p\) and \(n\le p+1\) [2408.11107]. A second instance is maximum rank distance. In the rank metric on \(M_{m,n}(\mathbb F_q)\), \[ d(X,Y)=\operatorname{rank}(X-Y), \] an \((m,n,q;s)\)-rank-metric code satisfies \(|\mathcal C|\le q^{\,n(m-s+1)}\), and codes meeting this bound are MRD codes. One paper studies puncturing generalized twisted Gabidulin codes: if \(\mathcal X\) is an \((n,n,q;s)\)-MRD code and \(A\) is an \(m\times n\) matrix of rank \(m\) with \(n-sn}(n-t_i)\right). \] Codes attaining it are optimum distance flag codes. The paper’s main criterion shows that one need inspect only one or two projected subspace codes: if \[ a=\max\{i:2t_i\le n\},\qquad b=\min\{i:2t_i\ge n\}, \] then \(\mathcal C\) is an ODFC iff \(\mathcal C_a\) and \(\mathcal C_b\) are constant-dimension codes of maximum distance and \[ |\mathcal C_a|=|\mathcal C_b|=|\mathcal C|. \] Specializing to orbit flag codes, this becomes a stabilizer criterion, and Singer groups are then used to construct families attaining both maximum distance and maximum possible size [2109.00270]. ## 7. Misnomers, near-neighbours, and common misconceptions Several papers are relevant precisely because they deny that they study a literal maximum-distance rule. In distance rationalization of social rules, the core construction is minimum-distance-to-consensus, \[ R(E)=\arg\min_r d(E,K_r), \] and a max-distance analogue is discussed only as an extrapolation outside the paper’s framework [1610.01902]. In local distance constrained bribery, the closest object is maximum displacement distance, \[ d_{\text{max displacement}}(\succ_1,\succ_2)= \max_{a\in\mathcal A}|\operatorname{pos}(a,\succ_1)-\operatorname{pos}(a,\succ_2)|, \] which is a per-voter feasibility constraint, not a voting rule [1901.08711]. Distance restricted manipulation studies robust and existential manipulation under per-voter Kendall–Tau balls: \[ d_{KT}(\succ_i,\succ_i')\le \delta_i\qquad \forall i. \] The “strong” version asks for a manipulative vote that succeeds for every profile in the uncertainty set, while the “weak” version asks only for one such profile. Again, this is a maximum-radius uncertainty model, not a winner-selection rule [1909.03162]. In brain connectomics, the paper explicitly distinguishes the exponential distance rule from any hard maximum-distance threshold. The basic law is \[ \rho_\lambda(s)=\lambda e^{-\lambda s}, \] with \(\lambda=1/\langle s\rangle\); long connections are exponentially suppressed, not forbidden. The paper therefore argues against reading the observed distance dependence as a literal maximum-distance rule [2307.06061]. The principal misconception is therefore semantic. The phrase often designates whatever role a maximal distance happens to play in a model—tail variable, admissibility radius, pivot normalization, covering threshold, or extremal metric optimum. A plausible implication is that cross-disciplinary uses should be treated as homonyms rather than as variants of one theory. Source: https://www.emergentmind.com/topics/maximum-distance-rule