Extended Edge Elimination in TSP

• Extended Edge Elimination Conditions are geometric criteria certifying that an edge cannot be part of any optimal tour in the Euclidean TSP.
• They extend traditional 2‑opt incompatibility by employing two independent witnesses and a 3‑opt replacement argument using cones and δ‑isolation.
• Probabilistic analysis shows that the Hougardy–Schroeder method reduces surviving edges to Θ(n), outperforming conventional Θ(n²) residual methods.

Searching arXiv for the cited work and closely related papers on Extended Edge Elimination Conditions in TSP. arxiv_search(query="Extended Edge Elimination Conditions traveling salesman problem Hougardy Schroeder local elimination", max_results=10) Extended Edge Elimination Conditions, commonly abbreviated EEC in the Euclidean traveling salesman problem, are sufficient conditions for certifying that a given edge is useless, meaning that it cannot belong to any optimum tour. In the TSP literature, the term is most closely associated with the Hougardy–Schroeder approach, which replaces purely pairwise 2-opt incompatibility by a stronger geometric certification based on cones, $\delta$-isolation, extremal points, and a 3-edge improvement argument (Hougardy et al., 2014). In the probabilistic setting of random Euclidean instances, these conditions admit a sharp asymptotic analysis: under independent points in the unit square with densities bounded above and below by positive constants, Hougardy–Schroeder elimination leaves an expected $\Theta(n)$ edges, whereas the non-recursive Jonker–Volgenant rule leaves $\Theta(n^2)$ edges (Zhong, 2018). The same phrase also appears in other parts of graph theory, notably in contraction-preserving $\mathcal{H}$-freeness and in edge-elimination graph polynomials, but those uses are conceptually distinct (Ibrahim et al., 2022, Trinks, 2011).

1. Edge elimination in the TSP and the emergence of EEC

For a symmetric TSP instance on a complete graph $G=(V,E)$ with length function $l:E\to\mathbb{R}_+$, an edge $e=pq$ is useless if it can be proved that no optimum Hamiltonian tour contains it. Edge elimination seeks a subset $E'\subseteq E$ that still contains all optimum tours while removing provably useless edges (Hougardy et al., 2014). The baseline local certificate is compatibility: two edges $pq$ and $xy$ are compatible, denoted $\Theta(n)$0, if

$\Theta(n)$1

Otherwise they are incompatible, and no optimum tour can contain both, since a 2-opt move would strictly improve the tour (Hougardy et al., 2014).

EEC arise as an extension of this pairwise logic. Instead of certifying uselessness through a single incompatible edge pair, they use two independent witnesses and a 3-opt-type contradiction. In the combinatorial formulation of Hougardy and Schroeder, one fixes an edge $\Theta(n)$2 and identifies two distinct potential points $\Theta(n)$3 and $\Theta(n)$4 with coverings that constrain how their neighbors can appear in any optimum tour containing $\Theta(n)$5. If two explicit inequalities hold, then whichever local adjacency pattern the tour chooses, one of two 3-opt replacements strictly shortens the tour, so $\Theta(n)$6 is useless (Hougardy et al., 2014).

The Euclidean specialization strengthens this abstract framework with geometry. In that setting, potentiality can be certified through cone constructions around witness points, and the main theorem can be checked in constant time per witness once local geometric parameters are known. This makes EEC a preprocessing method rather than only a structural theorem. The practical motivation is direct: eliminating edges before branch-and-cut or heuristic search reduces the candidate graph while preserving all optimum tours (Hougardy et al., 2014).

2. Geometric structure of the Euclidean EEC

The probabilistic analysis of EEC considers $\Theta(n)$7 independent random points $\Theta(n)$8 in the unit square $\Theta(n)$9, where each $\Theta(n^2)$0 has a density $\Theta(n^2)$1 satisfying

$\Theta(n^2)$2

for fixed constants $\Theta(n^2)$3, and uses Euclidean edge costs $\Theta(n^2)$4 (Zhong, 2018). For a point $\Theta(n^2)$5 and edge $\Theta(n^2)$6, the notation $\Theta(n^2)$7 denotes the orthogonal projection of $\Theta(n^2)$8 onto the line through $\Theta(n^2)$9 (Zhong, 2018).

The core Euclidean objects are $\mathcal{H}$0-isolated vertices, cones, and extremal points. Fix $\mathcal{H}$1; for the probabilistic analysis, $\mathcal{H}$2 is used (Zhong, 2018). For a vertex $\mathcal{H}$3 and any point $\mathcal{H}$4, let $\mathcal{H}$5 be the intersection of the ray $\mathcal{H}$6 with the circle centered at $\mathcal{H}$7 of radius $\mathcal{H}$8. Then for an edge $\mathcal{H}$9 and vertex $G=(V,E)$0, the Hougardy–Schroeder cones are

$G=(V,E)$1

$G=(V,E)$2

A vertex $G=(V,E)$3 is $G=(V,E)$4-isolated if for all $G=(V,E)$5, $G=(V,E)$6 (Zhong, 2018).

These cones encode a necessary adjacency restriction. If $G=(V,E)$7 is $G=(V,E)$8-isolated and $G=(V,E)$9 belongs to an optimal tour, then both neighbors of $l:E\to\mathbb{R}_+$0 in that tour lie in $l:E\to\mathbb{R}_+$1 (Zhong, 2018). On the circle $l:E\to\mathbb{R}_+$2 of radius $l:E\to\mathbb{R}_+$3 around $l:E\to\mathbb{R}_+$4, the arcs $l:E\to\mathbb{R}_+$5 and $l:E\to\mathbb{R}_+$6 determine extremal points: $l:E\to\mathbb{R}_+$7 is the point on $l:E\to\mathbb{R}_+$8 maximizing $l:E\to\mathbb{R}_+$9, and $e=pq$0 is defined analogously on $e=pq$1 (Zhong, 2018).

In the implementation-oriented treatment for EUC_2D, the same geometry is expressed through $e=pq$2, $e=pq$3, $e=pq$4, cone angles $e=pq$5, and a lower bound $e=pq$6 on the angle between the two incident tour edges at $e=pq$7 (Hougardy et al., 2014). With

$e=pq$8

and

$e=pq$9

a vertex $E'\subseteq E$0 is strongly potential with respect to $E'\subseteq E$1 if the cone-existence condition

$E'\subseteq E$2

holds and $E'\subseteq E$3 (Hougardy et al., 2014). This constant-time certification is the geometric engine that turns the abstract theorem into an efficient elimination routine.

3. The elimination conditions themselves

In the abstract combinatorial formulation, let $E'\subseteq E$4 and $E'\subseteq E$5 be distinct potential points for a fixed edge $E'\subseteq E$6, with coverings $E'\subseteq E$7 and $E'\subseteq E$8, respectively. If $E'\subseteq E$9 and $pq$0, and if

$pq$1

and

$pq$2

then $pq$3 is useless (Hougardy et al., 2014). The proof considers the two possible 3-opt replacements forced by the neighbor positions of $pq$4 and $pq$5 in any optimum tour containing $pq$6. The inequalities guarantee that both candidate improvements are strictly positive, so no such tour can be optimal (Hougardy et al., 2014).

The Euclidean EEC of Hougardy–Schroeder specialize these minima through cone geometry and extremal points. If $pq$7 and $pq$8 are two distinct $pq$9-isolated potential points with respect to $xy$0, with

$xy$1

and if

$xy$2

$xy$3

then $xy$4 is useless (Zhong, 2018). This is the formulation commonly called Extended Edge Elimination Conditions in the Euclidean TSP literature (Zhong, 2018).

EEC sit strictly above classical 2-opt incompatibility. The Jonker–Volgenant non-recursive rule eliminates $xy$5 if there exists a vertex $xy$6 such that for all $xy$7,

$xy$8

$xy$9

or equivalently, if the union of the improvement hyperbola regions

$\Theta(n)$00

$\Theta(n)$01

contains no vertex other than $\Theta(n)$02 (Zhong, 2018). The probabilistic contrast between EEC and this 2-opt-style rule is one of the central results of the later analysis.

The same 3-opt perspective also underlies the Close Point Elimination theorems and the metric excess refinement in the broader Hougardy–Schroeder framework (Hougardy et al., 2014). A plausible implication is that EEC are best understood not as a single inequality but as a family of local certificates whose common feature is the forced existence of a strictly improving local exchange whenever the target edge is present.

4. Probabilistic behavior on random Euclidean instances

Under the random geometric model with independent points in $\Theta(n)$03 and densities bounded by $\Theta(n)$04 and $\Theta(n)$05, let

$\Theta(n)$06

and

$\Theta(n)$07

The main asymptotic result is

$\Theta(n)$08

(Zhong, 2018).

For Hougardy–Schroeder EEC, the analysis replaces the exact certification by a weaker criterion based on certifying test regions that still implies the original EEC. Each test region $\Theta(n)$09 is a disk of radius $\Theta(n)$10 with $\Theta(n)$11 and contains two subregions $\Theta(n)$12 and $\Theta(n)$13, each of area $\Theta(n)$14; any pair $\Theta(n)$15, $\Theta(n)$16 are separated by at least $\Theta(n)$17 and at most $\Theta(n)$18 (Zhong, 2018). A test region is certifying if it contains exactly two vertices, one in each subregion, and no others. For long edges, meaning edges of length at least $\Theta(n)$19, the existence of a certifying test region implies the Hougardy–Schroeder conditions, hence the edge is useless (Zhong, 2018).

The probabilistic control comes from canonical test regions and negative correlation bounds. If $\Theta(n)$20 is the number of test regions placed along one side of a long edge, then each region has probability bounded away from $\Theta(n)$21 and $\Theta(n)$22 of being certifying as $\Theta(n)$23, and for some $\Theta(n)$24,

$\Theta(n)$25

Combining this with the distribution of edge lengths yields

$\Theta(n)$26

and therefore a linear expected number of surviving edges (Zhong, 2018).

For the non-recursive Jonker–Volgenant rule, the analysis instead lower-bounds the area of improvement hyperbolas. Using an $\Theta(n)$27-border strip $\Theta(n)$28 and layers $\Theta(n)$29, the paper proves that for endpoints $\Theta(n)$30 and suitable $\Theta(n)$31 there exists $\Theta(n)$32 such that

$\Theta(n)$33

Hence the probability that the hyperbola union is empty is at most

$\Theta(n)$34

and summing over layers shows that the elimination probability of a short edge is bounded by a constant multiple of $\Theta(n)$35 (Zhong, 2018). This leaves a constant fraction of all short edges uneliminated, implying $\Theta(n)$36 residual size.

This asymptotic separation formalizes a practical distinction. EEC are not merely somewhat stronger than a 2-opt-style hyperbola criterion; on the random Euclidean model analyzed in the paper, they change the expected residual graph from dense to sparse (Zhong, 2018).

5. Algorithmic realization and computational impact

Hougardy and Schroeder translate the abstract theorem into a multi-stage elimination algorithm whose main part runs in $\Theta(n)$37 time for an $\Theta(n)$38-vertex instance (Hougardy et al., 2014). Preprocessing builds a kd-tree and computes $\Theta(n)$39 values for all vertices in $\Theta(n)$40 total time. In Step 1, for each edge $\Theta(n)$41, candidate witnesses are chosen as vertices nearest to the midpoint of $\Theta(n)$42; the implementation caps the number of candidates at $\Theta(n)$43. Each candidate is tested for strong potentiality in constant time, and once two strongly potential witnesses are found, constant-time lower bounds for the minima in the Main Edge Elimination Theorem are used to try to eliminate $\Theta(n)$44 (Hougardy et al., 2014).

Step 2 combines the Main Edge Elimination and Close Point conditions on a reduced set of neighbor pairs, and Step 3 performs a bounded-depth backtrack search, exploring small local path configurations and pruning them whenever Close Point or Main Edge conditions eliminate a path edge (Hougardy et al., 2014). On TSPLIB EUC_2D instances, Step 1 typically reduces about $\Theta(n)$45 edges to about $\Theta(n)$46, and combining Step 2 and Step 3 reduces the remaining graph to approximately $\Theta(n)$47 edges (Hougardy et al., 2014). On instance d2103, Concorde alone needed $\Theta(n)$48 seconds on a 2.9GHz Xeon, while preprocessing for $\Theta(n)$49 seconds reduced the graph to $\Theta(n)$50 edges and Concorde then took $\Theta(n)$51 seconds, for a total speedup exceeding $\Theta(n)$52 (Hougardy et al., 2014).

A later line of work places EEC inside a more general local-elimination framework. An edge $\Theta(n)$53 is eliminated by constructing an $\Theta(n)$54-witness family $\Theta(n)$55 such that every tour containing $\Theta(n)$56 contains some witness set $\Theta(n)$57, and then proving that each $\Theta(n)$58 is nowhere $\Theta(n)$59-optimal, meaning that every tour class containing $\Theta(n)$60 admits a strictly improving $\Theta(n)$61-opt move (Cook et al., 2023). This framework strictly generalizes classical edge elimination and encompasses the Extended Edge Elimination conditions of Hougardy–Schroeder (Cook et al., 2023). It includes fast 2-opt and 3-opt tests,

$\Theta(n)$62

and

$\Theta(n)$63

together with brute-force $\Theta(n)$64 search and solver-assisted local improvement certification (Cook et al., 2023).

The computational outcomes reported for this local-elimination framework are on a larger scale than the original WG 2014 implementation. On all TSPLIB instances with at least $\Theta(n)$65 points, together with random and unsolved Euclidean instances, the complete-graph edge sets were reduced to under $\Theta(n)$66 edges in all but two cases, and for the three large unsolved instances repeated elimination reduced the graphs to under $\Theta(n)$67 edges; verified sparse sets later reached $\Theta(n)$68–$\Theta(n)$69 edges with average degree below $\Theta(n)$70 (Cook et al., 2023). This suggests that EEC-style local certification remains effective even when embedded in a broader witness-family and solver-assisted regime.

Method	Certification style	Reported residual size
Hougardy–Schroeder Step 1	Strongly potential witnesses and Main Edge Elimination	about $\Theta(n)$71 (Hougardy et al., 2014)
Hougardy–Schroeder full algorithm	Step 1 plus Close Point and bounded backtrack	approximately $\Theta(n)$72 (Hougardy et al., 2014)
Local elimination framework	Witness families and nowhere $\Theta(n)$73-optimality	under $\Theta(n)$74 in all but two large instances; under $\Theta(n)$75 for three unsolved instances (Cook et al., 2023)

6. Scope, limitations, and other uses of the term

The Euclidean probabilistic analysis is specific to the plane and to the random model with independent points and densities bounded above and below by $\Theta(n)$76 and $\Theta(n)$77 (Zhong, 2018). The arguments rely on cones, hyperbolas, circular arcs, border strips, and test regions in $\Theta(n)$78, and the paper does not generalize them to other metrics or higher dimensions (Zhong, 2018). Likewise, the asymptotic comparison with Jonker–Volgenant concerns only the non-recursive part of that method; recursive variants are excluded because of their additional complexity (Zhong, 2018). Even after EEC reduce the candidate graph to $\Theta(n)$79 edges in expectation, Euclidean TSP remains NP-hard, since there are instances with $\Theta(n)$80 non-useless edges (Zhong, 2018).

The more implementation-oriented Hougardy–Schroeder results also have a limited metric scope. The Main Edge Elimination and Close Point theorems apply to arbitrary symmetric TSP instances, but the constant-time geometric certification of potentiality uses EUC_2D-specific rounding bounds and cone geometry (Hougardy et al., 2014). In non-Euclidean or non-spatial metric instances, the same inequalities can still be used, but the geometric acceleration is unavailable (Hougardy et al., 2014). The local-elimination framework of 2023 removes some of this dependence by formulating elimination through witness families and nowhere $\Theta(n)$81-optimality for general symmetric lengths, using geometry only as a heuristic for choosing local neighborhoods (Cook et al., 2023).

Outside the TSP literature, the phrase “extended edge elimination conditions” is used in different senses. In contraction theory for hereditary graph classes, the paper “Edge Contraction and Line Graphs” gives a necessary and sufficient condition for strong $\Theta(n)$82-freeness: if $\Theta(n)$83 is $\Theta(n)$84-free and not critically $\Theta(n)$85-exist, then

$\Theta(n)$86

(Ibrahim et al., 2022). There the subject is preservation of forbidden-induced-subgraph structure under edge contraction, not TSP preprocessing. Similarly, in graph polynomial theory, the edge elimination polynomial $\Theta(n)$87 is the universal invariant satisfying a delete–contract–extract recurrence, and the covered components polynomial

$\Theta(n)$88

is a substitution instance of it (Trinks, 2011, 0801.1600). Those usages concern algebraic recurrences rather than elimination of candidate tour edges.

Taken together, these strands show that “Extended Edge Elimination Conditions” is not a uniform graph-theoretic term. In Euclidean TSP it denotes the Hougardy–Schroeder family of local 3-opt-based certificates, with rigorous geometric and probabilistic analyses and strong computational consequences (Zhong, 2018, Hougardy et al., 2014). In other areas it names broader elimination or contraction frameworks. The TSP usage remains the most specific and technically developed meaning attached to the term.