The Salesman's Improved Paths:
3
1
+
Integrality Gap and
2
34
Approximation Ratio
András Sebล and Anke van Zuylen
s-t Path TSP
Given:
vertex set ๐, distance metric c: ๐ × ๐ โ โโฅ0 , ๐ , ๐ก โ
๐
Find:
shortest path from ๐ to ๐ก that visits every vertex in ๐.
Why does Christofidesโs algorithm not directly give a
3/2-approximation?
Christofides for s-t path TSP
โข Find a minimum spanning tree ๐.
โข Add a ๐๐-join ๐ฝ, where ๐๐ is the set of vertices
with the wrong degree parity in ๐
โ ๐ , ๐ก โ ๐๐ if ๐ , ๐ก has even degree in ๐
โ ๐ฃ โ ๐๐ if ๐ฃ โ ๐\{๐ , ๐ก} has odd degree in ๐
Analysis:
โข ๐ ๐ โค ๐๐๐
โข ๐ ๐ฝ โค
1
๐๐๐
2
Does not hold for s-t
path TSP
LP based analysis
Feasible region is a
subset of the
spanning tree
polytope ๏ ๐ ๐ โค
๐๐๐๐ฟ๐
Subtour elimination LP
Min ๐๐๐ฅ
s.t. ๐ฅ ๐ฟ ๐ โฅ 2, for all ๐, ๐ โฉ {๐ , ๐ก} even
๐ฅ ๐ฟ ๐ โฅ 1, for all ๐, ๐ โฉ {๐ , ๐ก} = 1
0 โค ๐ฅ ๐ โค 1, for all ๐
๐๐-join polyhedron
Min ๐๐๐ฆ
s.t. ๐ฆ ๐ฟ ๐ โฅ 2, for all ๐, ๐ โฉ ๐๐ odd
๐ฆ ๐ โฅ 0, for all ๐
โ
๐ฅ /2 will only potentially
violate constraints for ๐
such that ๐ โฉ {๐ , ๐ก} = 1,
and
|๐ โฉ ๐๐| = odd
(i.e., if ๐บ has an even
number of edges in ๐น(๐ผ))
Conclusion: no 3/2 becauseโฆ
โ
๐ฅ /2 will violate the ๐๐-join constraints for ๐ -๐ก
cuts ๐ such that
โข ๐ has an even number of edges in ๐, and
โข ๐ is โnarrowโ: ๐ฅ โ ๐ < 2
Will show next: Simple proof of 8/5 ratio [Sebล
โ13] for โBest-of-Many-Christofidesโ (BOMC)
BOMC
[An, Kleinberg, Shmoysโ12]
S denotes also the incidence
vector of the edge set S
Randomized Algorithm:
โข Let ๐ฅ โ be an optimal solution to the subtour
elimination LP.
โข Express ๐ฅ โ as a convex combination of spanning
trees: ๐ฅ โ = ๐ ๐๐ ๐
โข Let ๐ฎ be a random spanning tree, ๐ ๐ฎ = ๐ = ๐๐
โข Add a ๐๐ฎ -join ๐ฅ๐ฎ
BOMC: Derandomize this algorithm (by iterating over
all trees with ๐๐ > 0) and output the cheapest solution
Bounding ๐ผ[๐ ๐ฅ๐ฎ ]
โข Let ๐ซ๐ ๐ก be the (incidence vector of the) ๐ โ ๐ก
path in ๐ฎ, and ๐โ = ๐ผ ๐ซ๐ ๐ก
โข Then, ๐ฎ\๐ซ๐ ๐ก is a ๐๐ฎ -join, so
๐ผ ๐ ๐ฅ๐ฎ โค ๐ผ ๐ฎ\๐ซ๐ ๐ก = ๐(๐ฅ โ ) โ ๐(๐โ )
โข Next, we will use ๐โ (in two ways) to complete
๐ฅ โ 2 to a (fractional) ๐๐ฎ -join
โ
Completing ๐ฅ 2 to a ๐๐-join
โข Option 1: โOne Size Fits Allโ
๐ฅโ
2
+
๐โ
2
is in the T-join polyhedron for any T of even size
โข Option 2: โTailored Completionโ
โข Natural way to assign ๐โ to s-t cuts ๐:
๐๐ ๐ = ๐(๐ฎ โฉ ๐ = {๐})
โข Need to complete only on narrow cuts ๐ such that |๐ โฉ ๐|
even: ๐(|๐ฎ โฉ ๐| even)โค ๐ ๐ฎ โฉ ๐ โฅ 2 โค ๐ฅ โ ๐ โ 1
โข Note that this implies
๐๐ ๐ = ๐ ๐ฎ โฉ ๐ = 1 โฅ 2 โ ๐ฅ โ (๐)
๐โ๐
So
๐ฅโ
2
๐๐
+ 2
is at least 1 on ๐
โ
Completing ๐ฅ 2 to a ๐๐-join
โข Option 1: โOne Size Fits Allโ
๐ผ ๐ฅ๐ฎ
๐ฅโ 1 โ
โค + ๐
2 2
Combine the two
options by replacing
this ½ by ๐พ
โข Option 2: โTailored Completionโ
๐ฅโ 1
โ Q โ 1)
๐
๐ผ ๐ฅ๐ฎ โค +
(๐ฅ
2 2 ๐๐
P(|๐ โฉ ๐ฎ| even)
The parity correction vector
โข โBasic Parity Correctionโ
๐ฅโ
+ ๐พ๐โ
2
Now, optimize over ๐พ
1
โ ๐พ= .
8
1
If ๐พ = 8 this multiplier is
โข โParity Completionโ
1
at
most
for any ๐
โ
8
๐ฅ ๐
1โ
โ๐พ
2
โ
๐
(๐ฅ
Q โ 1)
๐๐
โ
2 โ ๐ฅ (๐)
pQ Q = P ๐ โฉ ๐ฎ = 1 โฅ 2 โ ๐ฅ โ (๐)
P(|๐ โฉ ๐ฎ| even)
Bounding ๐ผ[๐ ๐ฅ๐ฎ ]
โข Already had: ๐ฎ\๐ซ๐ ๐ก is a ๐๐ฎ -join, so
๐ผ ๐ ๐ฅ๐ฎ โค ๐ผ ๐ ๐ฎ\๐ซ๐ ๐ก = ๐(๐ฅ โ ) โ ๐(๐โ )
โข Now, we also have
๐ผ ๐ ๐ฅ๐ฎ
๐ฅโ 1 โ 1 โ
โค ๐( + ๐ + ๐ )
2 8
8
Minimum of these two is at most 3/5๐(๐ฅ โ )
๏ ๐ผ ๐ ๐ฎ + ๐ ๐ฅ๐ฎ โค 8/5๐(๐ฅ โ )
How to improve? Observations:
๐ฅโ
2
โข Even when ๐ฎ โฉ ๐ = 1, we โpayโ + ๐พ๐โ for
parity correction across ๐
โข Parity completion uses
๐๐ ๐ = ๐(๐ฎ โฉ ๐ = {๐})
Idea:
โข Delete ๐ from ๐ฎ if ๐ฎ โฉ ๐ = {๐} for a narrow
cut ๐ to โsaveโ for later
๐ is a lonely edge (in ๐ in ๐) if ๐ โฉ ๐ = {๐} and
๐ฅโ ๐ < 2
Best-of-Many With Deletion
โข Let ๐ฅ โ be an optimal solution to the subtour
elimination LP.
โข Express ๐ฅ โ as a convex combination of
spanning trees: ๐ฅ โ = ๐ ๐๐ ๐
โข Let ๐ฎ be a random spanning tree, ๐ ๐ฎ = ๐ =
๐๐
โข Let โฑ(๐ฎ) be the forest after deleting the
lonely edges from ๐ฎ
โข Add a ๐โฑ(๐ฎ) -join ๐ฅโฑ(๐ฎ)
Game Plan for the Analysis
โข Bound ๐ฅโฑ(๐ฎ)
โข Note that โฑ(๐ฎ) โ ๐ฅโฑ(๐ฎ) may not be connected
โ Find the culprit (โbadโ edges in ๐ฅโฑ(๐ฎ) )
โ Bound the probability of bad edges by choosing the
convex combination that defines ๐ฎ carefully
โข Bound the cost of reconnecting โฑ(๐ฎ) โ ๐ฅโฑ(๐ฎ)
with a doubled spanning tree
Bounding ๐ฅโฑ(๐ฎ)
โข Basic Parity Correction
๐ฅโ
+ ๐พ๐โ
2
โข Parity Completion for ๐ s.t. โฑ(๐ฎ) โฉ ๐ โฅ 2, even
๐ฅโ ๐
1โ
โ๐พ
2
๐๐ 2 โ ๐ฅ โ (๐)
๐:|โฑ(๐ฎ)โฉ๐|โฅ2
โข Parity Completion for ๐ s.t. โฑ(๐ฎ) โฉ ๐ = 0
๐ฅโ ๐
๐(๐ฎโฉ๐) 1 โ
โ๐พ
2
๐: โฑ ๐ฎ โฉ๐ =0
Bounding ๐ผ[๐ฅโฑ(๐ฎ) ]
โข Basic Parity Correction
๐ฅโ
+ ๐พ๐โ
2
โข Parity Completion for ๐ s.t. โฑ(๐ฎ) โฉ ๐ โฅ 2, even
๐ฅโ ๐
1โ
โ๐พ
2
๐ฅ โ ๐ โ 1 ๐๐ 2 โ ๐ฅ โ (๐)
๐ ๐:|โฑ(๐ฎ)โฉ๐|โฅ2
โข Parity Completion for ๐ s.t. โฑ(๐ฎ) โฉ ๐ โฅ 0
โ
๐ฅ
๐
๐
๐(๐ฎ โฉ๐๐) 1 โ
โ๐พ
2
๐
๐:|โฑ(๐ฎ)โฉ๐|โฅ2
Problem: โฑ(๐ฎ) โ ๐ฅโฑ(๐ฎ) may not be
connected
Dashed: lonely edges removed from ๐
Red: ๐๐น(๐) -join
โข Bad edge: edge crossing more than one โlonelyโ
cut
โข Can we prevent that the ๐๐น(๐) -join has bad
edges?
May contribute
bad edges
Bounding ๐ฅโฑ(๐ฎ)
โข Basic Parity Correction
๐ฅโ
+ ๐พ๐โ
2
Contribute no bad
edges
โข Parity Completion for ๐ s.t. โฑ(๐ฎ) โฉ ๐ โฅ 2, even
๐ฅโ ๐
1โ
โ๐พ
2
๐๐ 2 โ ๐ฅ โ (๐)
๐:|โฑ(๐ฎ)โฉ๐|โฅ2
โข Parity Completion for ๐ s.t. โฑ(๐ฎ) โฉ ๐ = 0
๐ฅโ ๐
๐(๐ฎโฉ๐) 1 โ
โ๐พ
2
๐: โฑ ๐ฎ โฉ๐ =0
Bounding ๐ฅโฑ(๐ฎ)
โข Basic Parity Correction
๐ฅโ
+ ๐พ๐โ
2
How about ๐๐ ?
โข Parity Completion for ๐ s.t. โฑ(๐ฎ) โฉ ๐ โฅ 2, even
๐ฅโ ๐
1โ
โ๐พ
2
๐๐ 2 โ ๐ฅ โ (๐)
๐:|โฑ(๐ฎ)โฉ๐|โฅ2
โข Parity Completion for ๐ s.t. โฑ(๐ฎ) โฉ ๐ = 0
๐ฅโ ๐
๐(๐ฎโฉ๐) 1 โ
โ๐พ
2
๐: โฑ ๐ฎ โฉ๐ =0
No bad edges through parity
completion
In general, ๐๐ may contain bad edges, but if we
โ choose the convex combination of ๐ฅ โ carefully,
and
โ slightly redefine the notion of โlonely edgesโ and
๐๐ ,
โฆ we can make sure that parity completion does
not contribute bad edges
Layered Convex Combination
Let ๐ฅ โ = ๐ ๐๐ ๐, and for each ๐, let ๐ฌ ๐ โ {๐: ๐ โฉ ๐ =
{๐} for some narrow cut ๐}.
๐ฌ ๐ = narrow cuts that
contain the lonely edges
This is a layered convex combination if
1. For any ๐ such that ๐๐ > 0, and narrow cuts ๐, ๐ โฒ :
if ๐ โ ๐ฌ ๐ and ๐ฅ โ ๐ โฒ โค ๐ฅ โ ๐ then also Qโฒ โ ๐ฌ ๐ .
2. The ๐-weight of trees ๐ such that ๐ โ ๐ฌ ๐ is at least
2 โ ๐ฅ โ (๐) for every narrow cut ๐.
๐
Ensures that ๐ (๐) = ๐(๐ฎ โฉ
๐ = {๐}, ๐ โ ๐ฌ(๐ฎ)) satisfies
the properties we used of
the original ๐๐
๐
๐ contributes no bad edges
โข If ๐๐ contains a bad edge ๐ for ๐น(๐), then ๐ โ ๐โฒ for some
๐ โฒ โ ๐ฌ(๐).
โข If ๐๐ is used for our ๐๐น(๐) -join, then ๐ โฉ ๐น ๐ โฅ 2, even,
so ๐ โ ๐ฌ(๐).
By property 1: ๐ฅ โ ๐ > ๐ฅ โ (๐โฒ).
โข ๐๐ ๐ > 0, there exists ๐โฒ with ๐ ๐โฒ > 0 such that ๐ โฒ โฉ ๐ =
{๐}, ๐ โ ๐ฌ(๐ โฒ ).
โข But since ๐ โ ๐ โฒ , then ๐ โฒ โ ๐ฌ ๐ โฒ .
By property 1: ๐ฅ โ ๐โฒ > ๐ฅ โ (๐).
Contradiction.
Layered Convex Combination
โข First introduced by [Gottschalk and Vygen,
โ15] and used in analysis of BOMC
โข Proof of existence in [GV15], but no
polynomial time algorithm
โข We show: layered convex combination can be
found in strongly polynomial time using
Edmondโs matroid partition algorithm
Game Plan for the Analysis
๏ผ Bound ๐ฅโฑ(๐ฎ)
๏ผ Note that โฑ(๐ฎ) โ ๐ฅโฑ(๐ฎ) may not be connected
๏ผFind the culprit (โbadโ edges in ๐ฅโฑ(๐ฎ) )
๏ผBound the probability of bad edges by choosing the
convex combination that defines ๐ฎ carefully
(only
๐ฅโ
2
part of ๐ฅโฑ(๐ฎ) contributes bad edges)
โข Bound the cost of reconnecting โฑ(๐ฎ) โ ๐ฅโฑ(๐ฎ)
with a doubled spanning tree
Reconnection Bound
Dashed: lonely edges removed from ๐
Red: ๐๐น(๐) -join (assume all come from
Let ๐ฅ๐น(๐) be a random ๐๐น(๐) -join (with distribution given by the fractional
solution of our analysis).
Divide ๐ฅ๐น(๐) into bad edges (coming from
edges (all other edges).
๐ฅโ
2
part of the solution) and good
All but
one!
For each bad edge ๐ in ๐ฅ๐น(๐) , put back the lonely edges (doubled) for all the
lonely cuts that contain ๐.
For lonely cut ๐, the expected number of copies we put back of its lonely
edge is ๐ฅ โ (๐).
๐ฅโ ๐ โ 1
๐ฅโ
)
2
Reduce Reconnection Bound by 1 per
Cut: Transportation Problem
Edges ๐, supply ๐ฅ โ (๐)
Lonely cuts ๐, demand 1
๐
๐
Solution exists iff for any subset ๐ฌโฒ of the lonely cuts,
โ (๐) โฅ |๐ฌ โฒ |.
๐ฅ
๐โ๐โ๐ฌโฒ
This condition is easily checked to be satisfied
because of the subtour constraints
Game Plan for the Analysis
๏ผ Bound ๐ฅโฑ(๐ฎ)
๏ผ Note that โฑ(๐ฎ) โ ๐ฅโฑ(๐ฎ) may not be connected
๏ผ Find the culprit (โbadโ edges in ๐ฅโฑ(๐ฎ) )
๏ผ Bound the probability of bad edges by choosing the convex
combination that defines ๐ฎ carefully
(only
๐ฅโ
2
part of ๐ฅโฑ(๐ฎ) contributes bad edges)
๏ผ Bound the cost of reconnecting โฑ(๐ฎ) โ ๐ฅโฑ(๐ฎ) with a
doubled spanning tree: ๐ ๐๐ ๐ฅ โ ๐ โ 1
Putting it all together
Forest:
๐ผ๐ฎ โ
๐
โ
๐
=
๐ฅ
โ
๐
๐
๐
๐
Correcting parity:
โ ๐
๐ฅ
1โ
โ๐พ
๐ฅโ
2
โ
โ ๐ โ1
๐
+ ๐พ๐ +
๐ฅ
๐
2
2 โ ๐ฅ โ (๐)
โ ๐
๐ฅ
+
๐๐ 1 โ
โ๐พ
2
๐
Reconnection:
๐๐ ๐ฅ โ ๐ โ 1
๐
1
Choose ๐พ = , then the
16
combined multipliers for
๐๐ โs become โค 0.
โ expected total cost is at
3
1
most ๐ ๐ฅ โ + ๐(๐โ )
2
16
The new ratio
โข Already had: ๐ฎ โ (๐ฎ\๐ซ๐ ๐ก ) is a solution of cost
2๐(๐ฅ โ ) โ ๐(๐โ )
โข Now, we also have a solution of cost
3
1
โ
๐ ๐ฅ + ๐(๐โ )
2
16
Minimum of these two is at most
1
2
+
1
34
๐(๐ฅ โ )
© Copyright 2026 Paperzz