Computers & Operations Research 36 (2009) 637 – 654

www.elsevier.com/locate/cor

Invited review
A comparative analysis of several asymmetric traveling salesman
problem formulations
Temel Öncana , İ. Kuban Altınelb , Gilbert Laportec,∗
a Department of Industrial Engineering, Galatasaray University, Ortaköy, İstanbul 34357, Turkey
b Department of Industrial Engineering, Boğaziçi University, Bebek, İstanbul 34342, Turkey
c Canada Research Chair in Distribution Management, HEC Montréal, 3000, chemin de la Côte-Sainte-Catherine, Montréal, Canada, H3T 2A7

Available online 3 December 2007

Abstract
In this survey, a classification of 24 asymmetric traveling salesman problem (ATSP) formulations is presented. The strength of their
LP relaxations is discussed and known relationships from the literature are reviewed. Some new relationships are also introduced,
and computational results are reported.
! 2007 Elsevier Ltd. All rights reserved.

Keywords: Integer linear programming; Asymmetric traveling salesman problem; Formulations; Projections

1. Introduction

The asymmetric traveling salesman problem (ATSP) is defined on a directed graph G=(V , A), where V ={1, . . . , n}
is the vertex set, A = {(i, j ) : i, j ∈ V } is the arc set, and a non-symmetric cost matrix (cij ) is defined on A. To simplify
the notation, we associate variables to arcs (i, i) but we force these variables to be equal to 0 by setting cii = ∞. The
ATSP consists of determining a least cost Hamiltonian circuit or tour over G. The problem is commonly interpreted as
that of determining an optimal salesman’s tour over n cities. The ATSP is NP-hard even if the costs are Euclidean [1].
As is the case with most combinatorial optimization problems, exact algorithms for the ATSP combine polyhedral
results with enumeration. The efficiency of the enumeration depends on the strength of the LP (or linear) relaxation
of a given formulation. It is possible to state that given two formulations, the one yielding the larger relaxation value
is better. The strengths of LP relaxations, or equivalently the strengths of two formulations, can also be compared by
using polyhedral information. Suppose two different formulations F1 and F2 are stated in the same space of variables
x ∈ Rp and the objective is to minimize. Let P (F1 ) and P (F2 ) be the polyhedra associated with these formulations. If
P (F1 ) ⊂ P (F2 ), then F1 is a better formulation than F2 since the lower bound obtained by solving the LP relaxation
of F1 is at least equal to the one obtained by solving the LP relaxation of F2 .
Different formulations of a given problem can be frequently stated in terms of different sets of variables, as is the
case with some ATSP formulations that use an extended set of variables with respect to another formulation. If F3 is a
formulation with this property, it is possible to project the extended polyhedron P (F3 ) of F3 into the subspace of the

∗ Corresponding author.
E-mail addresses: [email protected] (T. Öncan), [email protected] (İ.K. Altınel), [email protected] (G. Laporte).

0305-0548/$ - see front matter ! 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cor.2007.11.008
638 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

original variables without losing any integer solution, namely any of the tours, and compare the lower bound obtained
over the projected polyhedron TP(F3 ). To be precise, given the polyhedron
P (F3 ) = {(x, y) ∈ Rp × Rq : Ax + By !b},
where A, B and b have m rows, the projection of P (F3 ) into the subspace of x variables, or into Rp , is
TP(F3 ) = {x ∈ Rp : there exists y ∈ Rq such that (x, y) ∈ P (F3 )},
and if TP(F3 ) ⊂ P (F1 ), then F3 is a better formulation since its LP bound is larger than that of F1 [2].
Several attempts have been made to compare a number of existing ATSP formulations, see, e.g., Wong [3], Padberg and
Sung [4], Langevin et al. [5], Gouveia and Voss [6], Altınel et al. [7] and Orman and Williams [8]. The main motivation of
this paper is to survey the existing ATSP formulations and to establish new relationships between them. The formulations
are described in Section 2. Section 3 summarizes the known relationships between these formulations, while Section 4
introduces some new ones. Comparative computational results are presented in Section 5, and conclusions follow in
Section 6.

2. Classical formulations and their subtour elimination constraints

Many ATSP formulations consist of an assignment problem with integrality and subtour elimination constraints.
They use binary xij variables equal to 1 if and only if arc (i, j ) belongs to the optimal solution. The basic model is as
follows:
!n !n
(ATSP) : Minimize cij xij (1)
i=1 j =1
n
!
subject to xij = 1, i = 1, . . . , n, (2)
j =1
! n
xij = 1, j = 1, . . . , n, (3)
i=1
0 !xij !1, i, j = 1, . . . , n, (4)
xij = 0, 1, i, j = 1, . . . , n, (5)
{(i, j ) : xij = 1, i, j = 2, . . . , n} does not contain subtours. (6)
In this formulation, the assignment constraints (2) and (3) ensure that each vertex is incident to one outgoing arc and
one incoming arc. Constraints (6) break subtours in the set {2, . . . , n}. Together with (2) and (3) they also eliminate
subtours containing vertex 1, and the subtour elimination constraints for i = 1 become redundant. To be precise with
notation, note that the projected polyhedron TP(F ) of an ATSP formulation F is stated in the space of xij variables.

2.1. The Dantzig, Fulkerson and Johnson (DFJ) formulation

In their seminal work Dantzig et al. [9] have formulated the subtour elimination constraints as
!
xij !|S| − 1, S ⊆ {2, . . . , n}, 2 !|S|!n − 1. (7)
i,j ∈S

These inequalities are facet defining [10] and can be gradually added to the formulation through a branch-and-bound
scheme, which results in an optimal solution. In practice, relatively few subtour elimination constraints are needed.

2.2. Circuit packing (CP) inequalities

The following inequalities are known as the circuit packing or simply circuit inequalities.
!
xij !|C| − 1 for all circuits DC = (VC , C), VC ⊆ {2, . . . , n}. (8)
(i,j )∈C
 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654 639

The formulation obtained by replacing inequalities (6) with circuit inequalities (8) is a valid ATSP formulation [10].

2.3. The Miller, Tucker and Zemlin (MTZ) formulation

The earliest known extended formulation of the ATSP is due to Miller et al. [11]. It was originally proposed for a
vehicle routing problem (VRP) where the number of vertices of each route is limited. It uses ui variables to define
the order in which each vertex i is visited on a tour. The subtour elimination constraints of the MTZ formulation are
stated as
ui − uj + (n − 1)xij !n − 2, i, j = 2, . . . , n, (9)
1 !ui !n − 1, i = 2, . . . , n. (10)
We should mention that the ui variables are unrestricted in the original paper. Simple bounds (10) were introduced
later on, and do not affect the LP bound obtained on P(MTZ).

2.4. The Desrochers and Laporte (DL) formulation

The compact polynomial representation of P(MTZ) extends naturally to a number of variants of the VRP [12].
Another advantage of the MTZ formulation is that the subtour elimination constraints (9) and (10) can be incorporated
into other type of problem formulations together with stronger constraints. Motivated by these facts, Desrochers and
Laporte [12] have lifted subtour elimination constraints (9) and (10) to obtain the stronger forms:
ui − uj + (n − 1)xij + (n − 3)xj i !n − 2, i, j = 2, . . . , n, (11)
n
! n
!
1 + (n − 3)xi1 + xj i !ui !n − 1 − (n − 3)x1i − xij , i = 2, . . . , n. (12)
j =2 j =2

They have shown that inequalities (11) are facet defining. The authors have also proved that (12) are valid inequalities,
which have been shown to be facet defining by Driscoll [13].

2.5. The Sherali and Driscoll (SD) formulation

Sherali and Driscoll [14] have also strengthened the relaxations of the MTZ formulation for the ATSP. The authors
apply only a partial first level version of the Reformulation–linearization technique (RLT) [15] to a nonlinear refor-
mulation of the MTZ subtour elimination constraints. The new variables yij assume a non-zero value representing the
order of arc (i, j ) on the tour. Their O(n2 ) subtour elimination constraints are as follows:
n
!
yij + (n − 1)xi1 = ui , i = 2, . . . , n, (13)
j =2

n
!
yij + 1 = uj , j = 2, . . . , n, (14)
i=2

xij !yij !(n − 2)xij , i, j = 2, . . . , n, (15)

uj + (n − 2)xij − (n − 1)(1 − xj i )!yij + yj i !uj − (1 − xj i ), i, j = 2, . . . , n, (16)
1 + (1 − x1j ) + (n − 3)xj 1 !uj !(n − 1) − (n − 3)x1j − (1 − xj 1 ), j = 2, . . . , n. (17)

2.6. The Gavish and Graves (GG) formulation

A large class of extended ATSP formulations are known as commodity flow formulations [5], where the additional
variables represent commodity flows through the arcs and satisfy additional flow conservation constraints. These models
640 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

belong to three classes: single commodity flow (SCF), two-commodity flow (TCF) and multi-commodity flow (MCF)
formulations.
The earliest SCF formulation is due to Gavish and Graves [16]. The additional continuous non-negative variables
gij describe the flow of a single commodity to vertex 1 from every other vertex. A set of O(n2 ) subtour elimination
constraints is given by
n
! n
!
gj i − gij = 1, i = 2, . . . , n, (18)
j =1 j =2

0 !gij !(n − 1)xij , i = 1, . . . , n; j = 2, . . . , n. (19)

An interpretation similar to the one given for the additional variables ui of the MTZ formulation can also be applied
to variables gij . These can be viewed as the number of arcs included on the path from vertex 1 and arc (i, j ) in the
optimal tour [17].
As an example of TCF formulation, we can mention the formulation due to Finke et al. [18] (FCG). For the sake of
brevity, we do not present FCG here since Langevin et al. [5] have shown that FCG and GG are equivalent.

2.7. Multi-commodity flow (MCF) formulations

Wong [3] was the first to formulate the ATSP as an MCF model using additional non-negative variables to describe
the flow of 2(n − 1) commodities between vertex 1 and the other vertices k, namely commodities k taking the value
(k,1) (1,l) (1,l)
wij , k = 2, . . . , n and commodities l taking the value wij , l = 2, . . . , n. The flow variable wij is equal to one if
(k,1)
and only if the commodity going from 1 to l flows on arc (i, j ). The flow variable wij is equal to one if and only if
the commodity going from k to 1 flows on arc (i, j ). His O(n3 ) subtour elimination constraints are defined as follows:
n
! n
!
(1,l) (1,l)
wij − wj i = 0, i, l = 2, . . . , n; i ̸ = l, (20)
j =1 j =1

n
! n
!
(1,l) (1,l)
w1j − wj 1 = 1, l = 2, . . . , n, (21)
j =2 j =2

n
! n
!
(1,i) (1,i)
wij − wj i = −1, i = 2, . . . , n, (22)
j =1 j =1

n
! n
!
(k,1) (k,1)
wij − wj i = 0, i, k = 2, . . . , n; i ̸ = k, (23)
j =1 j =1

n
! n
!
(k,1) (k,1)
w1j − wj 1 = −1, k = 2, . . . , n, (24)
j =2 j =2

n
! n
!
(i,1) (i,1)
wij − wj i = 1, i = 2, . . . , n, (25)
j =1 j =1

(1,l)
0 !wij !xij , i, j = 1, . . . , n; l = 2, . . . , n, (26)
(k,1)
0 !wij !xij , i, j = 1, . . . , n; k = 2, . . . , n. (27)
(1,l) (k,1)
Note that only the xij are subject to integrality since wij and wij variables are integer whenever the xij are, because
the flow matrix for commodities k and l is totally unimodular. In fact the variable xij can be interpreted as the capacity
bound on arc (i, j ).
 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654 641

Modifications of the Wong MCF model (WONG) were proposed by Langevin [19] and Loulou [20]. The Langevin
formulation (LANGEVIN) is a restriction of WONG. The Loulou formulation (LOULOU) is a restriction of both
LANGEVIN and WONG. Hence, WONG, LOULOU and LANGEVIN are equivalent. For the sake of conciseness, we
do not present LOULOU and LANGEVIN.
Another MCF formulation proposed by Claus [21] (CLAUS) uses only (n − 1) commodities. Langevin et al. [5]
observed that CLAUS can be obtained from WONG by eliminating half of the flow variables and related constraints.
(k,1)
Claus [21] uses only non-negative flow variables wij and all the constraints in which they appear. Hence, the Claus
subtour elimination constraints consists of (23)–(25) and (27).
2.8. The Fox, Gavish and Graves (FGG) formulations

The next formulations exploit a relationship between the ATSP and machine scheduling. Consider the following
single machine scheduling problem where n jobs must be scheduled on a single machine. The machine is in the initial
state which we denote by job 1. We assume that the machine will again be in that state after processing all of the
remaining n − 1 jobs. A set of n − 1 jobs, denoted by 2, . . . , n, are to be performed on a single machine and a set up
cost cij k is incurred when job j is processed immediately after job i in the kth position. The objective is to determine a
cheapest sequence of the n jobs. This problem is known as the Time Dependent TSP (TDTSP), a generalization of the
standard TSP where the cost of any given arc depends on its position in the tour.
Fox et al. [22] have proposed three time-dependent formulations. The one we present below has four set of constraints
and will be denoted as FGG4 in the sequel. Let rij k be a variable equal to 1 if and only if vertex j is visited immediately
after vertex i in the kth position. The FGG4 model is
n !
! n !
n
(FGG4) : Minimize cij k rij k (28)
i=1 j =1 k=1
!n ! n
subject to rij k = 1, i = 1, . . . , n, (29)
j =1 k=1
! n ! n
rij k = 1, j = 1, . . . , n, (30)
i=1 k=1
!n ! n
rij k = 1, k = 1, . . . , n, (31)
i=1 j =1
!n ! n n !
! n
kr ij k − kr j ik = 1, i = 2, . . . , n, (32)
j =1 k=2 j =1 k=1
rij k ∈ {0, 1}, i, j, k = 1, . . . , n. (33)
Fox et al. [22] have noted that constraints (31) are not needed and may be dropped. Their second formulation, namely
FGG3, consists of the objective function (28) and, constraints (29), (30), (32) and (33).
The third formulation by Fox et al. [22], denoted by FGG2, is more compact than the FGG4 formulation. It uses
(34) as an aggregation of constraints (29)–(31):
! n !
n ! n
rij k = n. (34)
i=1 j =1 k=1

To ensure the validity of the FGG2 formulation, Fox et al. [22] assume that rij 1 = rj in for i = 2, . . . , n, r1ik = 0 for
k = 2, . . . , n and ri1k = 0 for k = 1, . . . , n − 1.
2.9. The Gouveia and Pires (GP) formulations

Gouveia and Pires [17] have first proposed generalized formulations of the MTZ. They have shown that these formu-
lations are stronger than the MTZ formulation. Then, Gouveia and Pires [23] have devised even stronger formulations
by generalizing CLAUS.
642 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

2.9.1. Generalizations of the MTZ formulation

In their first extension of the MTZ formulation Gouveia and Pires [17], have shown that the set of inequalities

xij + vki − vkj !1, i, j, k = 2, . . . , n, (35)

xij − vij !0, i, j = 2, . . . , n, (36)

xij + vj i !1, i, j = 2, . . . , n (37)

eliminate subtours and, together with constraints (2)–(5), provide a valid ATSP formulation. Additional variables vij
indicate whether vertex i is on the path from vertex 1 to vertex j. This first formulation of Gouveia and Pires [17] will
be referred to as the GP1 formulation. Recall that variables ui of the MTZ formulation can be interpreted as the number
of intermediate vertices on the path from vertex 1 to vertex i of the optimal tour. These variables are related to the new
variables vij as follows:
n
!
ui = vki , i = 2, . . . , n. (38)
k=2

Constraints (35) can be lifted in two different ways [17]

xj i + xij + vki − vkj !1, i, j, k = 2, . . . , n, (39)

xkj + xik + xij + vki − vkj !1, i, j, k = 2, . . . , n. (40)

New formulations, called GP2 and GP3, obtained by replacing (35) first with (39) and then (40), respectively, are valid
for the ATSP. In their fourth formulation, which we refer to as GP4, Gouveia and Pires [17] consider constraints (39)
and (40) together with constraints (2)–(5), (36) and (37).

2.9.2. Generalizations of the CLAUS formulation

Gouveia and Pires [23] have considered an equivalent version of CLAUS. Although both formulations are equivalent
and yield the same LP bound, the order of indexing of commodity flow variables is different. They use fkij instead
(k,1)
of wij , which only affects their interpretation. Namely, fkij = 1 indicates that arc (i, j ) is on the path from vertex
k = 2, . . . , n to vertex 1. Their MCF formulation consists of the assignment constraints (2), (3), integrality restrictions
(5), and the following subtour elimination constraints [23]:
n
! n
!
fkj i − fkij = 0, j, k = 2, . . . , n; j ̸= k, (41)
i=1 i=1

n
!
fjj i = 1, j = 2, . . . , n, (42)
i=1

fkij !xij , i, j = 1, . . . , n; k = 2, . . . , n, (43)

fkij "0, i, j = 1, . . . , n; k = 2, . . . , n. (44)

Gouveia and Pires [23] have stated that the variables vij are related to the new variables fkij as follows:
n
!
fkij = vki , i, k = 2, . . . , n. (45)
j =1

It is also interesting to observe that the left-hand side of (45) counts the number of arcs leaving vertex i on the path
from vertex k to vertex 1.
 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654 643

Gouveia and Pires [23] have generalized inequalities (39) and (40), respectively, as follows:
!
xpq + vki − vkj !|S| − 1, i, j, k = 2, . . . , n, S ⊆ {2, . . . , n}, |S| "2; k ∈/ S; i, j ∈ S, (46)
p,q∈S
! ! !
vki + (xim + xkm ) + (xmj + xmk ) + xpq + xik + xkj + xij !1 + |S| + vkj ,
m∈S m∈S p,q∈S
i, j, k = 2, . . . , n; S ⊂ V \{1, i, j, k}. (47)

Note that for |S| = 2 constraints (46) reduce to (39) and, for |S| = 0 constraints (47) reduce to (40). Consequently,
the GP5 formulation is obtained by replacing inequalities (40) of GP3 with inequalities (47) and GP6 is obtained by
replacing inequalities (39) of GP2 formulation with inequalities (46) [23].
Finally, observing that inequalities (39) and the linear transformation (45) are not redundant for the polyhedron
P(CLAUS), Gouveia and Pires [23] have proposed the GP7 formulation which consists of (2)–(5), (3) and (41)–(45).
Gouveia and Pires [23] have further generalized the GP7 formulation by replacing constraints (39) of GP7 formulation
with (46). Hence, the GP8 formulation consists of (2)–(5) and (41)–(46).

2.10. The Sarin, Sherali and Bhootra (SSB) formulations

Sarin et al. [24] have proposed the following O(n3 ) path constraints for the ATSP. Let dij = 1 if and only if vertex i
precedes (not necessarily immediately) vertex j in a tour, and let

dij "xij , i, j = 2, . . . , n, (48)

dij + dj i = 1, i, j = 2, . . . , n, (49)

dij + dj k + dki !2, i, j, k = 2, . . . , n. (50)

Hence, the SSB1 formulation consists of inequalities (2)–(5) and (48)–(50). Note that the auxiliary variables dij
determine a precedence relationship between the vertices and play the same role as the vij variables of Gouveia and
Pires [17]. That is to say, vij = dij .
Sarin et al. [24] have proposed two stronger formulations, namely SSB2 (L1ATSPxy therein) and SSB3 (SL1ATSPxy
therein), which are polynomial size extensions of the SSB1 formulation by, respectively, replacing (50) with one of the
following inequalities

(dij + xj i ) + dj k + dki !2, i, j, k = 2, . . . , n, (51)

3(dij + dj k + dki ) + xj i + xkj + xik !6, i, j, k = 2, . . . , n. (52)

Inequalities (51) have also been discussed by Altınel et al. [7]. To further tighten the LP relaxations of the SSB2 and
SSB3 formulations, Sarin et al. [24] have included the following constraints to both of the SSB2 and SSB3 formulations:

x1j + xj 1 !1, j = 2, . . . , n. (53)

2.11. The Sherali, Sarin and Tsai (SST) formulations

The latest ATSP formulation is due to Sherali et al. [25]. They have started with the following lifted path constraints:

dij "x1i , i, j = 2, . . . , n, (54)

dj i "xi1 , i, j = 2, . . . , n, (55)

−(1 − xik )!dij − dkj !1 − xik , i, j, k = 2, . . . , n (56)

which they have combined with (48) and (49) to obtain the subtour elimination constraints of the SST1 (called ATSP0
therein) formulation.
644 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

Table 1
Classification of the ATSP formulations

Category Formulations Reference Variables Constraints

Exponential sized DFJ Dantzig et al. [9] O(n2 ) O(2n )

GP5, GP6, GP8 Gouveia and Pires [23]
CP Grötschel and Padberg [10]
Miller–Tucker–Zemlin based MTZ Miller et al. [11] O(n2 ) O(n2 )
DL Desrochers and Laporte [12]
SD Sherali and Driscoll [14]
Single commodity flow GG Gavish and Graves [16] O(n2 ) O(n2 )
Two-commodity flow FCG Finke et al. [18] O(n2 ) O(n)
Multi commodity flow WONG Wong [3] O(n3 ) O(n3 )
CLAUS Claus [21]
LANGEVIN Langevin [19]
LOULOU Loulou [20]
GP7 Gouveia and Pires [23]
SST2 Sherali et al. [25]
Time dependent FGG2, FGG3, FGG4 Fox et al. [22] O(n3 ) O(n)
Precedence variable based GP1, GP2, GP3, GP4 Gouveia and Pires [17] O(n2 ) O(n3 )
SSB1, SSB2, SSB3 Sarin et al. [24]
SST1 Sherali et al. [25]

These authors have also applied a first-order RLT to the SST1 formulation and derived the following constraints:

0 !tijk !xik , i, j, k = 2, . . . , n; i, k ̸ = j , (57)

n
!
tijk + xij = dij , i, j = 2, . . . , n, (58)
k=2;k̸ =j

n
!
x1k + tijk = dkj , k, j = 2, . . . , n, (59)
i=2;i̸=j

which are shown to be tighter than the MCF subtour elimination constraints. Sherali et al. [25] have proposed several
augmented formulations which are tighter than the SSB1 and SSB2. For the sake of conciseness, we will only present
here SST2 (called ATSP6 therein) which yields the tightest bounds. The SST2 formulation is defined with inequalities
(2)–(5), (49), (51), (54), (55) and (57)–(59).
The ATSP formulations presented in this section are summarized in Table 1.

3. Known relationships between ATSP formulations

We now summarize several known relationships between the ATSP formulations just described.

3.1. The MTZ formulation

In their analytical comparison of different ATSP formulations Padberg and Sung [4] have shown that the polyhedron
obtained by projecting P(MTZ) into the subspace of xij variables, namely TP(MTZ), is the weak circuit polytope
because of weak circuit inequalities which are a weaker version of the circuit inequalities (8). Hence they have shown
that P(CP) is a proper subset of TP(MTZ) and, the MTZ formulation is weaker than both CP and DFJ formulations.
The weakness of the MTZ formulation was also demonstrated by Langevin et al. [5]. Desrochers and Laporte [12]
have lifted some MTZ constraints, while Sherali and Driscoll [14] have proposed a reformulation of the MTZ constraints
which imply the DL constraints.
 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654 645

Table 2
Relative percent deviations between LP relaxation bounds and optimal values

Instance Optimum SST1 SSB1 SSB3 SSB2 SST2 GP1 GP2 GP3 GP4 CLAUS MTZ FGG2 FGG4 GG SD DL

ftv33 1286 7.574 4.769 4.589 4.426 0.000 4.769 4.646 0.000 0.000 0.000 7.642 75.492 6.959 7.031 4.782 5.351
ftv35 1473 5.974 3.288 3.138 3.123 0.651 3.404 3.235 1.322 1.288 1.064 6.118 77.065 5.530 5.594 3.903 4.039
ftv38 1530 5.752 3.194 3.058 2.913 0.641 3.201 3.094 1.272 1.240 1.024 5.869 76.952 5.838 5.437 3.264 3.454
ftv44 1613 5.456 2.036 1.957 1.922 na 2.118 1.992 1.837 1.837 1.744 5.555 72.229 5.158 5.175 2.433 2.433
ftv47 1776 6.729 2.528 2.451 2.365 na 2.528 2.412 1.896 1.858 1.542 6.767 74.347 6.521 6.544 2.747 2.835
ftv55 1608 10.759 4.740 4.484 4.167 na 4.740 4.571 2.539 2.332 1.493 10.554 74.184 10.288 10.311 5.891 6.049
ftv64 1839 6.417 4.002 3.954 3.800 na 4.001 3.982 3.144 3.144 1.713 6.314 74.792 5.823 5.845 4.003 4.241
ftv70 1950 9.436 4.463 4.379 4.251 na 4.463 4.374 3.536 3.523 2.103 9.260 74.279 8.729 8.751 4.637 4.691
ft70 38 673 1.211 0.700 0.673 0.598 na 1.271 1.252 1.146 1.144 0.053 1.773 37.470 1.085 1.103 0.798 0.878
ft53 6905 14.096 12.548 12.437 12.210 na 12.548 12.438 11.078 10.692 0.000 14.042 72.992 12.376 12.454 11.392 12.934

Average 7.340 4.227 4.112 3.977 0.431 4.304 4.200 2.777 2.706 1.073 7.389 70.980 6.831 6.825 4.385 4.691

3.2. The MCF formulations

We will next present the relationships associated with the MCF formulations which are among the most used ATSP
formulations because of their relative strength. To the best of our knowledge, there exist three results on the equivalence
of MCF formulations with the DFJ formulation. The first result is by Wong [3] who developed the first MCF formulation
and showed that it is equivalent to DFJ. The second result is due to Langevin et al. [5] who have shown the equivalence
of a class of MCF formulations: DFJ, LANGEVIN, LOULOU, WONG and CLAUS. The third result is due to Padberg
and Sung [4] who have derived the projection TP(CLAUS) of polyhedron P(CLAUS) into the subspace of xij variables
and have shown that it is equivalent to P(DFJ).

3.3. The SCF formulation

Langevin et al. [5] have demonstrated that by definition of the flow variables of the SCF and MCF formulations,
SCF is an aggregation of MCF. In fact the LP relaxation bound of the instances from Table 2 for the GG and CLAUS
formulations show that TP(MCF) is a proper subset of TP(SCF).
We now consider the GG formulation. As a consequence of a result by Gouveia [26] for the Capacitated Vehicle
Routing Problem, the polyhedron obtained by projecting P(GG) into the subspace of xij variables is the weak clique
polytope because of weak clique inequalities. These inequalities are weaker versions of inequalities (7). This result
shows that P(DFJ) is a proper subset of TP(GG) and therefore the GG formulation is weaker than the DFJ formulation.
However, GG is stronger than MTZ since TP(GG) is a proper subset of TP(MTZ). This follows from the fact that
weak clique inequalities include also constraints weak circuit inequalities. This result was shown by Wong [3] and by
Padberg and Sung [4].

3.4. The GP1, GP2, GP3 and GP4 formulations

We now summarize the projection results by Gouveia and Pires [17] on the strengths of the GP1, GP2, GP3 and
GP4 formulations. The authors have shown that the projection of the polyhedron P(GP1) into the subspace of original
xij variables, which is the set of feasible solutions in the LP relaxation of GP1, is exactly the circuit polytope P(CP).
Since P(CP) is a proper subset of TP(MTZ) this implies that GP1 is stronger than MTZ.
Gouveia and Pires [17] have also provided the projections TP(GP2) and TP(GP3) of the LP relaxations of GP2 and
GP3 formulations into the subspace of the original xij variables. Both TP(GP2) and TP(GP3) are proper subsets of the
circuit polytope P(CP), which is equivalent to TP(GP1). Hence this implies that the GP2 and GP3 formulations are
stronger that the GP1 formulation. Unfortunately, the GP2 and GP3 formulations are dominated by the DFJ formulation
because P(DFJ) is a proper subset of both TP(GP2) and TP(GP3) [17].
The authors have shown that the GP2 and GP3 formulations are incomparable. They have also stated that the GP4
formulation is stronger than the GP2 and GP3 formulations since it includes the constraints of both GP2 and GP3
646 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

formulations. Gouveia and Pires [17] have shown that a dominance relation between the GP4 and CLAUS formulations
does not exist. Hence, the GP4 and DFJ formulations are incomparable. Finally, they have proved that the GG and GP1
formulations are not comparable [17].

3.5. The GP5, GP6, GP7 and GP8 formulations

Gouveia and Pires [23] have analyzed the strength of the GP5, GP6, GP7 and GP8 formulations. They have proved
that the CLAUS formulation is stronger than the GP5 formulation, which is in its turn stronger than the GP3 formulation.
The equivalence of the GP6 and DFJ formulations has later been established by Myung [27]. Recall that Padberg and
Sung [4] have shown that TP(CLAUS), the projection of the polyhedron P(CLAUS) into the subspace of xij variables,
is the polyhedron P(DFJ). As a consequence TP(GP7), the projection of the polyhedron P(GP7) into the subspace of
xij variables, is contained in the set described by the inequalities (7), and therefore TP(GP7) is a proper subset of
TP(CLAUS) and also P(DFJ). This implies that GP7 is stronger than CLAUS. Finally, Gouveia and Pires [23] have
stated that the GP8 formulation is stronger than both the GP6 and GP7 formulations, since the constraints of GP8 imply
the constraints of the GP6 and GP7 formulations.

3.6. The time-dependent TSP formulation

The time-dependent TSP formulations were analyzed and compared by Gouveia and Voss [6]. In this overview
we have only considered the time-dependent formulation by Fox et al. [22]. Gouveia and Voss [6] have shown that
P(FGG4) is a proper subset of P(FGG3). Furthermore, Gouveia and Voss [6] have proposed an SCF formulation based
on the subtour elimination constraints of the GG formulation. Their SCF formulation was slightly stronger than the
GG formulation. They have proved that their SCF formulation is equivalent to the FGG3 formulation which implies
that the FGG3 formulation is stronger than the GG formulation. Another relationship regarding the FGG formulation
is due to Padberg and Sung [4] who have derived the projection of FGG2 formulation into the subspace of xij and then
shown that the FGG2 formulation is weaker than the DFJ formulation.

3.7. The SSB and SST formulations

Sarin et al. [24] have proposed three formulations: SSB1, SSB2 and SSB3. Two of them, SSB2 and SSB3, contain
lifted inequalities (51) and (52), respectively, which are liftings of constraints (50). Therefore, both the SSB2 and SSB3
formulations are stronger than the SSB1 formulation.
Sherali et al. [25] have observed that the LP relaxation of the SST1 formulation is as strong as the GP1 formu-
lation since the left-hand side of (56) reduces to that of constraint (35). On the other hand it is also possible to
show that the right-hand side of (56) reduces to dij + dj k "dik "xik which is implied by (48)–(49). Therefore SSB1
is as strong as SST1. By considering the instances from Table 2 we can say that TP(SSB1) is a proper subset of
TP(SST1) which is in turn a proper subset of TP(GP1). Finally, Sherali et al. [25] have shown that SST2 is stronger
than SSB2 since all the inequalities of SSB2 formulation are implied by the constraints of SST2 formulation.
Fig. 1 illustrates the relative strengths of the LP relaxations of the ATSP formulations we have presented.

4. New relationships between ATSP formulations

We now develop new relationships between some of the formulations presented in Section 2. These results do not
close the discussion on the comparison of ATSP formulations because there are still some missing relationships.

4.1. Time-dependent TSP formulation

We start with an observation on the time-dependent TSP formulation of Fox et al. [22].

Proposition 1. The FGG3 formulation is stronger than the FGG2 formulation.

 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654 647

A B A and B are incomparable

A B B is better than A

A B A and B are equivalent

GP8 SST2
[7]
[7]
[7] GP7

[1] [2] [8]

DFJ WONG CLAUS GP6 [11]
[7]
[3] GP5 [6]
[7]
SD GP4 SSB2
[3]
[1] [7] [6] [6]
[10] SSB3
[6]
[9] GP3 GP2 [10]
FGG4 SSB1
[5] [6] [6] [11]
[5]
FGG2 FGG3 GG DL GP1 CP SST1
[6]
[1] [6]
[4]
[11]
[1] Wong (1980) MTZ
[2] Langevin et al (1990)
[6] Gouveia and Pires (1999) [9] Sherali and Driscoll (1995)
[3] Padberg and Sung (1991)
[7] Gouveia and Pires (2001) [10] Sarin et al. (2005)
[4] Desrochers and Laporte (1991)
[8] Myung (2001) [11] Sherali et al. (2006)
[5] Gouveia and Voss (1995)

Fig. 1. Known relationships between 24 ATSP formulations.

Proof. Since the FGG2 formulation contains aggregated assignment constraint (34), which can be obtained by summing
up constraints (29)–(31), every solution feasible for the LP relaxation of the FGG3 is also feasible for the LP relaxation
of the FGG2. Indeed all the instances from Table 2 show that P(FGG3) is a proper subset of P(FGG2). #

4.2. The GG, SD and DL formulations

We now compare SD with GG and show that it is not possible to compare GG with DL.

Proposition 2. The SD formulation is stronger than the GG formulation.

Proof. Consider constraints (13)–(15) of the SD formulation and observe that the right-hand sides of (13) and (14) are
in fact equal. As a result we can write
n
! n
!
yij + (n − 1)xi1 = yj i + 1, i = 2, . . . , n. (60)
j =2 j =2

Then by adding this set of inequalities for S ⊆ V = {1, . . . , n} we obtain

! ! !
yij + (n − 1) xi1 = yij + |S|, (61)
i∈S\{1} i∈S\{1} i∈S\{1}
j ∈S\{1} j ∈S\{1}
648 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

where S = V \S. On the other hand the aggregation of constraints (15) over S gives
! !
(n − 2) xij " yij (62)
i∈S\{1} i∈S\{1}
j ∈S\{1} j ∈S\{1}

and
! !
yij " xij . (63)
i∈S\{1} i∈S\{1}
j ∈S\{1} j ∈S\{1}

Hence equality (61) and inequalities (62) and (63) imply

! ! !
(n − 2) xij + (n − 1) xi1 " xij + |S|, (64)
i∈S\{1} i∈S\{1} i∈S\{1}
j ∈S\{1} j ∈S\{1}

which results in
! !
(n − 2)D(S\{1}) + (n − 1) xi1 "(n − 1) xij + |S| (65)
i∈S\{1} i∈S\{1}
j ∈S\{1}
" " "
after adding (n − 2) i∈S\{1} xij to both sides of (64). Here D(S\{1}) = i∈S\{1} xij + i∈S\{1} xij is the number
j ∈S\{1} j ∈S\{1} j ∈S\{1}
of arcs having exactly one endpoint (head or tail but not both) in the subset S\{1}. Inequality (65) can be equivalently
rewritten as
!
(n − 2)D(S\{1}) + (n − 1) xij "(n − 1)D(S\{1}) + |S| (66)
i∈S\{1}
j ∈S\{1}
"
after adding (n − 1) i∈S\{1} xij to its right- and left-hand sides, from which
j ∈S\{1}
!
(n − 1) xij "D(S\{1}) + |S| (67)
i∈S\{1}
j ∈S\{1}

follows. Since
! !
xij + xij = |S| (68)
i,j ∈S i∈S
j ∈S

for all subsets S of V, we can write

! !
xij + xij !|S|, S⊆V (69)
i,j ∈S\{1} i∈S\{1}
j ∈S

or equivalently
! !
xij !|S| − xij , S ⊆ V. (70)
i,j ∈S\{1} i∈S\{1}
j ∈S

Now, using (70) we can rewrite (67) as

⎛ ⎞
!
(n − 1) ⎝|S| − xij ⎠ "D(S\{1}) + |S| (71)
i,j ∈S\{1}
 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654 649

the weak clique inequalities

|S| !
|S| − " xij , S ⊆ {2, . . . , n}, 2 !|S|!n − 1 (72)
n−1
i,j ∈S

since D(S\{1}) is non-negative. A direct consequence of this discussion is that the projection of a polyhedron that
includes P(SD) is included in TP(GG) and consequently TP(SD) is in TP(GG). The SD formulation is as strong as
GG. In addition it is always possible to find an ATSP instance for which the LP bound of the SD formulation is strictly
larger than that of the GG formulation on the same instance. One such example is ftv33 (Table 2). Therefore TP(SD)
is a proper subset of TP(GG) and SD is stronger than GG. #

Proposition 3. The DL and GG formulations are incomparable.

Proof. Consider instances ft53 and ft70 in Table 2. The percent gaps from the optimum of the LP bounds corre-
sponding to DL and GG formulations are, respectively, 12.934 and 12.454 for ft53. However, they are equal to 0.878
and 1.103 for ft70. #

4.3. The GP1, GP2, GP3, GP4, DL and SD formulations

Another interesting relationship exists between the GP1 and DL formulations. The fact that (11) is one of the subtour
elimination constraints of the DL formulation, and that it is derived from constraints (36), (37), (39) and the linear
transformation (38) may lead one to think that the GP1 formulation is stronger than the DL formulation. However, this
is not the case.

Proposition 4. The GP1, GP2, GP3 and GP4 formulations are not comparable to the DL formulation.

Proof. Observe that for ftv70 (Table 2) the LP bound of GP1 is strictly larger than the LP bound of DL. However,
for ft70 the LP bound of DL is strictly larger than that of GP4. #

A similar relationship exists between GP1, GP2, GP3, GP4 and SD.

Proposition 5. The GP1, GP2, GP3 and GP4 formulations are not comparable to the SD formulation.

Proof. For instance ftv55, the LP bound of GP1 is strictly larger than that of SD. However, for ft70, the LP bound
of SD is strictly larger than that of GP4. #

4.4. The SSB formulations

We now comment on the relative strengths of the SSB1, SSB2 and SSB3 formulations. Our first remark is on the
relation between the SSB1 and SSB2. Note that SSB2 is potentially stronger than SSB1 since inequalities (51) are lifted
versions of the inequalities (50). Our second observation is related with the inequalities (48) of SSB2 which are in fact
redundant. To see this, define XD = {(x, d) : xij !dij , dij + dj i = 1, i, j = 2, . . . , n} and also define the polyhedron
TP(XD) = {x ∈ Rn(n−1) : xij + xj i !1, i, j = 2, . . . , n}.

Proposition 6. The projection of XD into the subspace of the xij variables is TP(XD).

Proof. Let (x, d) be a vector in XD. Summing up the inequalities x ij !d ij and x j i !d j i and using the identities
d ij + d j i = 1 we obtain x ij + x j i !1. Hence, XD ⊆ TP(XD). Now select x from the polyhedron TP(XD). Then
xij !1 − xj i follows. In addition setting dij = xij and dj i = 1 − xij "xj i provides a feasible assignment, implying that
there exists a vector d for every x ∈ TP(XD) such that (x, d) ∈ XD. Therefore TP(XD) ⊆ XD. #

Now, adding inequalities dij + dj k + dki + xj i !2 and dj i + dik + dkj + xij !2 implies xij + xj i !1, namely TP(XD),
since dij + dj i = 1, dkj + dj k = 1 and dki + dik = 1. Hence, inequalities (51) together with inequalities (49) imply
inequalities (48). We have thus proved the following proposition.
650 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

Proposition 7. Inequalities (48) are redundant in SSB2.

The next two results show that the SSB2 dominates both of the GP2 and SSB3 formulations.

Proposition 8. TP(SSB2) is a proper subset of TP(SBB3).

Proof. Adding inequalities (51) for (i, j ), (j, k) and (k, i), i.e.,

dij + dj k + dki + xj i !2,

dj k + dki + dij + xkj !2,

dki + dij + dj k + xik !2

yields

3(dki + dij + dj k ) + xj i + xkj + xik !6,

which is exactly (52). This implies that SSB2 is also potentially stronger than SSB3. Furthermore, Table 2 contains
instances, such as ftv47, for which the LP bound SSB2 is strictly larger than that of SBB3. #

Proposition 9. TP(SSB2) is a proper subset of TP(GP2).

Proof. Consider a circuit DC = (VC , C). Adding the subtour elimination constraints

xij + dki + dij !dkj + 1, (i, j ) ∈ C with i, j, k = 2, . . . , n; k ∈

/C

and

xj i + dpj + dj i !dpi + 1, (i, j ) ∈ C with i, j, p = 2, . . . , n; i, j ̸ = p; p ∈ C,

and using circularity constraints dij + dj i = 1 we obtain inequalities

! !
xij + xj i !|C| − 1, p ∈ VC , VC ⊆ {2, . . . , n}. (73)
(i,j )∈C (i,j )∈C
i,j ̸=p

On the other hand, adding the following constraints and using dij + dj i = 1,

xij + dki + dij !dkj + 1 for i, j, k = 2, . . . , n,

xj i + dkj + dj i !dki + 1 for i, j, k = 2, . . . , n

we obtain

xij + xj i !1 for i, j = 2, . . . , n. (74)

Constraints (73) and (74) are known to be the inequalities describing the projection of the subtour elimination constraints
of GP2 formulation into the subspace of xij variables [17]. Therefore TP(SSB2) is a subset of TP(GP2). Moreover,
Table 2 contains ATSP instances, such as ftv33, for which the LP bound of SSB2 is strictly larger than that of GP2. #

Our final remark is on the incomparability of the SSB2 and GP4 formulations.

Proposition 10. The SSB2 and GP4 formulations are incomparable.

Proof. To see this it suffices to consider instances ft70 and ft53. For one of them GP4 give tighter LP bound but
for the other SSB2 yields a tighter LP bound. #
 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654 651

4.5. The SST formulations

We will now prove that SST2 is the strongest formulation proposed by Sherali et al. [25].

Proposition 11. The constraints (57)–(59) imply the lifted inequalities (40).

Proof. By definition of the flow variables fkij used in the constraints (41)–(44) [23] and by the definition of the flow
j
variables tik used in the tightened MCF constraints [25], we have the identity
j
tik = fkij . (75)
j
Recall that fkij is equal to one if and only if arc (i, j ) is on the path from node k to node 1 and tik is equal to one if the
commodity
" from 1 to k flows on arc (i, j ). On the other hand, like the transformation (45) by Gouveia and Pires [23],
namely nj=1 fkij = dki for i, k = 2, . . . , n, we have the following transformation:
n
! j
tik = dki , k, j = 2, . . . , n. (76)
j =1

Gouveia and Pires [23] have shown that it is possible to obtain constraints (40) from constraints (43), (44), and a weaker
version of (41). Following similar steps and by using (75) and (76), we obtain
xij + xik + xkj + dki !1 + dkj , i, j, k = 2, . . . , n, (77)
which are exactly the lifted inequalities (40) proposed by Gouveia and Pires [17] since dij = vij . #

Therefore, Proposition 11 states that the strongest formulation proposed by Sherali et al. [25] is SST2 since none of
their formulations contains any constraints stronger than (57)–(59).

Proposition 12. The SST2 formulation is stronger than the SSB2 formulation.

Proof. It is clear that SST2 is stronger than SSB2 since the former includes all the constraints of the latter, as well
as some additional non-redundant constraints. Hence, the LP relaxation of SST2 is tighter than that of SSB2. To see
that TP(SST2) is a proper subset of TP(SSB2) it suffices to consider the instance ftv33 of Table 2, for which the LP
bound of SST2 is strictly larger than that of SSB2. #

Proposition 13. The SST2 formulation is stronger than the GP7 formulation.

Proof. The thesis follows from the results by Sherali et al. [25] who have observed that constraints (57)–(59) are
stronger than the MCF constraints (41)–(44), and constraints (51) are liftings of constraints (35). To show that SST2 is
stronger than GP7 it suffices to provide a feasible (xij , vij , zkij ) vector of P(GP7) such that there does not exist a vector
(dj i , zkij ) for which (xij , dj i , zkij ) is included in P(SST2). Consider the following solution for a five-vertex instance
x13 = 23 , x15 = 13 , x21 = 23 , x25 = 13 , x34 = 23 , x35 = 13 , x41 = 13 , x42 = 23 , x52 = 13 , x53 = 13 , x54 = 13 . This solution
is feasible for GP7, but infeasible for SST2. Hence there exists at least one instance for which TP(SST2) is a proper
subset of TP(GP7). #

5. Computational experiments

To complete our comparative study, we report some results on the empirical quality of the LP bounds obtained
with some of the formulations presented in this paper. We have tested these formulations on standard ATSP in-
stances obtained from TSPLIB Reinelt [28]. In Table 2 we present relative deviations from the optimal tour lengths,
computed as
' ∗ (
z − z∗
100 × I P ∗ LP ,
zI P
652 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

where zI∗P is the length of an optimal tour and zLP ∗ is the lower bound obtained by solving the LP relaxation of the

models. All computations were performed on a Pentium IV PC with a 3 GHz CPU.

As can be observed from Table 2, the SST2 and CLAUS formulations provide the best bounds. This is not surprising
since SST2 is theoretically stronger than CLAUS, which is theoretically as strong as DFJ. The SSB2, SSB3, GP3 and
GP4 formulations are the second best, while the performance of FGG4 is the worst. Table 3 provides the CPU times
required for the computation of the lower bounds with the barrier solver of CPLEX 9.0. We use the barrier solver since
the instances are quite large and the use of interior point methods can decrease the solution time considerably. However,
barrier requires larger memory space which can cause difficulties for large linear programs as it is the case with the LP
relaxation of SST2. This explains why we could not solve the LP relaxation of ftv44 and of larger instances.

Table 3
CPU times (in seconds) with barrier option

Instance SST1 SSB1 SSB3 SSB2 SST2 GP1 GP2 GP3 GP4 CLAUS MTZ FGG2 FGG4 GG SD DL

ftv33 21.9 5.1 7.3 7.8 668.0 15.7 17.3 57.4 29.9 17.2 2.3 0.9 2.4 0.2 0.8 2.8
ftv35 27.6 7.9 10.0 10.8 1357.0 21.9 34.2 71.0 71.9 18.9 3.9 1.0 8.9 0.2 0.9 2.6
ftv38 44.9 13.7 15.8 22.5 1741.0 89.9 66.0 74.9 91.6 29.3 4.7 1.3 4.6 0.2 1.2 5.8
ftv44 103.7 22.7 27.9 34.0 na 131.0 101.9 218.0 148.0 56.6 10.0 2.9 7.2 0.3 1.6 12.4
ftv47 140.4 31.9 37.3 42.9 na 171.8 199.0 194.7 184.8 232.0 20.3 2.9 7.7 0.3 2.1 20.6
ftv55 314.2 87.7 92.4 8.3 na 242.0 263.0 283.8 507.0 346.0 36.4 5.9 16.9 0.4 3.5 40.2
ftv64 876.0 201.2 234.5 254.1 na 823.7 910.1 779.7 2806.7 602.1 52.8 8.9 29.1 0.5 5.7 75.0
ftv70 1384.9 310.3 408.4 493.0 na 1030.4 1193.6 1013.9 3707.3 917.6 112.2 12.8 41.7 0.6 14.5 152.9
ft70 671.0 410.5 433.7 447.0 na 780.7 1337.1 1144.8 2986.2 832.8 110.4 26.9 32.4 0.6 7.5 149.9
ft53 902.0 193.6 194.0 167.0 na 186.0 200.0 349.2 419.0 382.0 26.9 5.5 12.9 0.3 3.2 33.1

Average 448.7 128.5 146.1 148.7 1255.3 349.32 432.22 418.75 1095.25 343.46 37.99 6.9 16.4 0.4 4.1 49.5

A B A and B are incomparable

A B B is better than A

A B A and B are equivalent

GP8 SST2
Dashed lines denote new relationships

GP7

DFJ WONG CLAUS GP6

GP5

SD GP4 SSB2

SSB3
GP3 GP2
FGG4 SSB1

FGG2 FGG3 GG DL GP1 CP SST1

MTZ

Fig. 2. Relative strength of the 24 ATSP formulations.

 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654 653

6. Conclusions

This comparative review and analysis clearly illustrates the wealth of available models for the ATSP. We have
summarized the known relationships between these models and we have proved some new ones. Our study is sup-
ported by computational results on a subset of the TSPLIB instances. Fig. 2 depicts the relationships between 24
ATSP formulations, but also shows that some gaps exist. Clearly, there is scope for further research in this fascinating
area.

Acknowledgments

This research was partially supported by Boğaziçi University Research Fund Grant No. 04HA301D. Thanks are due
to three anonymous referees for their valuable comments which have helped improve the paper.

References

[1] Papadimitriou CH. The Euclidean traveling salesman problem is NP-complete. Theoretical Computer Science 1977;4:237–44.
[2] Balas E, Pulleyblank WR. The perfectly matchable subgraph polytope of a bipartite graph. Networks 1983;13:486–516.
[3] Wong RT. Integer programming formulations of the traveling salesman problem. Proceedings of the IEEE international conference of circuits
and computers, 1980. p. 149–52.
[4] Padberg M, Sung T. An analytical comparison of different formulations of the travelling salesman problem. Mathematical Programming
1991;52:315–57.
[5] Langevin A, Soumis F, Desrosiers J. Classification of travelling salesman problem formulations. Operations Research Letters 1990;9:127–32.
[6] Gouveia L, Voss S. A classification of formulations for the (time-dependent) travelling salesman problem. European Journal of Operational
Research 1995;83:69–82.
[7] Altınel İK, Öncan T, Ünal AT. New extended inequalities with precedence variables for the asymmetric travelling salesman problem. Research
Paper Series No. FBE-IE-02/2005-02, Department of Industrial Engineering, Boğaziçi University, İstanbul; 2005.
[8] Orman AJ, Williams HP. A survey of different integer programming formulations of the travelling salesman problem. In: Kontoghiorghes EJ,
Gatu C, editors. Optimisation, economics and financial analysis. Advances in computational management science, vol. 9. Heidelberg: Springer;
2007. p. 93–106.
[9] Dantzig GB, Fulkerson DR, Johnson SM. Solutions of a large-scale traveling-salesman problem. Operations Research 1954;2:363–410.
[10] Grötschel M, Padberg M. Polyhedral theory. In: Lawler EL, Lenstra JK, Rinnooy Kan AHG, Shymoys DB, editors. The traveling salesman
problem: a guided tour of combinatorial optimization. New York: Wiley; 1985.
[11] Miller CE, Tucker AW, Zemlin RA. Integer programming formulations and travelling salesman problems. Journal of the Association for
Computing Machinery 1960;7:326–9.
[12] Desrochers M, Laporte G. Improvements and extensions to the Miller–Tucker–Zemlin subtour elimination constraints. Operations Research
Letters 1991;10:27–36.
[13] Driscoll PJ. A new hierarchy of relaxation for 0–1 mixed integer problems with application to some specially structured problems. Doctoral
dissertation, Department of Industrial and Systems Engineering, Virginia Polytechnic Institute, Blacksburg, VA; 1995.
[14] Sherali HD, Driscoll PJ. On tightening the relaxations of Miller–Tucker–Zemlin formulations for asymmetric traveling salesman problems.
Operations Research 2002;50:656–69.
[15] Sherali HD, Adam WP. A reformulation–linearization technique for solving discrete and continuous nonconvex problems. Dordrecht, The
Netherlands: Kluwer; 1999.
[16] Gavish B, Graves SC. The travelling salesman problem and related problems. Working Paper GR-078-78, Operations Research Center,
Massachusetts Institute of Technology, 1978.
[17] Gouveia L, Pires JM. The asymmetric travelling salesman problem and a reformulation of the Miller–Tucker–Zemlin constraints. European
Journal of Operational Research 1999;112:134–46.
[18] Finke G, Claus A, Gunn E. A two commodity network flow approach to the travelling salesman problem. Congressus Numerantium 1984;1:
167–78.
[19] Langevin A. Planification des tournées de véhicules, PhD dissertation, École Polytechnique de Montréal, 1988.
[20] Loulou RJ. On multicommodity flow formulations for the TSP. Working paper, McGill University, Montréal; 1988.
[21] Claus A. A new formulation for the travelling salesman problem. SIAM Journal on Algebraic and Discrete Methods 1984;5:21–5.
[22] Fox KR, Gavish B, Graves SC. An n-constraint formulation of the (time-dependent) traveling salesman problem. Operations Research
1980;28:1018–21.
[23] Gouveia L, Pires JM. The asymmetric travelling salesman problem: on generalizations of disaggregated Miller–Tucker–Zemlin constraints.
Discrete Applied Mathematics 2001;112:129–45.
[24] Sarin SC, Sherali HD, Bhootra A. New tighter polynomial length formulations for the asymmetric travelling salesman problem with and without
precedence constraints. Operations Research Letters 2005;33:62–70.
654 T. Öncan et al. / Computers & Operations Research 36 (2009) 637 – 654

[25] Sherali HD, Sarin SC, Tsai P-F. A class of lifted path and flow-based formulations for the asymmetric travelling salesman problem with and
without precedence constraints. Discrete Optimization 2006;3:20–32.
[26] Gouveia L. A result on projection for the vehicle routing problem. European Journal of Operational Research 1995;85:610–24.
[27] Myung Y-S. Projections of extended formulations with precedence variables for the asymmetric traveling salesman problem. International
Journal of Management: A Journal of the Korean OR and MS Society 2001;7:1–12.
[28] Reinelt G. TSPLIB—A traveling salesman problem library. ORSA Journal on Computing 1991;3:376–84 (Downloadable from website
⟨http://www.iwr.uni-heidelberg.de/groups/comopt/software/TSPLIB95/⟩).