1 ON THE EXISTENCE OF PRICING STRATEGIES IN THE HETEROGENEOUS

2 SINGLE BOTTLENECK MODEL

3
4 Kien Doan
5 5107 Jonsson Engineering Center
6 Department of Civil and Environmental Engineering
7 Rensselaer Polytechnic Institute
8 Troy, NY 12180
9 [email protected]
10 (518) 276-8308
11
12 Satish Ukkusuri
13 Associate Professor
14 Purdue University
15 Office CIVL G175F
16 School of Civil Engineering
17 550 Stadium Mall Drive
18 West Lafayette, IN 47907-2051
19 [email protected]
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37 Number of words = 7184 + 8 figures + 0 tables = 9184
38
39 Submitted for publication in Journal of Transportation Research Record
40
41
 1
Kien Doan, Satish Ukkusuri

1 ABSTRACT
2 This paper formulates the dynamic system optimal conditions for the single bottleneck model with
3 heterogeneous commuters. This formulation contributes to the literature by solving the system optimal
4 problem with heterogeneous network users. The existence and uniqueness of the problem are shown and
5 discussed. Apart from a rigorous system optimal formulation, this paper develops feasible toll strategies
6 for the single bottleneck problem. The necessary and sufficient conditions for the existence of tolls
7 strategies within the single bottleneck are also developed. Extensive computational results are conducted
8 to demonstrate the results in this paper.
9
10
 2
Kien Doan, Satish Ukkusuri

1 1. INTRODUCTION AND LITERATURE REVIEW

2 The single bottleneck (SB) problem is identified as the morning rush hour congestion in which a fixed
3 number of commuters pass a route segment with a fixed capacity known as the single bottleneck. The
4 demand to pass this segment is larger than the capacity. In addition, the commuters have their own
5 preferred arrival times and cost functions. They pay a schedule delay cost if they arrive earlier or later
6 than the preferred arrival time. By choosing a specific departure time, the commuters can maximize their
7 utility when passing the bottleneck. Without toll, their time determination is based on the trade-off
8 between travel time and schedule delay that they experience. The seminal work of Vickrey (1) provided a
9 foundation for understanding the SB model. Since then, there have been many studies that explore this
10 model in various directions. Before discussing the contributions of this work, a brief literature review is
11 presented as follows:
12 1.1. Single bottleneck model with homogeneous commuters
13 In the homogeneous SB model, all commuters have the same characteristics, such as the preferred arrival
14 time and the schedule delay cost functions. Solving the equilibrium of the homogeneous commuter model
15 is straightforward. Arnott et al (2) considered the continuous model with homogeneous commuters. The
16 user equilibrium (UE) conditions and some congestion pricing conditions were taken into account. In
17 addition, this study examined the elastic demand and relative toll regimes. Daganzo (3) used the
18 bottleneck model with variable arrival curves. This study was concerned with the unique order of arrivals,
19 the smoothness of the arrival curve and the uniqueness of the equilibrium. Ramadurai and Ukkusuri (4)
20 show theoretically that Wardrop’s user equilibrium does not exist based on the assumption of discrete
21 players. Mun (5) analyzed the bottleneck model with traffic properties: speed, density, length of queue. A
22 model for homogeneous commuters was developed along with an optimal peak-load toll scheme. This
23 study used a more realistic traffic flow model but the analysis was limited to the homogeneous cases.
24 1.2. Single bottleneck model with heterogeneous commuters
25 Arnott et al (2) is one of the first papers to conclude that the heterogeneous SB model is not straight-
26 forward. There are few studies that explore solution procedures for the heterogeneous SB model under
27 specific conditions. Newell (6) used a similar model as in Daganzo (3), but considered the concept of
28 variable arrival time. Different travelers have different cost functions and different preferred arrival times.
29 The solution is obtained by a graphic technique. Arnott et al (7) provided the equilibrium conditions for
30 particular cases in the SB model: commuters with the same preferred arrival time and the same ratio of
31 late arrival schedule delay to early arrival schedule delay, commuters with same preferred arrival time,
32 travel cost, early arrival cost but different late arrival cost, and commuters with same cost functions but
33 different preferred arrival time. Lindsey (8), using the continuous time model, analytically showed that
34 the equilibrium in the SB model with heterogeneous commuters always exists. Recently, Ramadurai et al.
35 (9) (10) developed a linear complementarity (LCP) formulation for the SB model with homogeneity and
36 heterogeneity assumptions. The advantages of this formulation are: 1) It’s the first known analytical
37 formulation for the SB model accounting for the general form of heterogeneity in user classes, 2) The
38 existence and uniqueness are theoretically proven for both homogeneous and heterogeneous model, 3)
39 The formulation allows for efficient computation of equilibrium solutions; and 4) It allows further
40 extensions of the SB model especially for heterogeneous commuters. These papers above consider the
 3
Kien Doan, Satish Ukkusuri

1 user equilibrium conditions. However, they did not take into consideration the system optimal (SO)
2 assignment.
3 1.3. System optimal and toll scheme
4 In the first paper on SB problem, Vickrey (1) showed that the pricing structure could eliminate queuing
5 and lead to efficient use of SB in the morning rush hour. Many other papers considered tolls as a critical
6 issue to efficiently use the SB. Arnott et al (2) proposed a set of pricing regimes such as uniform toll,
7 coarse toll and fine toll to approximately transform the assignment to the system optimal conditions. This
8 study confirmed that the exact system optimal can be obtained with the fine toll. Moreover, the coarse toll
9 is more efficient than the uniform toll but less than fine toll. Arnott et al (11) considered the system
10 optimal and welfare for three particular cases: 1) individuals with different travel time penalty (α) and
11 early penalty (β), same ratio of late penalty to early penalty (γ/β), 2) individuals differing with respect to
12 γ, and 3) individuals with different preferred arrival time (t*). Laih (12) considered the multiple step tolls
13 for general SB model with homogeneous commuters. This study showed that at most n/(n+1) of the total
14 travel time can be eliminated with the n-step toll. Daganzo and Garcia (13) used time-dependent toll in
15 which some of commuters are exempted from paying. The paying commuters can avoid toll if they depart
16 outside of the time window. This paper stated that up to 25% of total user cost can be eliminated. Newell
17 (6) discussed time-dependent tolls for commuter heterogeneity model. The toll cT(t) is determined only by
18 the commuters who value schedule delay least or most relative to money. This study, however, did not
19 draw the concrete result of toll strategies. Finally, since Ramadurai et al (9) proposed the LCP
20 formulation to analytically solve the user equilibrium problem, it is promising to extend this model to test
21 the possibilities of feasible toll strategies in this paper.
22 1.4. Motivation and contribution of this paper
23 Within this context, the research in this paper is motivated by the lack of a rigorous system optimal
24 formulation for the heterogeneous single bottleneck model. In addition, this paper introduces a new
25 system state for traffic assignment which is at least as good as the system optimum and better than the
26 user equilibrium (no incentive for commuters to shift departure time) From now on, this state is called the
27 stable system state (SSS).
28 First, the no-toll system optimal is formulated for the SB model with heterogeneous commuters.
29 This paper fills in the gap in the literature by an analytical formulation with general parameter
30 assumptions. Then, this paper obtains the insight into the existence of solution, the number of solutions
31 and the parameters which affect the solutions. Some comparisons between SO formulation and UE
32 formulation are also described. The second contribution of this paper is finding the necessary conditions
33 for feasible toll schemes and the sufficient conditions for input parameters such that the SSS is obtained.
34 2. SYSTEM OPTIMAL OF SINGLE BOTTLENECK MODEL WITH HETEROGE-
35 NEOUS COMMUTERS
36 In transportation engineering, especially transportation network modeling, it is broadly known that a
37 system optimal is the condition in which the total social cost is minimized. In static traffic assignment
38 models, Sheffi (14) showed that “the only case in which the SO flow pattern can be used to represent
39 equilibrium are those special cases in which the SO solution is identical to the UE solution”. On the other
40 hand, the SO is realistic only if it was working in equilibrium condition. In large scale dynamic traffic
 4
Kien Doan, Satish Ukkusuri

1 assignment (Friesz et al., (15)), the system optimal of a network is formulated such that the total cost of
2 all links over time is minimized. In the following section, the paper will propose the formulation to solve
3 the system optimal problem for the SB model with heterogeneous commuters.
4 2.1. Formulation to find the no-toll SO in SB model with heterogeneous commuters.
5 Applying the definition of system optimal for the SB model with heterogeneous commuters, the SO
6 formulation is suggested as follows:
⎛ tg ⎞
*
G T
Minimize ∑ ⎜ ∑ rt , g β g (t g − t ) + ∑ rt , g γ g (t − t g ) ⎟
* *
7 (1)
⎜ ⎟
⎝
g =1 t =1 t =t *g +1 ⎠
G
8 Subject to ∑r
g =1
t,g ≤ s ∀t = 0..T (2)

T
9 ∑r
t =1
t,g = N g ∀g = 1..G (3)

10 rt , g ≥ 0 ∀t = 0..T , ∀g = 1..G (4)

11 rt , g : departure rate at time interval t of group g

12 β g : schedule delay penalty of group g if it arrives earlier than the preferred arrival time
13 γ g : schedule delay penalty of group g if it arrives later than the preferred arrival time
14 N g : total number of departures of group g
15 t : index for time interval, t = 0..T , T is sufficiently large.
16 tg* : preferred arrival time of group g
17 s : capacity of single bottleneck
18 With, g = 1..G . T is sufficiently large.

19 Parameter: β g , γ g , t g* , N g , s >0.

20 Different groups can have same β or same γ as long as their travel time penalties α are different. The
21 result from this formulation is the assignment in which the total system cost is minimized, thus system
22 optimal. Next, this paper will discuss the insight into the travel time the solutions properties from this
23 formulation.
24 2.2. Travel time in SO
25 In the SB model with homogeneous commuters, it is straight forward to see that there is no queue. Thus,
26 there is no travel time in the system optimal solution. This SO is derived from the UE by charging time-
27 varying toll. All the input parameters such as cost functions α, β, γ and preferred arrival time t* are the
28 same for these two states. However, in the SB model with heterogeneous commuters, it is impossible to
29 derive the SO from the UE by toll charge because the SO and UE are different in term of penalty
30 coefficient inputs. While the no-toll UE is determined by the ratio of β/α and γ/α, the SO depends on the
31 absolute value of β and γ (Arnott et al (11)). It has been hypothesized in this study that there is no travel
32 time parameter α in the SO assignment. In other words, in the SB system with heterogeneous commuters,
 5
Kien Doan, Satish Ukkusuri

1 if queue exists, the departure pattern will be different from the arrival pattern. Moreover, the arrival
2 pattern is always a straight line in which the arrival rate is equal to the capacity. Then the total system
3 cost is equal to the travel time cost plus the schedule delay cost determined by the arrival rate pattern.
4 Thus, the total system cost is always greater than or equal to the SO assignment.
5 2.3. Insight into the unique solution and the multiple global solutions
6 First, this paper shows that the formulation (1) to (4) always has solution. A linear system has a solution if
7 and only if it has at least one feasible point and the objective function is bounded. In the formulation (1)
8 to (4), constraints (2), (3), and (4) create a feasible region since variable rt,g is bounded by zero and s and
T
9 the condition ∑r
t =1
t,g = N g ∀g = 1..G is always satisfied with the assumption that T is sufficiently large.

10 Furthermore, the objective function (1) has lower bound equal to zero. Therefore the formulation (1) to
11 (4) always has solution.
12 Second, this paper intends to explore the uniqueness of the solutions since this has important
13 implications to developing the pricing strategies, which will be discussed in section 3 of this paper. This
14 paper shows that there are two approaches to define the number of solutions. In the first one, the number
15 of solutions depends on the input parameters such as preferred arrival time tg* and cost function β, γ.
16 Different values of cost functions will be analyzed in three scenarios. In the second one, the reduced cost
17 patterns, which are calculated from the optimization solvers such as CPLEX, can demonstrate the number
18 of solutions too.
19 This paper illustrates the SO of the trivial case, in which the problem of heterogeneous commuters
20 becomes multiple problems of homogeneous commuters. This trivial case happens when the preferred
21 arrival time tg* of each group is far away from that of the others such that its own departure times do not
22 affect the departure time of the others. In other words, the SO formulation will assign each group to
23 depart at its own departure time such that each group pays minimal SO cost. Thus, solving the SO for
24 heterogeneous problem becomes solving the SO for some homogeneous problems. The unique solution is
25 always founded. In the trivial case, the solution does not depend on the relationship between schedule
26 delay coefficients of all groups. Next, this paper considers the cases in which no separate group stands
27 separately in the SO assignment. Consequently, the relationship between schedule delay coefficients of all
28 groups is taken into account.
29 Each group has two schedule delay cost coefficients: β for early arrival, and γ for late arrival.
30 However, in the SO assignment, a group may depart before, or after, its preferred arrive time. Thus the
31 cost function β, or γ respectively, is used to generate the minimal objective cost (1). This paper defines
32 this coefficient as the “included coefficient”. For instance, if all commuters of group g depart after their
33 preferred arrival time in the SO assignment, γg is the included coefficient because γg is included in the
34 calculation of the total SO cost. βg is not the included coefficient because the objective value of
35 expression (1) do not include βg.
36 Theorem 1: The SO formulation (1) to (4) has a unique solution, in terms of the rate of departure, rt,g, if
37 and only if the included coefficients are different for all groups.
38 Proof:
39 The statement in Theorem 1 is equivalent to the statement: the SO formulation (1) to (4) has multiple
40 solutions if and only if the included coefficients are similar for at least two groups.
 6
Kien Doan, Satish Ukkusuri

1 Assume for the sake of contradiction that the SO formulation (1) to (4) has multiple solutions. It
2 implies that there are two different time intervals t1 and t2 in which two groups can exchange their
3 departure rates without increasing the optimal value. This paper will prove that these two groups must
4 have similar included coefficients.
5 The current departure rates of group g1 and group g2 at time t1 and t2 are rt1,g1, rt1,g2, rt2,g1, and rt2,g2
6 The current total cost at both time t1 and t2 is
7 c1 = rt1,g1δ1(t1*−t1) + rt1,g2δ2(t2*−t1) + rt2,g1δ1(t1*−t2) + rt2,g2δ2(t2*−t2) (5)
8 δ1 and δ2 are the included coefficients of group 1 and group 2.
9 If t1 ≤ t1*, δ1 = β1. Otherwise, δ1 = -γ1
10 If t2 ≤ t2*, δ2 = β2. Otherwise, δ2 = -γ2
11 The departure rates of group g1 and g2 at time t1 and t2 after exchanging an amount of departure rate σ>0:
12 rt1,g1+σ, rt1,g2−σ, rt2,g1−σ, and rt2,g2+σ
13 After departure rate exchanging, the total cost at both time t1 and t2 is:
14 c2 = (rt1,g1+σ)δ1(t1*−t1)+(rt1,g2−σ)δ2(t2*−t1)+(rt2,g1−σ)δ1(t1*−t2)+(rt2,g2+σ)δ2(t2*−t2) (6)
15 Because the exchanging does not increase the total system cost, then c1 = c2
16 From (5) and (6):
17 σδ1(t1*-t1) - σδ2(t2*-t1) - σδ1(t1*-t2) + σδ2(t2*-t2) = 0
18 ↔ σ(t2-t1)(δ1-δ2) = 0
19 ↔ δ1 = δ2 because σ > 0 and t1 ≠ t2
20 ↔ β1 = β2 or γ1 = γ2 because β1 = -γ2, or β2 = -γ1, never happens since penalty coefficient > 0
21 Therefore, the paper can conclude that the SO formulation (1) to (4) has multiple solutions if and
22 only if the included coefficients are similar to at least two groups. This conclusion is equivalent to
23 Theorem 1.
24 Theorem 1 shows that the number of solutions depends on the schedule delay parameters. Next, this
25 paper demonstrates that there is another approach to define the number of solutions. It is with respect to
26 the reduced cost patterns.
27 The reduced cost is the amount by which an objective function improves (decreases for the
28 minimization problem) before it would be possible for a corresponding variable to assume a positive
29 value in the optimal solution. It is the cost for increasing a variable by a small amount, i.e., the first
30 derivative from a certain point on the polyhedron that constraints the problem. In the SO formulation, the
31 reduced cost of group g at time t can provide information about how much the objective function
32 increases if a small amount of departure of group g moves to depart at time t. Thus, the reduced cost
33 patterns can conclude whether the SO formulation has a unique solution or multiple solutions. In the SO
34 formulation with g groups, there are g reduced cost patterns varying by time intervals. In each reduced
35 cost pattern, the basic variables must have a reduced cost that is exactly zero, and the non basic variables
36 have a reduced cost greater than zero. That means, for each group, the reduced cost is equal to zero at its
37 departure time and greater than zero otherwise. There are two circumstances of reduced cost patterns:
38 1) There is no time period in which two or more groups have the same zero reduced cost. In other
39 words, there is no overlap with zero-reduced cost in the reduced cost pattern of any two groups. It also
40 can be restated that every group has its reduced cost is equal to zero at the time period of its departure,
 7
Kien Doan, Satish Ukkusuri

1 while the reduced costs of other groups at that time period are greater than zero. From the definition of the
2 reduced cost, every exchange in departure time of two groups will make the total system cost increased.
3 In this case, the SO assignment is unique. Figure 2 is an illustration this case.

4
5 Figure 1 Unique solution illustrated by reduced cost pattern
6 In Figure 1, the red lines are the reduced cost pattern of groups 2. The reduced cost of group 2 is
7 equal to zero from time intervals 31 to 40 and greater than zero at remaining time intervals. From time
8 intervals 31 to 40, the reduced costs of two other groups are greater than zero. It implies that if some
9 commuters in group 2 change not to depart from time interval 31 to 40, they will increase the total system
10 cost, and if some commuters in group 1 or group 3 change to depart at time interval 31 to 40, they also
11 increase the total system cost. The same argument is used with the reduced cost patterns of group 1 and
12 group 3. In conclusion, this example has unique assignment of SO.
13 2) There is time period in which two or more groups have the same zero reduced cost. If both two
14 groups have a time period at which their reduced cost is equal to zero, they can freely exchange the
15 departure rate with each other without increasing the total system cost. Therefore the SO has multiple
16 solutions. Figure 2 is an illustration of the multiple solutions example.
17 In Figure 2, both group 1 and group 2 have reduced cost equal to zero from time intervals 30 to 60.
18 Then, the departure rate of two groups at that period can exchange to each other without increasing the
19 total system cost. Thus, there are an infinite number of assignments with the same minimal objective
20 value.
21 In conclusion, the three following statements are equivalent:
22 - SO assignment is unique
23 - The included coefficients are different for all groups
24 - There is no time period in which two or more groups have the same zero reduced cost.
25 The three following statements are equivalent:
26 - SO assignment is not unique
27 - The included coefficients are the same for at least two groups
 8
Kien Doan, Satish Ukkusuri

1 - There is time period in which two or more groups have the same zero reduced cost.

2
3 Figure 2 Multiple solutions illustrated by reduced cost pattern

4 2.4. Comparisons between the UE formulation and SO formulation

5 This sub-section compares the insight into the two discrete heterogeneous commuter models: the UE
6 model (Ramadurai et al, (9)), and the SO model (formulation (1) to (4)).
7 - First, the two models have different objectives. The UE model is to find the equilibrium state in
8 which there is no incentive for any commuter to shift to depart at other time. The assignment in the UE
9 model is to solve the trade-off between travel time and schedule delay of every commuter. On the other
10 hand, the SO model is to find the best assignment in terms of total cost. This assignment is determined by
11 combination in schedule delay of each group at each time interval. There is no travel time in SO model.
12 - The UE model has the unique solution in terms of Cg* of each group. The SO model has the unique
13 objective value in terms of total system cost.
14 - For departure rate vector rt,g, both UE model and SO model can have a unique solution or multiple
15 solutions. The number of UE solutions depends on the ratio of β/α and γ/α, while the number of SO
16 solutions depends on the absolute value of β and γ. In UE model, if two groups have same included β/α or
17 γ/α, they can depart at the same time without changing the equilibrium conditions. In SO model, if two
18 groups have similar included β or γ, they can depart at the same time without changing the minimal
19 objective value. Inversely, in the UE model, groups depart separately if there is no same included β/α and
20 γ/α. In SO model, groups depart separately if there is no similar included β or γ.
21 In summary, the UE assignments are more realistic although the commuters have to pay the travel
22 time cost. The SO assignments can minimize the total system cost but this assignment is not stable. In
23 section 3, the paper will analyze the pricing strategies to find a new state in which all commuters have
24 departure rate equal to the capacity but they have no incentive to shift their time.
25 3. Toll strategies
26 For the SB model with homogeneous commuters, Arnott et al. (2) showed that the SO can be obtained
27 when the travel time of every commuter is zero and the departure rate at every time is equal to the
 9
Kien Doan, Satish Ukkusuri

1 capacity. In order to get the SO, the “fine toll,” which is a piece-wise linear function, needs to be set. The
2 total cost of each driver in the SO with toll is similar to that in the no-toll UE conditions. This method to
3 directly find the SO from UE is straight forward for homogeneous commuter model. Both the UE and the
4 SO are unique and depend on the same parameters of cost functions. On the other hand, there is no
5 straight forward method to derive the SO from the UE in the heterogeneous commuter model. It is
6 because there is no connection between these two models. They have the different objective functions: the
7 UE is to find the equilibrium assignment in which there is no incentive for any commuter to shift their
8 departure time, while the SO is to find the minimal total system cost regardless the stableness of the
9 system. And most important, the solution of UE model depends on value of β/α and γ/α, while the
10 solution of the SO model depends on the absolute value of β and γ.
11 This paper develops a new approach to solve the SO problem by two steps
12 1) Find the SO assignment by formulation (1) to (4) in Section 2,
13 2) Find the toll strategies and set into the SO assignment to make it stable.
14 If the feasible toll is found, we can create a state in which the total system cost is minimized as in
15 the SO, and there is no incentive for any commuter shifting their departure time since as in the UE. To
16 avoid the confusion with the definition of UE and SO state, from now on this paper calls the new state as
17 stable system state (SSS). In the following sub-section, this paper will examine the conditions of the
18 feasible toll scheme for a general SB model with heterogeneous commuters in which different groups can
19 have different cost functions and preferred arrival time.
20 3.1. Necessary conditions for toll scheme
21 In this sub-section, this paper will provide the two conditions that the toll scheme must satisfy to
22 guarantee the equilibrium at the departure times of one group and two continuous groups.
23 3.1.1. Condition 1:
24 This paper states that: The toll must be set such that the total of toll and schedule delay cost for each
25 group is a constant.
26 In the SO assignment, commuters within a group pay different costs. In order to keep the
27 assignment of that group stable, the toll must complement the schedule delay cost such that the total is a
28 constant. Figure 3 is an illustration for Condition 1. After getting the SO assignment, the schedule delay
29 (SCD) is determined for each group. Thus, at each period of departure time, the slope of the toll scheme is
30 estimated such that the total cost (TC) is a constant for each group. The example in Figure 3a is
31 particularly for the unique SO solution. Condition 1 still works with a SO with multiple assignments.
32 Theorem 1 showed that in the overlap period, the penalty coefficients are similar. Thus, if there are two
33 continuous groups overlapping their departure time, the two toll patterns are similar (Figure 3b).
34 3.1.2. Condition 2:
35 This condition considers the equilibrium in the departure times of two continuous groups. Since the slope
36 of toll for each group is defined in condition 1, this sub-section takes into account the difference in toll
37 values at the transition between two groups. Since the SO formulation in this paper is a discrete time
38 model, there are two possible transitions as follows:
39 - Two groups share the transition time interval t in which the total departure rate of two groups is
40 equal to the capacity.
 10
Kien Doan, Satish Ukkusuri

1 - The first group finishes departing at time t and the other begins departing at time t+1.

2
3 Figure 3 Toll scheme pattern in Condition 1: a) Unique SO b) Multiple SO
4 In the first case, the tolls at time t-1 and time t have to satisfy Condition 1 for the first group. And
5 the tolls at time t and time t+1 have to satisfy Condition 1 for the second group. Therefore, the toll
6 patterns of two groups are continuous because they use same toll value at time t. In the second case with
7 the separate transition, the toll value at time t can defer the toll value at time t+1 without affecting the
8 equilibrium at departure time of these two groups. Theorem 2 will consider the separate transitions
9 between two different groups with different included coefficients. If these two groups have similar
10 included coefficient, the toll difference between two continuous time intervals is equal to the included
11 coefficient (Figure 3b).
12 Theorem 2: If two groups have separate transitions t and t+1, the toll requirements (p) at the transition
13 times between two groups are as follows:
14 Case A: t+1 ≤ min(t1*, t2*) → β1 ≤ pt+1 – pt ≤ β2 (7)
*
15 Case B: t ≥ max(t1 , t2*) → γ2 ≤ pt – pt+1 ≤ γ1 (8)
*
16 Case C: t ≥ t1 , t+1 ≤ t2* → -γ1 ≤ pt+1 – pt ≤ β2 (9)
17 Proof:
18 Three cases above are three situations in which the separate transitions t and t+1 can occur:
19 Case A: the transitions take place before t* of two groups
20 Case B: the transitions take place after t* of two groups
21 Case C: the transitions take place at time between t1* and t2* of two groups
22 * Case A: t+1 ≤ min(t1*, t2*): the transitions happens before t* of each group
23 ↔ t < t1*, t+1 ≤ t2*:
24 The total of schedule delay cost at t and t+1 is: c1 = β1 (t1* − t) + β2(t2* − t − 1) (10)
25 Since the assignment is optimal and has a unique solution → β1 ≠ β2
26 Because the current assignment is SO, this paper claims that β1 < β2 (11)
 11
Kien Doan, Satish Ukkusuri

1 Using contradiction, this paper proves that if β1 > β2, the original assignment is not optimal because
2 there is a new assignment with lower total cost. Particularly, commuters of group 1 shift to depart at t+1
3 and commuters of group 2 shift to depart at t, which can get lower system cost.
4 The cost after shifting is: c2 = β2 (t2* − t) + β1(t2* − t − 1) (12)
5 Then, subtracting (10) from (12)
6 c2 – c1 = [β2(t2* − t) + β1(t2* − t − 1)] – [β1 (t1* − t) + β2(t2* − t − 1)] = β2 – β1 < 0, (13)
7 ↔ c2 < c1 , which is the contradiction with SO assignment. Thus β1 < β2
8 In order to guarantee the equilibrium at departure time of two groups after the toll is set, this paper
9 claims that the toll difference between t and t+1 satisfies: β1 ≤ pt+1 – pt ≤ β2 (14)
10 Assume for the sake of contradiction that: pt+1 – pt > β2
11 The total cost of group 2 at time t is: β2 (t2* − t) + pt (15)
12 The total cost of group 2 at time t+1 is: β2(t2* − t − 1) + pt+1 (16)
13 Subtract (16) from (15): [β2 (t2* *
− t) + pt] – [β2(t2 − t − 1) + pt+1] = pt – pt+1 + β2 < 0 (17)
14 That means the commuters of group 2 at time t+1 have an incentive to shift to time t, which is a
15 contradiction to the equilibrium conditions of group 2. Thus, pt+1 – pt ≤ β2 (18)
16 Next, assume for the sake of contradiction that: β1 > pt+1 – pt
17 The total cost of group 1 at time t: β1 (t1* − t) + pt (19)
18 The total cost of group 1 at time t+1: β1(t1* − t − 1) + pt+1 (20)
* *
19 Subtract (20) from (19): [β1 (t1 − t) + pt] – [β1(t1 − t − 1) + pt+1] = pt – pt+1 + β1 > 0 (21)
20 That means commuters of group 1 at time t have incentive to shift to time t+1, which is a
21 contradiction to the equilibrium conditions of group 1. Thus, β1 ≤ pt+1 – pt (22)
22 Therefore, from (18) and (22), condition (14) holds, and β1 ≤ pt+1 – pt ≤ β2. That means the
23 difference between toll values at time t+1 and time t is not less than β1 and not greater than β2.

24
25 Figure 4 Toll conditions at transition times: a) Case A b) Case B
26 Figure 4a illustrates Case A in which the difference between the toll values pt and pt+1 at the
27 transition time intervals t and t+1 satisfy expression (14). If the toll at time t is too low (or the toll at time
28 t+1 is too high), the total cost of group 2 at time t will be less than the equilibrium cost of group 2 (the
 12
Kien Doan, Satish Ukkusuri

1 solid green line in Figure 4a, and group 2 will change its assignment. If the toll at time t is too high (or the
2 toll at time t+1 is too low), the total cost of group 1 at time t+1 will be less than the equilibrium cost of
3 group 1 (the solid red line in Figure 4a), and group 1 will change its assignment.
4 * Case B: t ≥ max(t1*, t2*):
5 Similar to case A, case B considers the transition time intervals after t* of each group
6 With the same arguments, this paper obtains the result γ2 ≤ pt – pt+1 ≤ γ1. That means the difference
7 between toll values at time t+1 and time t is not less than γ2 and not greater than γ1.
8 Figure 4b demonstrates Case B. If the toll at time t is too low (or the toll at time t+1 is too high), the
9 total cost of group 2 at time t will be less than the equilibrium cost of group 2 (the solid green line in
10 Figure 10), and group 2 will change its assignment.
11 * Case C: t > t1* and t+1 < t2* (the transition happens after t1* and before t2*)
12 Without loss of generality, assuming that group 1 departs before group 2, this paper claims that t1* < t2*.
13 If t1* = t2*, we return two Case A or Case B above when the transition times are before or after t* of
14 two groups. If t1* > t2*, this condition is not optimal. Commuters in group 1 can move to depart at t+1 or
15 after t+1, and commuters in group 2 can move to depart at t or before t to decrease the total system cost
16 (Figure 5). Therefore, t1* < t2*

17
18 Figure 5 An assignment which is not system optimal

19 The schedule delay cost of group 1 at time t is: SCDt1 = (t − t1*)γ1

20 The schedule delay cost of group 1 at time t+1 is: SCDt1+1 = (t + 1 − t1*)γ1

21 The schedule delay cost of group 2 at time t is: SCDt2 = (t2* − t)β2

22 The schedule delay cost of group 2 at time t+1 is: SCDt2+1 = (t2* − t − 1)β2
23 This paper claims that -γ1 ≤ pt+1 – pt ≤ β2 (23)
24 Assume for the sake of contradiction that pt+1 – pt > β2,
25 The total cost of group 2 at time t is: (t2* − t)β2 + pt (24)
*
26 The total cost of group 2 at time t+1 is: (t2 − t − 1)β2 + pt+1 (25)
27 Subtract (25) from (23): [(t2* − t)β2 + pt] – [(t2* − t − 1)β2 + pt+1] = pt – pt+1 + β2 < 0
28 That means commuters of group 2 at time t+1 have an incentive to move to time t, which is the
29 contradiction. Thus pt+1 – pt ≤ β2 (26)
30 Assume for the sake of contradiction that -γ1 > pt+1 – pt,
 13
Kien Doan, Satish Ukkusuri

1 The total cost of group 1 at time t is: (t − t1*)γ1 + pt (27)

2 The total cost of group 2 at time t+1 is: (t + 1 − t1*)γ1 + pt+1 (28)
3 Subtract (28) from (27)
4 [(t − t1*)γ1 + pt] – [(t + 1 − t1*)γ1 + pt+1] = pt – pt+1 − γ1 > 0
5 That means commuters of group 1 at time t have an incentive to move to time t+1, which is the
6 contradiction. Thus -γ1 ≤ pt+1 – pt (29)
7 From (26) and (29), expression (23) holds, and -γ1 ≤ pt+1 – pt ≤ β2

8
9 Figure 6 Toll conditions at transition times for Case C
10 In Figure 6, with the similar arguments, if the toll at time t is too low (or the toll at time t+1 is too
11 high), the total cost of group 2 at time t will be less than the equilibrium cost of group 2 (the solid green
12 line in Figure 12), and group 2 will change its assignment.
13 3.1.3. Condition 3:
14 This sub-section will prove that, within a special case in which a group departs at two periods of
15 time intervals, there always exist toll schemes such that this group is equilibrium in both two time
16 periods. The condition for toll scheme in this case is also described. This condition need be considered
17 because, without it, the toll scheme, which has already satisfy Condition 1 and 2, cannot guarantee the
18 same cost at these two periods of time intervals. It will make the assignment of this group is not stable.
19 Theorem 3: If one group departs at two separate periods of time, and both of the transitions are
20 combined by two continuous time intervals, there exist the toll schemes to guarantee the equilibrium cost
21 at two periods.
22 Proof:
23 There are two cases in which a group is separated by two periods of departure:
24 1) One transition time is less than t* and the other is greater than t* of that group (Figure 7a)
25 2) Both transition times are less than or greater than t* of that group (Figure 7b)
26 In order to obtain the equilibrium of group 1, the following expression must hold:
27 pt + SCDt = pt’+1 + SCDt’+1 (30)
28 Let pt be fixed, this paper shows that the toll pt’+1 is always found such that equation (30) holds.
 14
Kien Doan, Satish Ukkusuri

2
3 Figure 7 A group departs at two periods of time intervals

4 a) One transition time is less than t* and the other is greater than t*

5 b) Both transition times are less than or greater than t* of that group
6 * Case 1:
7 Since the assignment is the SO, the following inequalities hold:
8 β1(t1* − t) + γ2(t’ – t2*) < β2(t2* − t) + γ1(t’ – t1*) (31)
9 β2(t2* *
− t − 1) + γ1(t’ + 1 – t1 ) < β1(t1* − t − 1) + γ2(t’ + 1 – t2*) (32)
10 Inequality (31) shows that, at the SO assignment, any shift between departures at time t and time t’
11 will make the schedule delay cost increased. Similarly, inequality (32) states that any shift between
12 departures at time t+1 and time t’+1 will make the schedule delay cost increased.
13 Rearranging inequality (31) and (32), we have:
14 β1(t1* − t) − γ1(t’ – t1* + 1) < β2(t2* − t − 1) − γ2(t’ – t2*) + β2 – γ1 (33)
15 β2(t2* * *
− t − 1) − γ2(t’– t2 ) + β1 – γ2 < β1(t1 − t) – γ1(t’ + 1 – t1 ) *
(34)
16 Combine (33) and (34):
17 β2(t2* − t − 1) − γ2(t’– t2*) + β1 – γ2 < β1(t1* − t) – γ1(t’ + 1 – t1*) < β2(t2* − t − 1) − γ2(t’ – t2*) + β2 – γ1
18 ↔ β1 + (SCDt+1 – SCDt’) – γ2 < (SCDt – SCDt’+1) < β2 + (SCDt+1 – SCDt’) – γ1
19 ↔ β1 + pt + (SCDt+1 – SCDt’) – γ2 < pt + (SCDt – SCDt’+1) < β2 + pt + (SCDt+1 – SCDt’) – γ1 (35)
20 On the other hand, applying Theorem 2 to Case 1, we have:
21 β1 ≤ pt+1 – pt ≤ β2 (36)
22 γ1 ≤ pt’ – pt’+1 ≤ γ2 (37)
23 From (36): β1 + pt ≤ pt+1 ≤ β2 + pt (38)
24 Furthermore, to make group 2 in equilibrium, pt’ = pt+1 + (SCDt+1 – SCDt’) (39)
25 Add (SCDt+1 – SCDt’) to all expressions in (38)
26 → β1 + pt + (SCDt+1 – SCDt’) ≤ pt’ ≤ β2 + pt + (SCDt+1 – SCDt’) (40)
27 From (37): pt’ − γ2 ≤ pt’+1 ≤ pt’ − γ1
 15
Kien Doan, Satish Ukkusuri

1 Substitute pt’ from (40):

2 → β1 + pt + (SCDt+1 – SCDt’) – γ2 ≤ pt’+1 ≤ β2 + pt + (SCDt+1 – SCDt’) – γ1 (41)
3 From (35) and (41), since pt + (SCDt – SCDt’+1) and pt’+1 are bounded by the same values, we can
4 always find pt’+1 = pt + (SCDt – SCDt’+1). Thus, the equation (30) holds. That means we can always find
5 the toll such that the equilibrium condition at two periods of time intervals of group 1 is guaranteed.
6 From equation (30) and (39), this paper finds the relationship between toll pt+1 and toll pt’+1 because
7 that two tolls cannot be much different, as the result of Theorem 2 for the transition times.
8 Substitute pt = pt’+1 + (SCDt’+1 – SCDt) to inequalities (38):
9 β1 + pt’+1 + (SCDt’+1 – SCDt) ≤ pt+1 ≤ β2 + pt’+1 + (SCDt’+1 – SCDt)
10 or pt+1 – β2 – (SCDt’+1 – SCDt) ≤ pt’+1 ≤ pt+1 – β1 – (SCDt’+1 – SCDt)
11 * Case 2:
12 Case 2 is proven similarly as Case 1, with the properties of SO and the findings from Theorem 2 (β1 ≤ pt+1
13 – pt ≤ β2 and -γ2 ≤ pt’+1 – pt’ ≤ β1)
14 At the end, the relationship between toll pt+1 and toll pt’+1 is:
15 β1 + pt’+1 + (SCDt’+1 – SCDt) ≤ pt+1 ≤ β2 + pt’+1 + (SCDt’+1 – SCDt)
16 From the result of Theorem 3, there are two lemmas consider two cases in which the transition time
17 is shared by two continuous groups.
18 Lemma 1: If one group departs at two separate periods of time, and one transition is combined by two
19 continuous time intervals, one transition is a shared time interval, the toll can always be found to
20 guarantee the equilibrium cost at two periods.
21 Proof:
22 Lemma 1 contains two cases as in Figure 7.
23 1) One transition time is less than t* and the other is greater than t* of that group (Figure 7a)
24 2) Both transition times are less than or greater than t* of that group (Figure 7b)
25 However, one transition includes t and t+1 (group 1 departs at t, group 2 depart at t+1), the other
26 transition includes t’ (both group 2 and group 1 depart at t’). Thus pt’ – pt’+1 = γ1
27 Let pt be fixed, this paper proves that toll pt+1 can be found such that the equilibrium between two
28 periods of time intervals of group 1 is guaranteed.
29 * Case 1:
30 In case 1 of Theorem 3, γ1 ≤ pt’ – pt’+1 ≤ γ2
31 In case 1 of Lemma 1, pt’ – pt’+1 = γ1
32 All other steps are the same. Similar to (35), we have:
33 β1 + pt + (SCDt+1 – SCDt’) – γ1 < pt + (SCDt – SCDt’+1) < β2 + pt + (SCDt+1 – SCDt’) – γ1
34 Similar to (41), we have:
35 β1 + pt + (SCDt+1 – SCDt’) – γ1 ≤ pt’+1 ≤ β2 + pt + (SCDt+1 – SCDt’) – γ1
36 Thus, there exist pt’+1 = pt + (SCDt – SCDt’+1), and the equilibrium hold for group 1 at two periods of
37 time intervals.
38 The relationship between tolls pt+1 and toll pt’+1 is:
39 β1 + pt’+1 + (SCDt’+1 – SCDt) ≤ pt+1 ≤ β2 + pt’+1 + (SCDt’+1 – SCDt)
 16
Kien Doan, Satish Ukkusuri

1 or pt+1 – β2 – (SCDt’+1 – SCDt) ≤ pt’+1 ≤ pt+1 – β1 – (SCDt’+1 – SCDt)

2 * Case 2:
3 Case 2 is proven similarly as Case 1, with findings from Theorem 2 (β1 ≤ pt+1 – pt ≤ β2) and property of
4 new condition (pt’+1 – pt’ ≤ β1)
5 At the end, the relationship between toll pt+1 and toll pt’+1 is:
6 β1 + pt’+1 + (SCDt’+1 – SCDt) ≤ pt+1 ≤ β2 + pt’+1 + (SCDt’+1 – SCDt)
7 or pt+1 – β2 – (SCDt’+1 – SCDt) ≤ pt’+1 ≤ pt+1 – β1 – (SCDt’+1 – SCDt)
8 Lemma 2: If one group departs at two separate periods of time, both of two transitions are the shared
9 time intervals, the toll can always be found to guarantee the equilibrium cost at two periods.
10 Proof:
11 There are two possible cases as in Theorem 2 and Lemma 1. Like Lemma 1, this paper proves Case 1.
12 Case 2 can be proven by the same arguments.
13 * Case 1:
14 Without loss of generality, assume that time t and t’+1 are the shared transition time intervals. Both two
15 groups depart at times t and t’+1. Then these following two equations must hold:
16 pt+1 – pt = β2 (pt and pt+1 satisfy the equilibrium of group 2 at time t and t+1) (42)
17 pt’ – pt’+1 = γ1 (pt’ and pt’+1 satisfy the equilibrium of group 1 at time t’ and t’+1) (43)
18 Furthermore, to satisfy the equilibrium of group 2, pt’ = pt+1 + SCDt+1 – SCDt’ (44)
19 Substitute (44) to (43): pt+1 + SCDt+1 – SCDt’ – pt’+1 = γ1 (45)
20 Substitute pt+1 from (42) to (45): β2 + pt + SCDt+1 – SCDt’ – pt’+1 = γ1
21 ↔ pt – pt’+1 = γ1 + SCDt’ – SCDt+1 – β2 = SCDt’+1 – SCDt (46)
22 That means in Case 1, toll at t and t’+1 always satisfy the equilibrium of group 1 at two periods of
23 time intervals.
24 3.2. Sufficient condition for SSS after adding toll (Condition 4)
25 This sub-section provides the conditions of the input parameters with which the toll scheme could be
26 feasible. After adding the toll scheme with all conditions in sub-section 3.1, the final step is to check if
27 each group pays the minimal cost in its cost pattern. There are special cases in which although toll scheme
28 satisfy all of the necessary conditions in sub-section 3.1, the equilibrium of some groups is not guaranteed
29 because they can shift to other departure time to make them better off. This paper clarifies some
30 definitions:
31 - The schedule delay pattern of each group is the V-shape pattern representing the schedule delay
32 cost of that group at each time.
33 - The toll pattern is formed by the toll at each time interval.
34 - The cost pattern of each group is the total of the schedule delay and the toll.
35 This condition states that:
36 The sufficient condition for a feasible toll is that the equilibrium cost of each group must be the
37 minimal cost in their own cost patterns.
 17
Kien Doan, Satish Ukkusuri

1 Assume for the sake of contradiction that the equilibrium cost of group g is not the minimal cost in
2 group g cost pattern. Then all commuters in group g have an incentive to move to the time with which
3 they can pay less. Thus, the equilibrium condition does not hold. This theorem is to confirm that every
4 group pays minimal cost in their own cost pattern so that the equilibrium of every group is guaranteed.
5 Actually there are some situations in which no toll can satisfy the sufficient condition. The
6 following example illustrates the infeasible of toll strategies for SO assignment.
Cost

Time
7
8 Figure 8 Example of no feasible toll scheme
9 In this example, group 1 has high value of early coefficient β. However, since the preferred arrival
10 time of group 2 is earlier than other two groups, group 1 with higher β departs first. It make the high slope
11 of toll at the period group 1 departs, which satisfies the Condition 1. For group 2 cost pattern, the high
12 slope of toll makes the group 2 cost very low at some time intervals. In Figure 8, the minimal point in
13 group 2 cost pattern is even smaller than the equilibrium cost (see the red line). Thus group 2 is not in the
14 equilibrium state. In conclusion, even though the toll scheme satisfy all necessary conditions (Condition
15 1,2,3), the sufficient condition (Condition 4) is not guaranteed for group 2. Therefore, there is no feasible
16 toll scheme to create the SSS.
17 3.3. Conclusion
18 This paper has formulated the system optimal conditions by linear programming. The existence of
19 solutions and the insight into the number of solutions are analyzed. Then, the pricing scheme is studied to
20 obtain a stable system optimal assignment. The necessary conditions of tolls and the sufficient conditions
21 of input parameters are studied to guarantee that the stable system state can be obtained. This paper
22 considers the most general cases of heterogeneity parameter assumptions. Thus, it is an important
23 contribution to the literature in which previous studies consider only some limited set of heterogeneity
24 parameter assumptions.
 18
Kien Doan, Satish Ukkusuri

1 REFERENCES
2 1. Vickrey, W. S., 1969. Congestion theory and transport investment. American Economic Review 59
3 251-261.
4 2. Arnott, R., A., dePalma, R., Lindsey, 1993. A structural model of Peak-priod congestion: a traffic
5 bottleneck with elastic demand. The American Economic Review, Vol. 83, No. 1 (Mar., 1993), pp. 161-179.
6 3. Daganzo, C., 1985. The uniqueness of a time-dependent equilibrium distribution of arrivals at a SB.
7 Transportation Science. Vol. 19, No. 1, February 1985.
8 4. Ramadurai, G., S., Ukkusuri, 2006. Dynamic traffic equilibrium: theoretical and experimental
9 network game results in the SB model. Transportation research board record. Aug 2006.
10 5. Mun, S., 1999. Peak-load pricing of a bottleneck with traffic jam. Journal of Urban Economics 46,
11 323-349 (1999).
12 6. Newell, G., 1987. The morning commute for nonidentical travelers. Transportation Science. Vol. 21,
13 No. 2, May 1987.
14 7. Arnott, R., A., dePalma, R., Lindsey, 1989. Schedule delay and departure time decisions with
15 heterogeneous commuters. Transportation Research Record 1197 56-67.
16 8. Lindsey, R., 2004. Existence, uniqueness, and trip cost function properties of user equilibrium in the
17 bottleneck model with multiple user classes. Transportation Science 38(3) 293-314.
18 9. Ramadurai, G., S.V., Ukkusuri, S.J, Pang, Z., Zhao, 2009. Linear commplementarity formulation for
19 the SB model with heterogeneous users. Accepted for publication in Transportation research Part B.
20 10. Ramadurai, G., 2009. Novel dynamic user equilibrium models: analytical formulations, multi-
21 dimensional choice, and an efficient algorithm. PhD dissertation.
22 11. Arnott, R., A., dePalma, R., Lindsey, 1994. A Welfare congestion tolls with heterogenous commuters.
23 Journal of Transportation Economics and Policy. Vol. 28, No. 2 (May, 1994), pp. 139-161.
24 12. Laih, C.H., 1994. Queueing at a bottleneck with single-and multi-step tolls. Transportation research
25 Part A. Vol. 28A, No.3, pp. 197-208, 1994.
26 13. Daganzo, C., R., Garcia, 2000. A pareto improving strategy for the time-dependent morning commute
27 problem. Transportation Science. Vol. 34, No. 3, August 2000.
28 14. Shaffi, J., 1885. Urban transportation networks: Equilibrium Analysis with Mathematical
29 Programming Methods.
30 15. Friesz, L. T., J. Luque, R. L. Tobin, B. Wie, 1989. Dynamic network traffic assignment considered as
31 a continuous time optimal control problem. Operations Research; Nov/Dec 1989; 37, 6
32