Eigenvalue and Eigenvector Analysis of Stability

for a Line of Traffic

By Liang Wang, Berthold K. P. Horn, and Gilbert Strang

Many authors have recognized that traffic under the traditional car-following
model (CFM) is subject to flow instabilities. A recent model achieves sta-
bility using bilateral control (BCM)—by looking both forward and backward
[1]. (Looking back may be difficult or distracting for human drivers, but is
not a problem for sensors.) We analyze the underlying systems of differen-
tial equations by studying their eigenvalues and eigenvectors under various
boundary conditions. Simulations further confirm that bilateral control can
avoid instabilities and reduce the chance of collisions.

1. Introduction

In the traditional car-following model (CFM), the state (including relative

position and relative speed) of the current car is controlled to more closely
match the state of the leading car [2–7]. To avoid the instability and
collision inherent in the car-following system, Horn proposed the bilateral
control model (BCM) [1], in which the state of the current car is controlled
to more closely match the average of the states of the leading and following
cars. (This is more easily implemented with automatic vehicular control.)
In this paper, we give a theoretical analysis of both the CFM and
the BCM, by analyzing the eigenvalues and eigenvectors of two ordinary
differential equations. Moreover, we give analytical solutions of both models
under different boundary conditions, something that does not appear in

Address for correspondence: Prof. Berthold K. P. Horn, Department of Electrical Engineering and
Computer Science, Massachusetts Institute of Technology, Cambridge, MA 02139, USA; e-mail:
[email protected]

DOI: 10.1111/sapm.12144 1
STUDIES IN APPLIED MATHEMATICS 00:1–30

C 2016 Wiley Periodicals, Inc., A Wiley Company
2 L. Wang et al.

the original paper [1]. We prove that bilateral control suppresses traffic
instabilities under all the different boundary conditions—infinite line, cir-
cular boundary, fixed-fixed boundaries, free-free boundaries, and fixed-free
boundaries. In contrast, the CFM is unstable for any and all of the boundary
conditions. For the CFM under the fixed boundary condition, the corre-
sponding “big matrix” of the ODE system is similar to a block Jordan form.
There are only two eigenvectors. The “big matrix” is not diagonalizable. We
analyze this interesting case and show that its performance is very similar
to the case of a circular boundary condition—in which discrete Fourier
transform (DFT) analysis can be applied. Thus, the overall car-traffic system
will become unstable and collisions will occur.
Simulation results confirm the instability of the CFM. A tiny perturbation
in an equilibrium state will soon cause a traffic jam and collisions. Bilateral
control can suppress the traffic flow instability and collision very effectively.
The traffic system soon returns to an equilibrium state, even in the case
that bilateral control is only turned on just before a traffic jam becomes
imminent.

2. Related work

A previous attempt at solving traffic flow problems was in the context

of a platoon [8–12]. More theoretical analysis of various platoon models
was developed in Refs. 13, 14. We should mention that the implementation
of bilateral control is very different from that of a platoon. There are
significant differences between the two models:
• All cars in a platoon are led by the first car (for one directional
platoon) or both first and last cars (for bi-directional platoon).
However, bilateral control has no leaders. Every car adjusts its state
(position and speed) based on information about the states of its
neighbors.
• For platoon models, information about the states of cars in the pla-
toon is transmitted via global communication. For bilateral control,
global communication between the cars is not used. The informa-
tion about the distance and relative speed of the neighboring cars
is measured using sensors on the current car. (Note that bilateral
control does not exclude the possibility of global communication
between the vehicles. Information about neighboring cars obtained by
a communication system can be fused with measurement from the
sensors to generate more accurate estimations of the states and to
increase the robustness even further [15].)
Bilateral control is more flexible than the platoon model. No other cars
are allowed to insert themselves into a platoon, while, other cars can merge
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 3

into the traffic flow under bilateral control. Moreover, at least as presently
envisaged, the size of each platoon is small (several cars in general), while
bilateral control focuses on the whole traffic flow. It seems more practical to
use sensor-based control technology (e.g., bilateral control) to solve traffic
problem than to build vehicular networking infrastructure.
Mathematically, bilateral control and bi-directional platoon control use the
information of the positions of both leading and following cars for control
[11, 12]. However, bi-directional platoon control uses the velocity of the first
car (or a prespecified velocity) to adjust the states of all cars in the platoon
[11, 12]; while bilateral control uses the speed of the neighboring cars to
adjust the state of the current car.
Before the invention of bilateral control, there were attempts at using
bi-directional information. Nakayama et al. use bi-directional information
to improve the traditional optimal velocity (OV) model [16]. Treiber and
Helbing use bi-directional information to improve throughput of the traffic
system [17]. Horn pointed out that (1) the information from the leading
and following cars should contribute equally and (2) adding “dampers” will
cause traveling waves to die out [1]. He called this new model bilateral
control and proved its ability to eliminate the “phantom traffic jams” and
“stop-and-go” instabilities (caused by the CFM).

3. Summary of the CFM and BCM

Let xi (t) be the position of the i-th car, and vi (t) = ẋi (t) be its velocity.1
The pair {xi (t), vi (t)} gives the state of the i-th car, which is adjusted
through the acceleration ai (t) = ẍi (t) commanded by the control system. For
the CFM,2

ai = kd (xi−1 − xi − s) + kv (vi−1 − vi ), (1)

where s is known as the safe distance, kd > 0 and kv > 0 are the
proportional and derivative gains, respectively. In this model, control of car
i is based only on the relative position and relative velocity of car i − 1
immediately ahead. For the BCM,
1 1
ai = kd ((xi−1 − xi ) − (xi − xi+1 )) + kv ((vi−1 − vi ) − (vi − vi+1 )). (2)
2 2

1 Note that x
i−1 and x i denote the positions of the leading and current cars. The positive direction is
chosen as the direction in which cars are moving, thus, xi−1 − xi > 0 (see Fig. 1).
2 Here, we consider only the case where the velocity of the cars is all between zero and the speed
limit of the highway. Otherwise, the difference between the car’s current velocity and a desired
velocity, denoted by vi − vdes , should be added as a control input [1].
4 L. Wang et al.

Figure 1. Illustration of the car-following model and bilateral control model. The blocks
with “L”, “C,” and “F” denote the leading car, current car, and following car. (a)
Car-following control is based only on the state of the leading car “L”. (b) Bilateral control
uses the states of both leading car “L” and following car “F”.

Now the control of car i is based on the relative positions and relative
velocities of both car i − 1 ahead and car i + 1 behind. Figure 1 shows the
two models.
In the ideal case, all the cars are spaced the safe distance s apart and
move at the same speed v0 . In this case, all the accelerations ai in Eqs.
(1) and (2) are zero, and the traffic continues in the equilibrium state. The
important question then is whether this equilibrium is stable, meta-stable, or
unstable. If there is a small perturbation in xi (t) or vi (t), will the traffic
system return to the equilibrium state or will there be increasing departures
from the equilibrium, which ultimately lead to a traffic jam? To answer this
question, we will analyze Eqs. (1) and (2).

4. The eigenvalue-and-eigenvector–based analysis

For convenience, we change variables:

yi (t) = xi (t) + i × s − v0 t. (3)
Then Eqs. (1) and (2) can be written in the form:
CFM: ÿi + kv ẏi + kd yi = kv ẏi−1 + kd yi−1 . (4)

1 1
BCM: ÿi + kv ẏi + kd yi = kv ( ẏi−1 + ẏi+1 ) + kd (yi−1 + yi+1 ) . (5)
2 2
Now {yi (t), ẏi (t)} are the state of the i-th car in the relative reference
system (relative position and relative speed). Note that the left-hand sides of
Eqs. (4) and (5) are the same, and correspond to a spring-damper-mass
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 5

Figure 2. The second-order ODEs (4) and (5) describe a spring-damper-mass system. The
right-hand side of Eqs. (4) and (5) are the external inputs. (a) If the right-hand side of Eq.
(4) or (5) is zero, then the system is attached to a fixed wall. (b) The external inputs of
the system can be imagined as a “moving” wall. For the CFM, the state of the “moving”
wall is chosen as the state of the leading car. For BCM, the state of the “moving” wall is
chosen as the average of the states of the leading and following cars. (c) Bilateral control
obeys Newton’s third law. All the spring-damper-mass modules can be cascaded physically
as a large system. That is not the case for the CFM.

system. The right-hand sides of Eqs. (4) and (5) describe the external input
to that spring-damper-mass system. The external inputs of the system can be
imagined as a “moving” wall (see Fig. 2b). For the CFM, the state of the
“moving” wall is chosen as the state of the leading car. For BCM, the state
of the “moving” wall is chosen as the average of the states of the leading
and following cars.
First, note that the CFM does not correspond directly to a real mechanical
system, because it does not obey Newton’s third law. The state of the
leading car acts on the effective spring and damper connected to the i-th car,
while the car i − 1 does not react to changes from the following driver i.
It is as if a “buffer” device was inserted that replicates the position of the
leading car. The spring and damper would be attached to this “buffer,” not
the leading car directly. The BCM on the other hand does correspond to a
real physical system and obeys Newton’s third law. This makes it possible
to employ physical intuition for BCM. For instance, the spring-damper-mass
modules can be cascaded physically as one large system (see Fig. 2c).
Intuitively, the waves caused by perturbation of one of the masses will travel
in two opposite directions, and be damped as they travel.

4.1. The “big ODE system”

For easy analysis, we rewrite the second-order ODE as two first-order ODEs.
Equations (4) and (5) can be written as:

d
CFM: Yi = NY i−1 + MY i , (6)
dt
6 L. Wang et al.

d 1 1
BCM: Yi = NY i−1 + MY i + NY i+1 , (7)
dt 2 2
where
     
yi 0 0 0 1
Yi = , N= , M= .
ẏi kd kv −kd −kv
Let Y = (· · · , Yi−1
T
, YiT , Yi+1
T
, · · · )T be the state vector containing the states
of all the cars in the traffic system. From Eqs. (6) and (7), we can build two
big ODE systems:
d d
(CFM) Y = AY and (BCM) Y = BY. (8)
dt dt
A and B are both block-constant-diagonal (or block-Toeplitz) matrices:
⎛ ⎞ ⎛ ⎞
.. .. ..
. . .
⎜. ⎟ ⎜ ⎟
⎜ .. ⎟ ⎜ . ⎟
⎜ M ⎟ ⎜ .. M N /2 ⎟
⎜ ⎟ ⎜ ⎟
A=⎜ N M ⎟ , B = ⎜ N /2 M N /2 ⎟.
⎜ ⎟ ⎜ ⎟
⎝ N M ⎠ ⎜ N /2 M
..
. ⎟
⎝ ⎠
.. .. . .
. . .. ..

The solutions are3

(CFM) Y(t) = etA Y(0) and (BCM) Y(t) = etB Y(0), (9)
where Y(0) is the initial state of the traffic system. The traffic’s stability can
be determined by analyzing the eigenvalues of the “big matrices” A and B
[18].

4.2. Eigenvalues and eigenvectors of the “Big Matrix”

First, if M and N were scalars (rather than 2 × 2 matrices), then A
and B would both be Toeplitz matrices (i.e., linear convolution systems).
The discrete-time Fourier transform (DTFT) will solve the eigenvalue-and-
eigenvector problem [19]. The eigenvector u(ω) should be of the form:
u(ω) = (· · · , wl−1 , wl , wl+1 , · · · )T ,
where w = w(ω) = e− jω , and the corresponding eigenvalue (for CFM)
would be
λ(ω) = M + N w. (10)


3 Fora doubly infinite Toeplitz matrix A, we can define etA Y(0) = S eλ(ω)t S −1 Y(0) . Here, S −1 Y(0)
denotes the DTFT of Y(0), S( f (ω)) denotes the inverse DTFT of f (ω), and λ(ω) —called the
eigenvalue (function) of A—is the DTFT of one row of A.
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 7

In our case, M and N are both 2 × 2 matrices. One more step is needed.
The eigenvector U (ω) is the Kronecker tensor product of u(ω) and a 2 × 1
vector p(ω) = (a(ω), b(ω))T :
U (ω) = u(ω) ⊗ p(ω)
= (· · · , a(ω)wl−1 , b(ω)wl−1 , a(ω)wl , b(ω)wl , a(ω)wl+1 , b(ω)wl+1 , · · · )T , (11)
where w = e− jω as above. For CFM, the eigenvalue λ(ω) corresponding to
U (ω) can be found by solving
λ(ω) p(ω) = M p(ω) + w(ω)N p(ω).
That is, λ(ω) and p(ω) are the eigenvalue and eigenvector of the 2 × 2
matrix M + w(ω)N . The corresponding characteristic equation is
λ2 (ω) + (1 − w(ω))kv λ(ω) + (1 − w(ω))kd = 0. (12)
Note that the sum of the real parts of the two eigenvalues is smaller or equal
to zero. Thus, the real part of one of the eigenvalues must be smaller or
equal to zero. It is not as easy to determine the sign of the real part of the
other eigenvalue. However, the result is as follows:
• the real part of the other root of Eq. (12) is greater than zero, except
in the extreme case when kd = 0.

One way of seeing this conclusion is to consider the case that ω is pretty
small, such that 1 − e− jω ≈ jω. Then Eq. (12) can be approximated by
λ2 (ω) + jωkv λ(ω) + jωkd = 0.
The real part of the two roots are
√ 
2
2
± −ω kv + ω2 kv2 + 16ω2 kd2 .
2 2
4
The real part of both roots is zero if and only if kd = 0. Otherwise, the real
part of one root must be positive. In Appendix A, we give a rigorous proof
of the following result:
• the eigenvalues have nonpositive real part for the CFM only in the
extreme case kd = 0. Then, the control system ignores the distance to
the leading car and uses only the relative speed of the leading car.

However, this special control stratagem is pretty dangerous—there is

no assurance that the trajectory of a car will not “cross over” those of
the cars ahead and behind (after all, no attention is paid to the relative
positions). Thus, this method cannot be used. Even if kd is small, for
example, kv /kd = 100, some eigenvalues will have positive real part and
traffic instabilities will occur.
8 L. Wang et al.

For BCM, the eigenvectors are also of the form U (ω) = u(ω) ⊗ p(ω),
however, the eigenvalues are found by solving
1
λ(ω) p(ω) = M p(ω) + (e− jω + e jω )N p(ω)
2
= (M + N cos(ω)) p(ω).

Similarly, λ(ω) and p(ω) are the eigenvalue and eigenvector of the 2 × 2
matrix M + N cos(ω). The corresponding characteristic equation is

λ2 (ω) + (1 − cos(ω))kv λ(ω) + (1 − cos(ω))kd = 0. (13)

The sum of these two eigenvalues is −(1 − cos(ω))kv ≤ 0, and the product
of these two eigenvalues is (1 − cos(ω))kd ≥ 0. Thus, the real parts of both
eigenvalues are smaller than or equal to zero. The only case where the
eigenvalue is zero is when ω = 0, which corresponds to the equilibrium
state where the cars are spaced equally and are moving at the same speed.
The car-state mode corresponding to other eigenvectors U (ω) (with ω = 0)
will decay to zero. Thus, the traffic system goes to the equilibrium state
from an arbitrary initial state when using bilateral control. That is, bilateral
control can suppress traffic flow instabilities.

5. Boundary conditions

In Section 4, we have given an eigenvalue-and-eigenvector analysis of both

car-following and BCMs, and have shown the bilateral control’s advantage
of suppressing traffic flow instabilities. For easy analysis, we did not yet
consider boundary conditions. In the above, both matrices A and B were
doubly infinite. We should also consider the case when the total number of
cars is finite, and this means that we need to consider boundary conditions
on the line of traffic. Next, we analyze both models under various boundary
conditions.

5.1. Circular boundary condition

Suppose that there are totally K cars moving on a circle. That is, car K is
immediately ahead of car 1. Now, both A and B become finite (2K × 2K )
block circulant matrices:
⎛ ⎞ ⎛ ⎞
M N M N /2 N /2
⎜N M ⎟ ⎜ .. ⎟
⎜ ⎟ ⎜ N /2 M . ⎟
A=⎜ .. .. ⎟ , B = ⎜ ⎟.
⎝ . . ⎠ ⎝ ..
.
..
. N /2 ⎠
N M N /2 N /2 M
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 9

0.25 0.05
max Real(λ (k ,k ))
d v

0.2

kv: from 0 to 0.3

0.1
k

0.15
0.15
0.1
k

0.05 0.2

0
0.3 0.25

0.2
0.25 0.3
0.1 0.2
0.15 0 0.05 0.1 0.15 0.2 0.25 0.3
0.1
k : from 0 to 0.3 0 0.05
v
0 kd: from 0 to 0.3 k : from 0 to 0.3
d

(a) Small system ( K = 10) (b) Support set of (a)

Figure 3. The maximum real part of the eigenvalues for the car-following model, denoted
by λmax = maxk {Real(λk (kd , kv ))}, under the circular boundary condition. Both kd and kv
are in the range from 0 to 0.3. Even when K is small enough, for example, K = 10, the
region such that λmax = 0 is still pretty small (the “white region” in (b)). The choice of kd
in the “white region” of (b) is too close to zero to be useful in real applications.

The Fourier-transform approach used in Section

4 still works, only now we
use the DFT. That is, ω is sampled as 0, ω, 2ω, · · · , kω, · · · , (K −
T
1)ω —K equally spaced points with K ω = 2π . Now, u(ω) becomes a
finite vector u k with K entries:

T
u k = 1, wk , wk2 , · · · , wkl , wkl+1 , · · · , wkK −1 ,

where wk = e−2π jk/K (for k = 0, 1, 2, · · · , K − 1). The eigenvector Uk is

the Kronecker tensor product of u k and a 2 × 1 vector pk = (ak , bk )T :

U k = u k ⊗ pk

T
= ak , bk , wk ak , wk bk , · · · , wkl ak , wkl bk , · · · , wkK −1 ak , wkK −1 bk . (14)

For CFM, the eigenvalue equation λk pk = (M + wk N ) pk leads to

λ2k + (1 − wk )kv λk + (1 − wk )kd = 0. (15)

In Appendix A, we will show that the condition for all {λk } ≤ 0 is

 
kv2 1 1 K2
≥ − 1 ≈ .
kd 2 sin2 (π/K ) 2π 2
In practice it is not possible to satisfy this condition. For instance, in a
small traffic system containing just 10 cars, if kv = 0.2, then kd must be
less than 0.008 (which is too small to keep safe distance). Figure 3 shows a
simulation by MATLAB.
10 L. Wang et al.

In bilateral control, λk and pk are the eigenvalue and eigenvector of the

2 × 2 matrix M + (1 − cos(2π k/K ))N . The eigenvalue λk comes from

λ2k + (1 − cos(2π k/K ))kv λk + (1 − cos(2π k/K ))kd = 0. (16)

Still all λk ≤ 0, with λk = 0 only when k = 0 (corresponding to the

equilibrium state). The traffic goes to the equilibrium state from arbitrary
initial state when using bilateral control.

5.2. Other boundary conditions for the BCM

For other boundary conditions, for example, fixed-fixed, free-free, and
fixed-free, the eigenvalue decomposition analysis still works to analyze
the BCM under these three boundary conditions. Suppose that there are
K + 2 cars in the traffic system. The state of the first car, that is,
{y0 , ẏ0 }, and the state of the last car, that is, {y K +1 , ẏ K +1 }, provide the
boundary condition. The states of the other K cars form the state vector
Y = (y1 , ẏ1 , y2 , ẏ2 , · · · , yi , ẏi , · · · , y K , ẏ K )T . The form of B depends on the
particular type of boundary conditions.

5.2.1. Fixed-fixed boundary. Suppose that car 0 and car K + 1 are

moving at the (same) constant speed (or both fixed in the relative reference
system), that is, y0 = y K +1 = 0 and ẏ0 = ẏ K +1 = 0. Then, the matrix B is
⎛ ⎞
M N /2
⎜ .. ⎟
⎜ N /2 M . ⎟
B=⎜ ⎟.
⎝ . .. . . . N /2 ⎠
N /2 M
The vector u k is exactly the k-th basis vector of the K -point Discrete Sine
Transform (DST):
u k = (sin (kh) , sin (2kh) , · · · , sin (K kh))T ,
with k = 1, 2, · · · , K and h = π/(K + 1). The eigenvector Uk = u k ⊗ pk is
the Kronecker tensor product of u k and a 2 × 1 vector pk = (ak , bk )T . The
eigenvalue equation λk pk = (M + N cos(kh)) pk leads to

λ2k + (1 − cos (kh)) kv λk + (1 − cos (kh)) kd = 0. (17)

Now, all the eigenvalues are with negative real part. Thus, the final state is
Y(∞) = 0, that is, a special case of the equilibrium state. Note that the first
and last cars are fixed, thus, the total length of the traffic system does not
change. The final speed of the traffic system is the same as the speed of the
first car (and the last car).
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 11

5.2.2. Free-free boundary. Suppose that the state of the boundary car 0
is the same as the state of the car 1 behind (in the relative reference),
that is, y0 = y1 and ẏ0 = ẏ1 , moreover, the state of the boundary car K + 1
is also the same as the state of the car K ahead, that is, y K +1 = y K and
ẏ K +1 = ẏ K . Note that the states of the boundaries (cars 0 and car K + 1)
are changing to match the states of their neighbors, rather than fixed. Now,
the matrix B is
⎛ ⎞
M1 N /2
⎜ .. ⎟
⎜ N /2 M . ⎟
B=⎜ ⎟,
⎝ ..
.
..
. N /2 ⎠
N /2 M1

where
 
0 1
M1 = .
−kd /2 −kv /2

The vector u k is in the K -point Discrete Cosine Transform (DCT) basis:

u k = (cos ((1 − 1/2)kh) , cos ((2 − 1/2)kh) , · · · , cos ((K − 1/2)kh))T ,

with k = 0, 1, 2, · · · , K − 1 and h = π/K . The eigenvector Uk = u k ⊗ pk is

the Kronecker tensor product of u k and a 2 × 1 vector pk = (ak , bk )T . The
eigenvalue equation λk pk = (M + N cos(kh)) pk leads to

λ2k + (1 − cos (kh)) kv λk + (1 − cos (kh)) kd = 0. (18)

The real part of λk is less than zero for k = 1, · · · , K − 1. The only case
that λk = 0 is k = 0. The corresponding state is Y(t) = (a0 + b0 t, b0 , a0 +
b0 t, b0 , · · · , a0 + b0 t, b0 )T , when all the relative positions (and also relative
speeds) are shifted by the same amount. We can choose new relative
position and speed such that Y(t) = 0. Thus, free-free boundaries model
the traffic system from one equilibrium state to another equilibrium state
by bilateral control, during which the relative speed and the space between
the successive cars can be adjusted to match the demanded values. This
illustrates the flexibility of BCM.

5.2.3. Fixed-free boundary. Another interesting boundary condition is

fixed-free. Car 0 (one boundary) is moving at a constant speed, that is,
y0 = ẏ0 = 0. The state of car K + 1 (the other boundary) matches the state
12 L. Wang et al.

of its neighboring car K , that is, y K +1 = y K and ẏ K +1 = ẏ K . Now, the

matrix B is
⎛ ⎞
M N /2
⎜ .. ⎟
⎜ N /2 M . ⎟
B=⎜ ⎟.
⎝ ..
.
..
. N /2 ⎠
N /2 M1

And the vector u k is equally spaced sampling K points of sin((2k − 1)ω) in

the range of 0 < ω < π/2

u k = (sin ((k − 1/2)h) , sin ((k − 1/2)2h) , · · · , sin ((k − 1/2)K h))T ,

with k = 1, 2, · · · , K and h = π/(K + 1/2). The eigenvector Uk = u k ⊗ pk

is the Kronecker tensor product of u k and a 2 × 1 vector pk = (ak , bk )T .
The eigenvalue equation λk pk = (M + N cos((k − 1/2)h)) pk leads to

λ2k + (1 − cos ((k − 1/2)h)) kv λk + (1 − cos ((k − 1/2)h)) kd = 0. (19)

Now, the real part of λk is negative for all k = 1, 2, · · · , K . Thus, the final
state is Y(∞) = 0. All the cars are equally spaced and move at the same
speed. The state of the first car is fixed (used as the reference). The last car’s
state matches its leading car (i.e., what the CFM tries to implement). The
final speed of the traffic system is the same as the speed of the first car in
this case. Moreover, the total length of the traffic system can be changing
during the process of going to the equilibrium state (which is different from
the case of the fixed-fixed boundaries)
Figure 4 shows the numerical results of the real part of the eigenvalues
{λk } for the BCM under the various four boundary conditions: circular,
fixed-fixed, free-free, and fixed-free. The parameters are set as kd = kv =
0.2 and K = 100. All the eigenvalues are with nonpositive real part.4
Comparing Eq. (13) with Eqs. (16), (17), (18), and (19), respectively, we
can see that the eigenvalues of the BCM under different boundary conditions
(circular, fixed-fixed, free-free, and fixed-free) are exactly the equally spaced
sampling results of the (continuous) eigenvalues in the infinite boundary
condition (i.e., it is not necessary to consider the boundary condition), with
the step size ω chosen as 2π/K , π/(K + 1), π/K , and π/(K + 1/2),
respectively. The BCM is stable under the infinity boundary condition, thus,
it will also be stable under these four different finite boundary conditions.

4 The eigenvalues calculated by MATLAB command “eig( )” are sorted on their norm. Then, we
resort the eigenvalues by their real part.
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 13

0 0

−0.05 −0.05

−0.1 −0.1

−0.15 −0.15

−0.2 −0.2
0 50 100 150 200 0 50 100 150 200

(a) Circular boundary condition (b) Fixed-fixed boundary condition

0 0

−0.05 −0.05

−0.1 −0.1

−0.15 −0.15

−0.2 −0.2
0 50 100 150 200 0 50 100 150 200

(c) Free-free boundary condition (d) Fixed-free boundary condition

Figure 4. The real-part of the eigenvalues of the bilateral control model under various
boundary conditions. The parameters are chosen as kd = kv = 0.2 and K = 100. All the
eigenvalues have nonpositive real part. The labels under the horizontal axis are the indices
of the eigenvalues.

5.3. Other boundary conditions for the CFM

From the analysis in Section 5.2, we can see that there are two steps to find
the eigenvalues and eigenvectors of the “big matrix” A and B:
1. Find the eigenvector u k of the matrix generated by replacing the
“blocks” M and N in the “big matrix” with two scalars.
2. Eigenvector Uk is the Kronecker tensor product of u k and a 2 × 1
vector pk , which is an eigenvector of a 2 × 2 matrix.

In the case of BCM, if the “blocks” M and N are treated as two scalars,
then B will become a symmetric matrix for all of the three special boundary
conditions (fixed-fixed, free-free, and fixed-free). Thus, B is diagonalizable
[18], that is, there are enough eigenvectors to span the whole space.
However, for the CFM, the eigenvalue decomposition analysis does
not work for other boundary conditions. The “big matrix” A cannot be
diagonalized when M and N are treated as scalars. For instance, in the fixed
boundary condition,5 the “big matrix” A is
⎛ ⎞
M
⎜N M ⎟
⎜ ⎟
A=⎜ ⎜
N M ⎟.
⎟
⎝ .. ..
. . ⎠
N M

5 Note that the boundary condition for the CFM is just the state of the first car, rather than the
states of both of the first and last cars (as used in the BCM).
14 L. Wang et al.

If M and N are treated as scalars, then the matrix A will be similar to a

Jordan block. There are two matrices D and J :
⎛ ⎞ ⎛ ⎞
1 M
⎜ N ⎟ ⎜1 M ⎟
⎜ ⎟ ⎜ ⎟
⎜ N 2 ⎟ ⎜ 1 M ⎟
D=⎜ ⎟, J = ⎜ ⎟
⎜ . ⎟ ⎝ . . ⎠
⎝ .. ⎠ .. ..
K −1 1 M
N
such that
D −1 AD = J. (20)
Note that J is the transpose of a K × K Jordan block. Thus, there are K
repeated eigenvalues λ0 = M, but only one eigenvector
u 0 = (0, 0, · · · , 0, 0, 1)T .
In this case,6 the states of the cars can be calculated by
Y(t) = etA Y(0) = D −1 et J DY(0), (21)
where et J = etλ0 L, and L is a lower triangular Toeplitz matrix:
⎛ ⎞
1
⎜ t 1 ⎟
⎜ ⎟
⎜ t2 ⎟
⎜ ⎟
⎜ t 1 ⎟
⎜ 23 ⎟
⎜ 2 ⎟
L=⎜ t t ⎟. (22)
⎜ t 1 ⎟
⎜ 3! 2 ⎟
⎜ .. .. .. .. .. ⎟
⎜ . . . . . ⎟
⎜ ⎟
⎝ t K −1 t3 t2 ⎠
··· t 1
(K − 1)! 3! 2
The entries in the first column of L are exactly the first K terms of the
Taylor expansion of et . If λ0 = M is negative, then all the entries in the
matrix et J (and the sum of each row) go to zero when t is sufficiently large.
However, when K is infinite, the sum of each row of the matrix et J becomes
et(λ0 +1) , which could go to infinity even if λ0 is negative. Moreover, when K
− jω
is infinite, the eigenvalue of the matrix et J becomes et(λ0 +e ) . The stability
depends on the value of λ0 + e− jω , rather than λ0 itself.
When M and N are both 2 × 2 matrices, the two eigenvalues λ1 and λ2
of M can be found from
M pk = λk M, (k = 1, 2). (23)

6M is treated as a scalar temporarily, we will give the solution when M is a 2 × 2 matrix later.
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 15

The corresponding characteristic equation is

λ2k + kv λk + kd = 0. (24)
The two roots λ1 and λ2 both have negative real parts. Note that λ1 and
λ2 are the two (K repeated) eigenvalues of the “big matrix” A, and the
corresponding two eigenvectors are Uk = u 0 ⊗ pk (with k = 1, 2). However,
as mentioned above, having negative real part for both λ1 and λ2 does not
imply that the traffic system is stable. While the size of the matrix A is
2K × 2K . Eigenvalue decomposition analysis does not work here. Actually,
A is similar to a block-Jordan Form J, that is, A = EJE−1 , where
 
J1
J= ,
J2
and J1 and J2 are both K × K Jordan blocks:
⎛ ⎞ ⎛ ⎞
λ1 λ2
⎜ 1 λ1 ⎟ ⎜1 λ2 ⎟
⎜ ⎟ ⎜ ⎟
J1 = ⎜ 1 λ1 ⎟ , J2 = ⎜ 1 λ2 ⎟.
⎜ ⎟ ⎜ ⎟
⎝ .. ..
. . ⎠ ⎝ ..
.
..
. ⎠
1 λ1 1 λ2
In Appendix B, we give the proof and the detailed form of the 2K × 2K
invertible matrix E. The analytical solution of the CFM under the fixed
boundary condition is:
 tJ 
tJ −1 e 1
Y(t) = e Y(0) = Ee E Y(0) = E
tA
E−1 Y(0), (25)
et J2
where et Ji = etλi L (for i = 1, 2), and L is the lower triangular matrix in
(22). Even though both λ1 and λ2 have negative real parts, these matrices
grow large in finite time before decaying to zero (see Fig. 6). Traffic
instability will still occur.
Figure 5 shows the real parts of the eigenvalues {λk } of the CFM under
the circular and fixed boundaries (by MATLAB). The parameters are set
as kd = kv = 0.2 and K = 100. Note that, although there are theoretically
only 2 eigenvalues in the fixed boundary case, the numerical algorithm gives
some “eigenvalues” with positive real part.7
Figure 6 shows the norm of etA —the maximum singular value of
tA
e —for circular and fixed boundaries. Note that the vertical axis uses the

7 Note that the numerical algorithm is not guaranteed to find the correct eigenvalues. Nevertheless,
the numerical solutions for some of the eigenvalues predict the performance of the system. Some
“eigenvalues” estimated by the numerical algorithm have positive real parts (as shown in Fig. 5b).
Thus, the corresponding components of a perturbation will be amplified very quickly when they
are multiplied by the matrix A for finite times. Thus, car-following control with fixed boundary
conditions will also lead to traffic jams (or result in collisions).
16 L. Wang et al.

0.2 0.2

0.1 0.1

0 0

−0.1 −0.1

−0.2 −0.2

0 50 100 150 200 0 50 100 150 200

(a) Circular boundary condition (b) Fixed boundary condition

Figure 5. The numerical results of the real-part of the eigenvalues of the car-following
model under various boundary conditions (kd = kv = 0.2 and K = 100). Some of the
“eigenvalues” (obtained by the numerical algorithm) have positive real parts.

100
10
circular boundary
80
10 fixed boundary
||exp(t A)||

60
10

40
10

20
10

0
10
0 200 400 600 800 1000 1200 1400
Time t: from 0 to 1400 seconds

Figure 6. The norm of etA in the circular boundary (dash line) and fixed boundary (solid
curve). In the circular boundary, ||etA || increases exponentially with time t. In the fixed
boundary, ||etA || increases approximately exponentially at the beginning, and decays only
when t is large enough.

logarithmic coordinates. In the circular boundary condition, ||etA || increases

exponentially with time t. In the fixed boundary condition, ||etA || increases
approximately exponentially when t is not large, for example, t < 250, and
decays only when t is large enough, for example, t > 900. In the fixed
boundary condition, ||etA || will decay finally because both of the eigenvalues
have negative real part, However, ||etA || becomes very large (more than
1020 ) before decaying. Thus, traffic jams will still occur.
Here, we should mention that a modification of CFM that is closer to
human driver’s actual behavior is known as “constant time headway” control,
in which the safe distance s is chosen adaptively according to the car’s
speed, that is, si = vi T (with constant T ), rather than a fixed headway
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 17

distance s. The eigenvalue-eigenvector analysis used above can be used to

analyze this new model as well. The details are summarized in Appendix C.
Comparing this to the simple CFM in (1), the stability condition (A5) is
relaxed as shown in (C4) when the adaptive safe distance is used. However,
condition (C4) provides a critical limitation on the traffic throughput (see
Eq. (C5)). It corresponds to our experience that at low densities there may
be few problems, but that at high densities flow instabilities (“phantom
traffic jams”) will occur.

6. Simulation results

In this paper, we use the forward Euler method to do the numerical

simulations.8 The corresponding finite difference schemes are
(n) (n−1) (n−1)
CFM: Yk = (I + Mt)Yk + tNY k−1 , (26)

(n) (n−1) 1  (n−1) (n−1)


BCM: Yk = (I + Mt)Yk + t N Yk−1 + Yk+1 , (27)
2
(n)
where Yk is the state of cars k at time nt and I is a 2 × 2 identity
(n)
matrix. For small t (e.g., t = 0.1 second), the discrete simulation Yk
approaches the result of the continuous function Yk (nt). If the initial states
(0)
of all the cars are all zeros (i.e., Yk = 0 for all k), and the state of the
(n)
boundary (car 0) is also kept on as zeros (i.e., Y0 = 0 for all n), then the
states of all the cars will be all zeros all the time. That is, the traffic system
will maintain the equilibrium state shown Fig. 7. All the cars are spaced
equally and move at the same speed.
If there is a small perturbation in the initial states Y(0) , the car-following
system will become unstable quickly. Suppose that all the cars are spaced
by 30 m and moving at 90 km/h (i.e., 25 m/s) at the beginning. Let
kd = 0.2 (1/second), kv = 0.2 (1/second2 ), and K = 50. Figure 8 shows the
simulation result with small perturbations of the space between the cars,
that is, a random number in the range ±1.5 m. The car-following system
amplifies the small errors, and the traffic jam happens quickly. Moreover, the
CFM cannot avoid collision. Bilateral control can suppress traffic instability
effectively. The traffic system quickly becomes stable again, even if bilateral

8 Some more complicated schemes, for example, the fourth-order Runge-Kutta method, can also be
used to solve the ODEs in (8) with much higher order accuracy. However, more memory space is
needed, and the computational cost is much higher. In real application, we want the implementation
of the control system to be as simple as possible. Thus, in this paper, we use the forward Euler
method for the numerical simulations.
18 L. Wang et al.

50 50

45 45

40 40

35 35
time (second)

time (second)
30 30

25 25

20 20

15 15

10 10

5 5

0 10 20 30 40 50 0 10 20 30 40 50
car index: from 0 to 50 car index: from 0 to 50
(a) yk (t) (k = 1, 2, · · · , 50) (b) ẏk (t) (k = 1, 2, · · · , 50)

Figure 7. If the traffic is stable, then all the cars will be spaced equally and move at
the same speed. The relative position yk (t) and relative speed ẏk (t) doesn’t vary with time.
Thus, yk (t) and ẏk (t) are all vertical lines. The bold lines are the state of the first car (i.e.,
boundary).

control is turned on only just before the traffic chaos caused by the
car-following system sets in.
Figure 9 shows the simulation result with small perturbations of the
relative speed of the cars, that is, a random number in the range ±1 m/s.
The CFM becomes unstable quickly and the traffic jam happens. Bilateral
control suppresses traffic instability again. The BCM is used just before the
jam caused by the CFM, and then the traffic system quickly becomes stable
again.
Even if the initial condition is the equilibrium state. Under car following
control, the tiny small oscillation in the state of the first car (boundary)
is amplified in the state of the second car, and then is amplified again
in the state of the third car, and so on. Figure 10 shows the simulation
result with small perturbations of the speed of the boundary car 0, that
is, a random number between ±1 m/s. In the beginning, yk (0) = ẏk (0) = 0
for all k = 1, 2, · · · , 50. The car-traffic system under car-following control
becomes unstable quickly. If BCM is used in time, traffic instability can be
suppressed efficiently.

7. Conclusion

Bilateral control can suppress traffic instabilities. In this paper, we provide a

theoretical analysis from the viewpoint of the eigenvalues and eigenvectors
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 19

50 50
45 45
40 40

35 35
time (second)

time (second)
30 30

25 25

20 20

15 15

10 10

5 5

0 0
0 10 20 30 40 50 0 10 20 30 40 50
car index: from 0 to 50 car index: from 0 to 50
(a) {yk (t)} (car-following) (b) {ẏk (t)} (car-following)

70 70

60 60

50 50
time (second)

time (second)

40 40

30 30

20 20

10 10

0 0
0 10 20 30 40 50 0 10 20 30 40 50
car index: from 0 to 50 car index: from 0 to 50
(c) {yk (t)} (bilateral control) (d) {ẏk (t)} (bilateral control)

Figure 8. If there is small perturbations of the space between cars, then the car-following
model (CFM) will become unstable quickly (car collisions start at about 25 seconds). If
bilateral control is used before collision, then the car-traffic system quickly becomes stable
again. The first 20 seconds results in (c) and (d) are generated by CFM, then bilateral
control (fixed-free boundaries) is used promptly. The state of the first car (boundary) is set
to be zero.

of the corresponding linear system of ODEs. We show that the traditional

car-following system is unstable because some eigenvalues of the “big
matrix” A, corresponding to its ODE system, have positive real parts. In
contrast, the eigenvalues of the “big matrix” B corresponding to the ODE
20 L. Wang et al.

50 50

45 45

40 40

35 35

time (second)
time (second)

30 30

25 25

20 20

15 15

10 10

5 5

0 0
0 10 20 30 40 50 0 10 20 30 40 50
car index: from 0 to 50 car index: from 0 to 50
(a) yk (car-following) (b) ẏk (t) (car-following)

70 70

60 60

50 50
time (second)
time (second)

40 40

30 30

20 20

10 10

0 0
0 10 20 30 40 50 0 10 20 30 40 50
car index: from 0 to 50 car index: from 0 to 50
(c) yk (bilateral control) (d) ẏk (t) (bilateral control)

Figure 9. If there is small perturbation of the speed of the cars, then the car-following
model (CFM) will be unstable, and collision happens quickly (at about 22 seconds). If
bilateral control is used before collision, then the car-traffic system becomes stable quickly.
The first 18 seconds results in (c) and (d) are generated by CFM, then bilateral control
(fixed-free boundaries) is used promptly. The state of the first car (boundary) is set to be
zero.

system for the BCM all have nonpositive real part, thus, the traffic system
will become stable from arbitrary initial state when using bilateral control.
We also analyze the CFM and BCM under various boundary conditions.
“Don’t tailgate.” That is, do not come closer to the car in front of you
than the distance between you and the car behind you. This may be the big
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 21

70 70

60 60

50 50
time (second)

time (second)
40 40

30 30

20 20

10 10

0 0
0 10 20 30 40 50 0 10 20 30 40 50
car index: from 0 to 50 car index: from 0 to 50
(a) yk (car-following) (b) ẏk (t) (car-following)

90 90

80 80

70 70

60 60
time (second)
time (second)

50 50

40 40

30 30

20 20

10 10

0 0
0 500 1000 1500 0 10 20 30 40 50
car index: from 0 to 50 car index: from 0 to 50
(c) yk (bilateral control) (d) ẏk (t) (bilateral control)

Figure 10. Under car following control, the tiny small oscillation in the state of the first
(boundary) car is amplified in the state of the second car, and then is amplified again in
the state of the third car, and so on. The collision happens quickly (at about 40 seconds).
If BCM is used before collision, for example, at 32 seconds, then the car-traffic system
becomes stable quickly. The initial condition is the equilibrium state.

idea behind bilateral control. It is good for the community of drivers and
provides advantages, such as freedom from traffic flow instabilities, which
can be proven mathematically.
In this paper, we only considered the simplest models—the linear
approximation of both the CFM and BCM. In real application, the speed of
22 L. Wang et al.

the car is limited to a range from vmin to vmax , and the possible acceleration
(and deceleration) of the car is also limited to some range from amin
to amax . Moreover, kd and kv need not be constant, but can be chosen
to be some functions of the relative distances and relative speeds. For
example, the feedback gains kd and kv can be made inversely proportional
to the distances and speed difference so as to provide stiffer control when
neighboring cars are near. This appears to be what drivers actually do,
mostly to avoid collisions in the car following situation.
In contrast, even the simplest linearized BCM—with both kd and kv
constants—can suppress traffic instabilities effectively. However, the BCM
can be improved beyond that simple linear version. For instance, both kd and
kv in the BCM can be made to be functions of the states of the neighboring
cars to increase the safety margin, make the system more robust to sensor
errors, increase fuel efficiency, and more aggressively damp traffic flow
instabilities. Moreover, communication between cars can be used to further
improve the safety margin and robustness of the system [15]. We plan to
work on these aspects in future.

Acknowledgment

This work is sponsored by Toyota Motor Corporation (LP-C000765-SR).

Appendix A: Stable Condition of the CFM

The two roots of Eq. (12) are9

ω  ω ω
− jω/2
λ(ω) = − jkv e sin ± −e− jω sin2 kv2 − 2 jkd e− jω/2 sin .
2 2 2
First, The real part of the sum of the two roots is smaller than zero,
  ω  ω
c(ω) = − jkv e− jω/2 sin = −kv sin2 < 0. (A1)
2 2
Thus, one of the λ(ω) must be with nonpositive real part. Furthermore, we
can calculate
 ω  ω 
R(ω) = −e− jω sin2 kv2 − 2 jkd e− jω/2 sin
2 2
     
2 ω 2 ω
= − sin 1 − 2 sin kv + 2kd ,
2
(A2)
2 2

9 Inthis paper, the square root of a complex number is chosen as the one with positive real part. (In
our case, the real part of the square root is not zero.)
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 23

 ω  ω 
I (ω) =  −e− jω sin2 kv2 − 2 jkd e− jω/2 sin
2 2
ω ω    
2 ω
= 2 cos sin sin k − kd ,
2
(A3)
2 2 2 v
where { f } and { f } denote the real and imaginary part of the complex
number f . Then, we can calculate
 
d(ω) = R(ω) + j I (ω)

1
= ±√ R(ω) + R 2 (ω) + I 2 (ω).
2
For easy analysis, let p = sin2 (ω/2). The condition of nonpositive real part
of λ(ω) is that the inequality d 2 (ω) ≤ c2 (ω) holds for all 0 ≤ p ≤ 1, that
is,

I 2 (ω) ≤ 4c4 (ω) − 4c2 (ω)R(ω) (A4)

By tedious calculation, we can obtain the equivalent expression of Eq.(A4).

That is, for all 0 ≤ p ≤ 1
 
1
p p−
2  ≥ 0. (A5)
2 kv /kd + 1

Equation (A5) implies that kv2 /kd → ∞. Note that kv is a finite number (in
general, kv < 1). Thus, the condition kv2 /kd → ∞ means kd = 0. That is,
only the car’s speed is used by the control system, and finally all cars move
at the same speed. (Note that no guarantee of avoiding collision during this
process.)
In the case of circular boundary condition, ω is sampled as K points
{2π k/K }, and p = sin2 (ω/2) is also sampled as K points:
pk = sin2 (kπ/K ) , (k = 0, 1, 2, · · · , K − 1).
The corresponding equivalent condition of “non-positive real part for all
eigenvalues” is that Eq. (A5) holds for p chosen as all of the K sampled
points { pk }. Note that p0 = 0, the secondary smallest number is p1 . Thus,
all sample points { pk } will satisfy Eq. (A5) if and only if p1 satisfies Eq.
(A5). Now, we obtain the following condition for the stabilization of the
“car-following” model:
1

 ≤ sin2 (π/K ), (A6)
2 kv2 /kd + 1
24 L. Wang et al.

that is,
 
kv2 1 1 K2
≥ − 1 ≈ . (A7)
kd 2 sin2 (π/K ) 2π 2
As K increases, this condition will be more and more difficult to satisfy.

Appendix B: Similarity to the Block Jordan Form

For the CFM with fixed boundary condition, the “big matrix” A is
⎛ ⎞
M
⎜N M ⎟
⎜ ⎟
A=⎜ ⎜
N M ⎟,
⎟
⎝ .. ..
. . ⎠
N M
where M and N are 2 × 2 matrices. We want to diagonalize both M
and N by a similarity transform. Start with the eigenvalue-eigenvector
decomposition of the 2 × 2 matrix M:
 
−1 λ1
M = PP = P P −1 .
λ2
Then, all blocks M in A are changed to  by a simple similarity, and all
blocks N are changed to A = P −1 N P:
S−1 AS = F, (B1)
where
⎛ ⎞ ⎛ ⎞
P 
⎜ P ⎟ ⎜A  ⎟
⎜ ⎟ ⎜ ⎟
S=⎜ P ⎟, F=⎜ A  ⎟
⎜ ⎟ ⎜ ⎟
⎝ ..
. ⎠ ⎝ ..
.
..
. ⎠
P A 
and
   
λ1 −1 a11 a12
= , A=P NP = .
λ2 a21 a22
Now we want to change the 2 × 2 submatrices A to diagonal matrices D.
This is achieved by a second similarity transform of the big matrix F using
2 × 2 blocks C:
FW = WL, (B2)
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 25

where
⎛ ⎞ ⎛ ⎞
I 
⎜C I ⎟ ⎜D  ⎟
⎜ ⎟ ⎜ ⎟
W=⎜ C I ⎟, L=⎜ D  ⎟
⎜ ⎟ ⎜ ⎟
⎝ ..
.
..
. ⎠ ⎝ ..
.
..
. ⎠
C I D 

and
   
c11 c22 d1
C= , D= .
c21 c22 d2

Equation (B2) is equivalent to a Sylvester equation for C:

C + A = D + C, or C − C = A − D. (B3)

Solutions C and D are not unique, but they exist. In fact only c12 and c21
are determined by Eq. (B3)
   
0 −c12 a11 − d1 a12
(λ1 − λ2 ) = .
c21 0 a21 a22 − d2

The entries c11 and c22 are arbitrary and we choose c11 = c22 = 0. The
diagonal matrix D and the matrix C are
   
a11 1 0 −a12
D= and C = . (B4)
a22 λ1 − λ2 a21 0

Now Eq. (B2) holds. The diagonal blocks I lead to determinant of W equal
to 1, thus W is invertible and F is similar to L.
Then the combined effect of the similarities in Eq. (B1) and (B2) is to
produce diagonal blocks  and D to replace M and N in the big matrix:

(SW)−1 A(SW) = L. (B5)

The exponentials etA show the fast growth of perturbations in the car-
following system. Using this similarity, the same information is in etL —and
it is simpler to analyze because  and D are diagonal.
From the matrix L, it is a short step to the official Jordan form J. Then
etJ has the same growth information as etL . The Jordan form J has just 2
blocks with the eigenvalues λ1 and λ2 —each repeated K times with only
one eigenvector,
 
J1
J= ,
J2
26 L. Wang et al.

where J1 and J2 are both K × K Jordan blocks:

⎛ ⎞ ⎛ ⎞
λ1 λ2
⎜ 1 λ1 ⎟ ⎜1 λ2 ⎟
⎜ ⎟ ⎜ ⎟
J1 = ⎜ 1 λ1 ⎟ , J2 = ⎜ 1 λ2 ⎟.
⎜ ⎟ ⎜ ⎟
⎝ .. ..
. . ⎠ ⎝ ..
.
..
. ⎠
1 λ1 1 λ2

First, G comes from L by a simple permutation of the rows and the same
permutation of the columns:

• Rows (and columns) 1 to K of G come from rows (and columns)

1, 3, 5, · · · , 2K − 1 of L.
• Rows (and columns) K + 1 to 2K of G come from rows (and
columns) 2, 4, 6, · · · , 2K of L.

Thus, G is similar to L:

G = PLP−1 = PLPT . (B6)

The form of G is
 
G1
G= ,
G2

where G 1 and G 2 are both K × K matrices:

⎛ ⎞ ⎛ ⎞
λ1 λ2
⎜ d1 λ1 ⎟ ⎜ d2 λ2 ⎟
⎜ ⎟ ⎜ ⎟
G1 = ⎜ d1 λ1 ⎟ , G2 = ⎜ d2 λ2 ⎟
⎜ ⎟ ⎜ ⎟
⎝ .. ..
. . ⎠ ⎝ ..
.
..
. ⎠
d1 λ1 d2 λ2

and the form of the permutation matrix P is

 
P1
P= ,
P2

where P1 and P2 are both K × 2K matrices:

⎛ ⎞
1 0 0 0 0 0 0 ··· 0 0
⎜0 0 1 0 0 0 0 ··· 0 0⎟
⎜ ⎟
P1 = ⎜ 0 0 0 0 1 0 0 ··· 0 0⎟
⎜. .. .. .. .. .. .. .. .. .. ⎟
⎝ .. . . . . . . . . .⎠
0 0 0 0 0 0 0 ··· 1 0
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 27

⎛ ⎞
0 1 0 0 0 0 0 ··· 0 0
⎜0 0 0 1 0 0 0 ··· 0 0⎟
⎜ ⎟
P2 = ⎜ 0 0 0 0 0 1 0 ··· 0 0⎟.
⎜. . . . . . . .. .. .. ⎟
⎝ .. .. .. .. .. .. .. . . .⎠
0 0 0 0 0 0 0 ··· 0 1

This permutation matrix P is famous for its use in the fast Fourier transform
(FFT) algorithm [19].
A final similarity with a diagonal matrix D will replace d1 and d2 in G by
1—to produce the official Jordan form of the car-following matrix A. The
form of the diagonal matrix D is
 
D1
D= ,
D2

where D1 and D2 are both K × K diagonal matrices:

⎛ ⎞ ⎛ ⎞
d1 d2
⎜ d 2 ⎟ ⎜ d22 ⎟
⎜ 1 ⎟ ⎜ ⎟
⎜ d13 ⎟ ⎜ d23 ⎟
D1 = ⎜ ⎟ , D2 = ⎜ ⎟.
⎜ .. ⎟ ⎜ .. ⎟
⎝ . ⎠ ⎝ . ⎠
K
d1 d2K

We can check that

D−1 GD = J. (B7)

Again, our purpose is to construct a matrix J similar to the original A with

a convenient exponential:

A = EJE−1 (B8)

and E = SWPT D.

Appendix C: CFM with Adaptive Safe Distance

In simple CFM (1) — also known as the “constant headway” control,

the safe distance is a fixed number s. A more complicated model—which
is closer to human driver’s behavior—is known as “constant time headway”
control policy in leader following, in which the safe distance is chosen
adaptively according to the car’s speed:

si = vi T, (C1)
28 L. Wang et al.

where T is known as the response time, that is, the time taken by the
current car to react to sudden braking of the leading car.10 In general, T is
chosen to be about 1 second. Now, the simple CFM (1) becomes:
ai = kd (xi−1 − xi ) + kv (vi−1 − (1 + τ )vi ), (C2)
where τ = T kd /kv is positive. The eigenvalue-eigenvector decomposition
method still works for analyzing this new CFM (C2) with adaptive safe
distance. Only a small modification is needed. First, let s = 0 in (3), and
then let the 2 × 2 matrix M be
 
0 1
M= .
−kd −(1 + τ )kv
We can build the same “big ODE” system as in (8). The eigenvalues of the
block-Toeplitz matrix A are the roots of the following characteristic equation
(see Eq. (12)):
λ2 (ω) + (1 − w(ω) + τ )kv λ(ω) + (1 − w(ω))kd = 0, (C3)
where w(ω) = e− jω . The c(ω) in (A1) becomes
ω τ
c(ω) = −kv sin2 − kv < 0
2 2
and the R(ω) in (A2) and I (ω) in (A3) become
 ω    ω   τ2
R(ω) = − sin2 1 − τ − 2 sin2 kv2 + 2kd + kv2 ,
2 2 4
ω ω    τ 2
2 ω
I (ω) = 2 cos sin sin k − kd + kv ,
2
2 2 2 v 2
correspondingly. The stability condition (A5) gives the following con-
straint:


1 − (τ 2 /2 + τ ) kv2 /kd
p p−
2  ≥ 0, (C4)
2 kv /kd + 1 + τ kv2 /kd

where p = sin2 (ω/2) is in the range from 0 to 1. Thus, the stability

condition is


1 − (τ 2 /2 + τ ) kv2 /kd ≤ 0.
Or (by substituting τ = T kd /kv ),
kd T 2 + 2kv T − 2 ≥ 0.

10 A more detailed analysis takes into account possible differences in the speeds of the vehicles as
well.
 Eigenvalue and Eigenvector Analysis of Stability for a Line of Traffic 29

That is,

2kd + kv2 − kv
T ≥ . (C5)
kd
Comparing this with the original “constant headway” control (1)—whose
stability condition (A5) can not be satisfied—we note that the stability
condition (C4) of the “constant time headway” control (C2) can be satisfied
under certain conditions. However, the (equivalent) stability condition (C5)
is still difficult to satisfy in a real traffic system. For instance, when both
kd and kv are in the (reasonable) range from 0 to 0.2, the “response time”
T must be larger than 2.3166 seconds (that minimum for T occurs when
kd = 0.2 and kv = 0.2). We get a value close to 1 second for T only when
both kd and kv are larger than 0.66.
Note that the stability condition (C5) provides a critical limitation
 on the
traffic throughput.11 If the traffic throughput is larger than kd /( 2kd + kv2 −
kv ) (see Eq. (C5)), then traffic flow instabilities will arise (or collisions
will result). This corresponds to our experience that at low densities there
may be few problems, but that at high densities flow instabilities (“phantom
traffic jams”) will occur.

References

1. B. K. HORN, Suppressing traffic flow instabilities, 16th International IEEE Conference

on Intelligent Transportation Systems (ITSC 2013), IEEE, 2013 pp. 13–20.
2. H. GREENBERG, An analysis of traffic flow, Oper. Res. 7:79–85 (1959).
3. R. E. CHANDLER, R. HERMAN, and E. W. MONTROLL, Traffic dynamics: Studies in car
following, Oper. Res. 6:165–184 (1958).
4. R. HERMAN, E. W. MONTROLL, R. B. POTTS, and R. W. ROTHERY, Traffic dynamics:
Analysis of stability in car following, Oper. Res. 7:86–106 (1959).
5. L. C. EDIE, Car-following and steady-state theory for noncongested traffic, Oper. Res.
9:66–76 (1961).
6. P. I. RICHARDS, Shock waves on the highway, Oper. Res. 4:42–51 (1956).
7. T. NAGATANI, Traffic jams induced by fluctuation of a leading car, Phys. Rev. E
61:3534 (2000).
8. W. S. LEVINE and M. ATHANS, On the optimal error regulation of a string of moving
vehicles, IEEE Trans. Automat. Contr. 11:355–361 (1966).
9. D. N. GODBOLE and J. LYGEROS, Longitudinal control of the lead car of a platoon,
IEEE Trans. Veh. Technol. 43:1125–1135 (1994).
10. C. CHIEN, Y. ZHANG, and C. CHENG, Autonomous intelligent cruise control using both
front and back information for tight vehicle following maneuvers, American Control
Conference, Proceedings of the 1995, vol. 5 IEEE, 1995 pp. 3091–3095.

11 The traffic throughput is defined as v/(l + vT ) (which is less than 1/T ), where v is the car’s
speed and l is the car length.
30 L. Wang et al.

11. P. BAROOAH, P. G. MEHTA, and J. P. HESPANHA, Mistuning-based control design to

improve closed-loop stability margin of vehicular platoons, IEEE Trans. Automat.
Contr. 54:2100–2113 (2009).
12. F. LIN, M. FARDAD, and M. R. JOVANOVIC, Optimal control of vehicular formations
with nearest neighbor interactions, IEEE Trans. Automat. Contr. 57:2203–2218 (2012).
13. M. R. JOVANOVIC and B. BAMIEH, On the ill-posedness of certain vehicular platoon
control problems, IEEE Trans. Automat. Contr. 50:1307–1321 (2005).
14. B. BAMIEH, M. R. JOVANOVIC, P. MITRA, and S. PATTERSON, Coherence in large-scale
networks: Dimension-dependent limitations of local feedback, IEEE Trans. Automat.
Contr. 57:2235–2249 (2012).
15. T. A. BARAN and B. K. HORN, A robust signal-flow architecture for cooperative
vehicle density control., ICASSP, Proceedings of IEEE, Vancouver, British Columbia,
Canada. 2013 pp. 2790–2794.
16. A. NAKAYAMA, Y. SUGIYAMA, and K. HASEBE, Effect of looking at the car that follows
in an optimal velocity model of traffic flow, Phys. Rev. E 65:016112-1–016112-6
(2001).
17. M. TREIBER and D. HELBING, Hamilton-like statistics in one-dimensional driven
dissipative many-particle systems, Eur. Phys. J. B 68:607–618 (2009).
18. G. STRANG, Introduction to Linear Algebra, Wellesley-Cambridge Press, Wellesley,
MA, 2016.
19. G. STRANG, Computational Science and Engineering, Wellesley-Cambridge Press,
Wellesley, MA, 2007.
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
(Received May 18, 2016)