Maximum Likelihood Method for MIMO
Mobile-to-mobile Channel Parameter Estimation
Alenka G. Zajić and Gordon L. Stüber
School of Electrical and Computer Engineering
Georgia Institute of Technology, Atlanta, GA 30332 USA
Abstract— A three-dimensional reference model for wideband
multiple-input multiple-output mobile-to-mobile channels is re-
viewed. To allow comparison between the proposed model and
measured data, a new maximum likelihood based stochastic
estimator is derived. The proposed estimator extracts the relevant
model parameters from the measured data. The performance
of the new estimator is evaluated by deriving the Cramér-Rao
lower bound (CRLB) and by comparing the mean square error
of the parameter estimates to the CRLB. Simulation results
show that the proposed estimator has an asymptotically optimal
performance, since it reaches the CRLB for a small number of
samples.
I. I NTRODUCTION
Mobile-to-mobile (M-to-M) communications play an im-
portant role in mobile ad-hoc wireless networks, intelligent
transportation systems, and relay-based cellular networks. The
statistical properties of M-to-M channels are quite different
from conventional fixed-to-mobile (F-to-M) cellular land mo-
bile radio channels [1]. M-to-M communication systems are
equipped with low elevation antennas and have both the trans-
mitter and receiver in motion. Akki and Haber [1] proposed a
two-dimensional (2-D) reference model for single-input single-
output (SISO) M-to-M Rayleigh fading channels. Reference
models for 2-D narrowband multiple-input multiple-output
(MIMO) M-to-M channels have been proposed in [2], [3]. To
appropriately model an urban environment, we have recently
proposed the three-dimensional (3-D) reference model for
narrowband and wideband MIMO M-to-M multipath fading
channels [4], [5].
Realistic M-to-M channel models are necessary for success-
ful design of M-to-M systems. In order to validate channel
models, comparison between the channel models and measure-
ments, and consequently estimation of the model parameters,
is necessary. Deterministic parameter estimation techniques
employ deterministic channel models that have a large number
of discrete waves with unknown amplitudes, phases, and
frequencies (e.g., Space Alternating Generalized Expectation-
maximization (SAGE) algorithm) [6], [7]. Such an estimation
approach leads to maximization of highly non-linear likelihood
functions with many local optima, which causes convergence
0
Prepared through collaborative participation in the Collaborative Technology
Alliance for Communications & Networks sponsored by the U.S. Army
Research Laboratory under Cooperative Agreement DAAD19-01-2-0011. The
U.S. Government is authorized to reproduce and distribute reprints for
Government purposes notwithstanding any copyright notation thereon.
978-1-4244-2324-8/08/$25.00 © 2008 IEEE.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2008 proceedings.
problems [8]. On the other hand, stochastic channel param-
eter estimation techniques employ statistical channel models,
where the only unknowns are the parameters of the distribution
functions used to characterize the angles of departure and the
angles of arrival. The first stochastic estimator, which employs
the 2-D MIMO F-to-M channel model in [9] and estimates
the parameters in the von Mises distribution function, is
introduced in [10]. Compared to the deterministic parameter
estimation approach, this method leads to lower complexity
and faster convergence [11], [12].
In this paper, a new maximum likelihood based stochastic
estimator is derived to extract the parameters needed for our
3-D wideband MIMO M-to-M channel reference model in [5]
from measured data. The new estimator jointly estimates the
parameters of the distribution functions used to characterize
the azimuth and elevation angles of departure, the azimuth
and elevation angles of arrival, and the parameters that specify
how much the single- and double-bounced rays contribute
to the total averaged received power. This estimator is an
extension of the stochastic estimator in [10], which only
estimates the parameters of the distribution function used to
characterize the azimuth angles of arrival. The performance
of the new estimator is evaluated by deriving the Cramér-
Rao lower bound (CRLB) and by comparing the mean square
error of the parameter estimates to the CRLB. Simulation
results show that the proposed estimator has an asymptotically
optimal performance, since it reaches the CRLB for a small
number of samples.
The remainder of the paper is organized as follows. For ease
of reference, Section II reviews our 3-D wide-band MIMO M-
to-M reference model and its space-time-frequency correlation
function (STF-CF). Section III presents the new maximum
likelihood based stochastic estimator. Section IV derives the
CRLB for the proposed estimator. Section V compares the
mean square error of the parameter estimates to the CRLB
and evaluates the performance of the proposed estimator
by comparing the analytical and measured STF-CF. Finally,
Section VI provides some concluding remarks.
II. A 3-D T HEORETICAL M ODEL FOR W IDEBAND MIMO
M OBILE - TO -M OBILE C HANNELS
This section reviews our 3-D theoretical model for wideband
MIMO M-to-M multipath fading channels proposed in [5]. We
consider a wideband MIMO communication system with Lt
transmit and Lr receive omnidirectional antenna elements. It is
1
 (m,l) (n,k)
assumed that both the Tx and Rx are in motion and equipped imuth angles of departure and arrival, αT and αR , are
with low elevation antennas. The radio propagation occurs in characterized by the von Mises probability density function
outdoor metropolitan environments that are characterized by (pdf) [13]. The elevation angles of departure and arrival,
(m,l) (n,k)
3-D non-isotropic scattering with either line-of-sight (LoS) βT and βR , are characterized by the pdf [14] f (ϕ) =
or non-line-of-sight (NLoS) conditions between the Tx and π cos (πϕ/(4ϕm )) /(4|ϕm |) for | ϕ |≤| ϕm |≤ π/2 and
Rx . The MIMO channel can be described by an Lr × Lt f (ϕ) = 0 otherwise. The parameter ϕm is the maximum
matrix T(t, f ) = [Tij (t, f )]Lr ×Lt of the time-variant transfer elevation angle and lies in the range 0◦ ≤| ϕm |≤ 20◦ [5].
(l) (k)
functions. Finally, the radii Rt and Rr are characterized by the pdf
2 2
The 3-D reference model is derived using the 3-D geomet- f (R) = 2R/(R2 − R1 ) [5].
rical “concentric-cylinders” model shown in Fig. 1. Around The normalized STF-CF between two time-variant transfer
the Tx , M fixed omnidirectional scatterers occupy a volume functions Tpq (t, f ) and Tp̃q̃ (t, f ) can be written as [5]
between cylinders of radii Rt1 and Rt2 . It is assumed that the
M scatterers lie on L cylindric surfaces of radii Rt1 ≤ Rt ≤
(l) Rpq,p̃q̃ (Δt, Δf ) = Rpq,
SBT
p̃q̃ (Δt, Δf ) + Rpq,p̃q̃ (Δt, Δf )
SBR

th
Rt2 , where 1 ≤ l ≤ L. The l cylindric surface contains + DB
Rpq,p̃q̃ (Δt, Δf ) + Rpq,p̃q̃ (Δt, Δf ),(2)
LoS

M (l) fixed omnidirectional scatterers, and the (m, l)th transmit

p̃q̃ (Δt, Δf ), Rpq,p̃q̃ (Δt, Δf ), Rpq,p̃q̃ (Δt, Δf ), and
(m,l) SBT SBR DB
scatterer is denoted by ST , where 1 ≤ m ≤ M (l) . where Rpq,
p̃q̃ (Δt, Δf ) denote the normalized STF-CFs of the single-
LoS
Similarly, around the Rx , N fixed omnidirectional scatterers Rpq,
occupy a volume between cylinders of radii Rr1 and Rr2 . It bounced transmit, single-bounced receive, double-bounced,
is assumed that the N scatterers lie on P cylindric surfaces of and LoS components, respectively, and p, p̃ ∈ {1, . . . , Lt },
(k)
radii Rr1 ≤ Rr ≤ Rr2 , where 1 ≤ k ≤ P . The k th cylindric q, q̃ ∈ {1, . . . , Lr }. The expressions for Rpq, SBT
p̃q̃ (Δt, Δf ),
surface contains N (k) fixed omnidirectional scatterers, and Rpq,p̃q̃ (Δt, Δf ), Rpq,p̃q̃ (Δt, Δf ), and Rpq,p̃q̃ (Δt, Δf ) are
SBR DB LoS
(n,k) given in Appendix A.
the (n, k)th receive scatterer is denoted by SR , where
(k)
1 ≤ m ≤ N . The parameters in Fig. 1 are defined in Table I.
TABLE I
z z′ D EFINITION OF PARAMETERS IN F IGURE 1.

D The distance between the centers of the Tx and Rx cylinders.

Rt 2 Rr 2 Rt(l ) , Rr(k ) The radius of the l th Tx and k th Rx cylinder, respectively.
Rt(l ) Rr(k ) dT , d R The spacing between two adjacent antenna elements at the Tx and
Rt1 ST(m,l ) Rr1
Rx, respectively.
ε p, m,l ε m,l ,n,k The orientation of the Tx and Rx antenna array in the x-y plane
ε p , n, k θT , θ R
S R(n, k ) (relative to the x-axis), respectively.
AT( p ) ε p, q ε m,l ,q
AR(q ) ε n ,k ,q ψT , ψ R The elevation of the Tx’s and Rx’s antenna array relative to the x-y
ψT βT( m,l ) plane, respectively.
vT , v R The velocities of the Tx and Rx, respectively.
OT ψR β R( n,k )
~ γT , γ R The moving directions of the Tx and Rx, respectively.
AT( p ) OR The azimuth angles of departure (AAoD) of the waves that impinge
hp α T( m, l ) , α T( n, k )
(q~ ) on the scatterers S T(m, l ) and S R(n, k ) , respectively.
vT vR hq AR
α R( m, l ) , α R( n, k ) The azimuth angles of arrival (AAoA) of the waves scattered from
( n ,k ) S T(m, l ) and S R(n, k ) , respectively.
γT γ R αR
α RLoS The AAoA of the LoS paths.
OT ' π − θT π − α R(m ,l ) OR ' π − θR x
αT(n ,k ) π − α RLoS The elevation angles of departure (EAoD) and the elevation angles
β T( m, l ) , β R( n, k )
αT(m,l ) of arrival (EAoA), respectively.
D ε p , m , l , ε m, l , q , ( ) ( ) ( )
The distances d AT( p) , ST(m, l ) , d ST( m, l ) , AR( q ) , d AT( p) , S R(n, k ) ,

y
ε p , n, k , ε n , k , q ,
( ) d (S , A ), d (S , S ),
d ST( m) , S R( n) , ( n, k )
R
(q)
R
(m, l )
T
(n, k )
R and
d (A , A ) , respectively
ε m, l , n, k and ε pq ( p) ( q)
T R
Fig. 1. The “concentric-cylinders” model with LoS, single-bounced transmit, h p , hq The distances d (O , O ) and d (O , O ), respectively.
T
'
T R
'
R
single-bounced receive, and double-bounced rays for a MIMO M-to-M
channel with Lt = Lr = 2 antenna elements.

(p) (q) III. M AXIMUM L IKELIHOOD BASED PARAMETER

The time-variant transfer function of the link AT - AR is
defined as a superposition of the LoS, single-bounced transmit, E STIMATION
single-bounced receive, and double-bounced rays [5] The parametric nature of our wide-band channel model
Tpq (t, f ) = (1) makes it adaptable to a variety of propagation environments.
To verify our reference model with measured data, we need
Tpq (t, f ) + Tpq (t, f ) + Tpq (t, f ) + Tpq (t, f ),
LoS SBT SBR DB
to estimate the model parameters from the measurements.
where the single-bounced transmit, single-bounced receive, Hence, a new maximum likelihood based stochastic estimator
double-bounced, and LoS components of the time-variant is derived to extract the parameters needed for our reference
transfer function are defined in (3), (4), (8), (11) of [5], model from the measured data. The new estimator jointly
respectively. estimates the parameters of the distribution functions used
The model in [5] assumes that the angles of departure to characterize the azimuth and elevation angles of depar-
and the angles of arrival are random variables. The az- ture, the azimuth and elevation angles of arrival, and the

978-1-4244-2324-8/08/$25.00 © 2008 IEEE. 2

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2008 proceedings.
 parameters that specify how much the single- and double-
bounced rays contribute to the total averaged received power,
i.e., Θ = [βTm , kT , μT , βRm , kR , μR , ηT , ηR ]. This estimator
is an extension of the stochastic estimator in [10], which only
estimates the parameters of the distribution function used to
characterize the azimuth angles of arrival. In contrast to the
estimator in [10], which estimates its parameters from the
spatial correlation function, our estimator uses the STF-CF
to estimate the parameters.
By observing that both the theoretical and measured input
delay-spread functions are complex Gaussian random pro-
cesses, we can estimate the parameters Θ = [βTm , kT , μT ,
βRm , kR , μR , ηT , ηR ] using a maximum-likelihood (ML) esti-
mator with the log-likelihood function

L{ΘT,F } = −Ns ln π Lr Lt + ln |R(ΘT,F , T Δts , F Δfs )|
 × R(ΘT,F , T Δts , F Δfs )−1 DΘi (ΘT,F , T Δts , F Δfs )
−1
the Fisher
 information matrix can be written as [F(Θ)]i,j =
−E ∂ 2 L(Θ)/∂Θi ∂Θj , where i, j ∈ {1, . . . , 8}. To obtain
the elements of the Fisher information matrix, we calculate the
first and the second derivative of the log-likelihood function
in (3) as follows:
∂L(ΘT,F )
= (5)
∂Θi

− Ns tr ILt LR −R(ΘT,F , T Δts , F Δfs )−1 R̂(T Δts , F Δfs )
−1
× R(ΘT,F , T Δts , F Δfs ) DΘi (ΘT,F , T Δts , F Δfs ) ,
∂ 2 L(ΘT,F )
= (6)
∂Θi ∂Θj

− Ns tr R(ΘT,F , T Δts , F Δfs )−1 DΘj (ΘT,F , T Δts , F Δfs )
× R(ΘT,F , T Δts , F Δfs )−1 R̂(T Δts , F Δfs )

+ tr{R(ΘT,F , T Δts , F Δfs ) R̂(T Δts , F Δfs )} , (3)

where ln( · ) denotes the natural logarithm operation,
tr( · ) denotes the matrix trace, Δts and Nt denote the
sampling period and the number of samples with respect to
the time variable t, Δfs and Nf denote the sampling period
and the number of samples with respect to the frequency
variable f , Ns = Nt Nf is the total number of samples,
T ∈ {0, . . . , tmax − 1}, F ∈ {0, . . . , fmax − 1}, tmax =
1/(Bd Δts ), fmax = 1/(Δfs τmax ), Bd is the Doppler
−1
−1 fmax
spread, τmax is the maximum relative delay, and ΘT,F denotes
the vector of estimated parameters in the time and frequency
instances T Δts and F Δfs 1 . The matrix R̂(T Δts , F Δfs )
is the Lt Lr × Lt Lr matrix of measured correlation func-
tions, whereas R(ΘT,F , T Δts , F Δfs ) is the Lt Lr × Lt Lr
matrix of theoretical correlation functions, having elements
[R(ΘT,F , T Δts , F Δfs )]q+Lr (p−1),q̃+Lr (p̃−1) =
Rpq,p̃q̃ (ΘT,F , T Δts , F Δfs ) that are defined in (2). After
removing the constant terms that are not dependent on the
channel, the ML estimates are obtained as
 omitted for brevity.
Θ̂T,F = argmaxΘ̂T ,F − ln |R(ΘT,F , T Δts , F Δfs )|

−1
− tr{R(ΘT,F , T Δts , F Δfs ) R̂(T Δts , F Δfs ) , (4)
with the constraint ηT + ηR + ηT R = 1. The expression in
(4) is optimized using the sequential quadratic programming
algorithm [15]. The final vector 
of estimated parameters Θ is
−1 fmax −1
obtained as Θ̂ = (tmax fmax )−1 Tmax
t
=0 F =0 Θ̂T,F .
IV. C RAM ÉR -R AO B OUND
To study the performance of the proposed estimator, we
derive the Cramér-Rao lower bounds (CRLB) for the estimated
parameters Θ = [Θi ]8i=1 = [βTm , kT , μT , βRm , kR , μR , ηT ,
ηR ]. The CRLB bound of the estimated parameter Θi is the
diagonal element of the inverse Fisher information matrix
that corresponds to the parameter Θi . The (i, j)th element of
1 Operation · denotes rounding up to the next integer, B is estimated
d
from the measured sD-psd, and τmax is estimated from the measured psds.
978-1-4244-2324-8/08/$25.00 © 2008 IEEE.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2008 proceedings.
+ ILt LR − R(ΘT,F , T Δts , F Δfs )−1 R̂(T Δts , F Δfs )

∂
R(ΘT,F , T Δts , F Δfs )−1 DΘi (ΘT,F , T Δts , F Δfs ) ,
∂Θj
DΘi (ΘT,F , T Δts , F Δfs ) = ∂R(ΘT,F , T Δts , F Δfs )/∂Θi ,
DΘj (ΘT,F , T Δts , F Δfs ) = ∂R(ΘT,F , T Δts , F Δfs )/∂Θj ,
and ILt Lr is the Lt Lr × Lt Lr identity matrix. By observ-
ing that E[R̂(T Δts , F Δfs )] = R(ΘT,F , T Δts , F Δfs ), the
(i, j)th element of the Fisher information matrix becomes
 2
tmax
∂ L(Θ) Ns
[F(Θ)]i,j = −E =
∂Θi ∂Θj tmax fmax
T =0 F =0

tr R(ΘT,F , T Δts , F Δfs )−1 DΘj (ΘT,F , T Δts , F Δfs )

R(ΘT,F , T Δts , F Δfs )−1 DΘi (ΘT,F , T Δts , F Δfs ) , (7)
The expressions for the derivatives DΘi (ΘT,F , T Δts , F Δfs )
can be obtained by taking the partial derivatives of the STF-
CF in (2) with respect to Θi/j . Here, these expressions are
V. E VALUATION OF THE P ROPOSED E STIMATOR
In this section, we evaluate the performance of the proposed
estimator by comparing the analytical and measured STF-CF
and by comparing the mean square error of the parameter
estimates to the CRLB.
To verify our reference model with measured data, we need
to estimate the model parameters from the measurements. The
measured data used in this paper is reported in [16]. The dis-
tances between antenna elements (dT and dR ), the directions
and speeds of the Tx and Rx (γT , γR , vT , and vR ), the azimuth
angles of the Tx ’s and Rx ’s antenna arrays (θT and θR ), and
the elevation angles of the Tx ’s and Rx ’s antenna arrays (ψT
and ψR ) are obtained from the antenna array geometry and
video camera recordings. The Rice factor is estimated using
the moment-method in [17]. Furthermore, the parameters Rt2
and Rr2 are estimated from the measured psds by setting
Rt2 = Rr2 = c0 τmax /2, where τmax is the maximum relative
3
 1.0
delay and c0 is the speed of light. The parameters Rt1 and Rr1 Theoretical Re{R12,22[Δt,Δf=72 MHz]}

Space-time-frequency Correlation Functions

are chosen to be Rt1 = Rr1 = 0.1Rt2 . Finally, the parameters 0.8 Theoretical Im{R12,22[Δt,Δf=72 MHz]}
Θ = [βTm , kT , μT , βRm , kR , μR , ηT , ηR ] are estimated using 0.6
Measured Re{R12,22[Δt,Δf=72 MHz]}
the proposed maximum likelihood based stochastic estimator. Measured Im{R12,22[Δt,Δf=72 MHz]}
0.4
Figs. 2 - 4 compare the theoretical and measured correlation
functions. The parameters used to obtain curves in Figs. 2 - 4 0.2
are βT m = 5.1◦ , βRm = 10.2◦ , μT = 31.3◦ , μR = 141.7◦ ,
0.0
kT = 12.5, kR = 10.2, ηT = 0.213, ηR = 0.234, ηT R =
0.553, θT = θR = 0◦ , ψT = ψR = 0◦ , γT = γR = 90◦ , λ = -0.2
0.123 m, Rt1 = Rr1 = 3.1 m, Rt2 = Rr2 = 31 m, ΔH = 0, -0.4
D = 300 m, γ = 4, Lt = Lr = 2, dT = dR = 2.943 λ,
fT max = fRmax = 90.86 Hz, and K = 2.41. -0.6

-0.8
1.0
Theoretical Re{R12,22[Δt,Δf=0]} -1.0
0.8 Theoretical Im{R12,22[Δt,Δf=0]} 0 1 2 3 4 5
Space-time Correlation Functions

Measured Re{R12,22[Δt,Δf=0]} Normalized Time Delay [ωDΔt]

0.6
Measured Im{R12,22[Δt,Δf=0]}
0.4 Fig. 4. Comparison of the theoretical and measured space-time-frequency
correlation functions R12,22 (Δt, Δf = 72 Hz).
0.2
100
0.0
CRLB_kT
-0.2 10 MSE_kT
CRLB_μT
-0.4 MSE_μT
1
CRLB_βT
-0.6 MSE_βT
m

0.1 m
CRLB_ηT
-0.8
MSE_ηT
MSE

0.01
-1.0
0 1 2 3 4 5
Normalized Time Delay [ωDΔt] 1E-3

Fig. 2. Comparison of the theoretical and measured space-time correlation 1E-4

functions R12,22 (Δt, Δf = 0).
1E-5

1.0 1E-6
Theoretical Re{R11,11[Δt,Δf=72 MHz]} 100 1000 10000 100000 1000000
0.8 Theoretical Im{R11,11[Δt,Δf=72 MHz]} Ns
Time-frequency Correlation Functions

Measured Re{R11,11[Δt,Δf=72 MHz]}

0.6
Measured Im{R11,11[Δt,Δf=72 MHz]}
Fig. 5. Comparison of the MSE and CRLB for the parameters [Θi ]4i=1 =
0.4 [βTm , kT , μT , ηT ].

0.2

0.0 [βTm , kT , μT , βRm , kR , μR , ηT , ηR ], where Θ̂i denotes the ex-

-0.2 act value and Θ̃i denotes the estimated value of the pa-
rameter Θi . Since the results obtained for the parameters
-0.4
[βRm , kR , μR , ηR ] are almost identical with those obtained for
-0.6 the parameters [βTm , kT , μT , ηT ], Fig. 5 plots the CRLB and
-0.8
the MSE only for the parameters [Θi ]4i=1 = [βTm , kT , μT , ηT ].
The curves in Fig. 5 are obtained with the parameters in
-1.0
Figs. 2 - 4. The simulation results show that the proposed esti-
0 1 2 3 4 5
mator has asymptotically optimal performance, since it reaches
Normalized Time Delay [ωDΔt]
the CRLB for a small number of samples, i.e., Ns = 103 . The
Fig. 3. Comparison of the theoretical and measured time-frequency correla- close agreement between the theoretical and empirical curves
tion functions R11,11 (Δt, Δf = 72 Hz). confirms the utility of the proposed estimator.

To illustrate the performance of the proposed estimator, VI. C ONCLUSIONS

we compare the derived CRLB with the mean square er- In this paper, our 3-D reference model for wide-band MIMO
ror MSE = (Θ̂i − Θ̃i )2 for the parameters [Θi ]8i=1 = M-to-M channels is reviewed. To allow comparison between

978-1-4244-2324-8/08/$25.00 © 2008 IEEE. 4

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2008 proceedings.
 the proposed model and measured data, a new maximum
likelihood based stochastic estimator is derived. The proposed
estimator extracts the relevant model parameters from the mea-
sured data. The performance of the new estimator is evaluated
by deriving the CRLB and by comparing the mean square
error of the parameter estimates to the CRLB. Simulation
results show that the proposed estimator has an asymptotically
optimal performance, since it reaches the CRLB for a small
number of samples. Finally, the close agreement between the
theoretical and empirical curves confirms the utility of the
proposed estimator.
D ISCLAIMER
The views and conclusions contained in this document
are those of the authors and should not be interpreted as
representing the official policies, either expressed or implied,
of the Army Research Laboratory or the U. S. Government.
A PPENDIX A
T HE STF-CF S OF THE SINGLE - BOUNCED ,
DOUBLE - BOUNCED , AND L O S C OMPONENTS
In [5] is shown that the STF-CFs of the single-bounced
transmit, single-bounced receive, double-bounced, and LoS
components can be closely approximated as, respectively,
p̃q̃ (Δt, Δf ) ≈
SBT
Rpq,
 2π  outdoor propagation environments,” IEEE Trans. on Veh. Tech., vol. 49,
ηT cos λ βT m dT z −j 2π (q−q̃)dRx −j2πΔtfRmax cos γR
q̃)dRy ΔT /λ + ΔtfT max sin γT + ΔtfRmax ΔT sin γR ] + kT
sin μT , xSBR ≈ j2π[(q − q̃)dRx /λ + ΔtfRmax cos γR −
Δf Rr /c0 ] + kR cos μR , ySBR ≈ j2π[(q − q̃)dRy /λ + (p −
p̃)dT y ΔR /λ + ΔtfRmax sin γR + ΔtfT max ΔR sin γT ] + kR
sin μR , xDB ≈ j2π[(p − p̃)dT x /λ + ΔtfT max cos γT + Δf Rt /
c0 ]+kT cos μT , yDB ≈ j2π[(p− p̃)dT y /λ+ΔtfT max sin γT ]+
kT sin μT , zDB ≈ j2π[(q − q̃)dRx /λ + ΔtfRmax cos γR − Δf
Rr /c0 ]+kR cos μR , wDB ≈ j2π[(q − q̃)dRy /λ + ΔtfRmax
sin γR ] + kR sin μR , ADB = ηT R cos(2πβT m dT z /λ) cos(2π
βRm dRz /λ)2e−j2πΔf D/c0 /I0 (kT )I0 (kR )(1−(4βT m dT z /λ)2 )
(1−(4βRm dRz /λ)2 )(Rt2 2 2
−Rt1 2
)(Rr2 2
−Rr1 ), dT z = dT sin ψT ,
dT x = dT cos ψT cos θT , dT y = dT cos ψT sin θT , dRz =
dR sin ψR , dRx = dR cos ψR cos θR , dRy = dR cos ψR sin θR .
R EFERENCES
[1] A. S. Akki and F. Haber, “A statistical model for mobile-to-mobile land
communication channel,” IEEE Trans. on Veh. Tech., vol. 35, pp. 2–10,
Feb. 1986.
[2] M. Pätzold, B. O. Hogstad, N. Youssef, and D. Kim, “A MIMO mobile-
to-mobile channel model: Part I-the reference model,” Proc. IEEE
PIMRC, pp. 573–578, Berlin, Germany, Sept. 2005.
[3] A. G. Zajić and G. L. Stüber, “Space-time correlated MIMO mobile-to-
mobile channels,” Proc. IEEE PIMRC, Helsinki, Finland, Sept. 2006.
[4] A. G. Zajić and G. L. Stüber, “A three-dimensional MIMO mobile-to-
mobile channel model,” IEEE WCNC, Hong Kong, March 2007.
[5] A. G. Zajić and G. L. Stüber, “A three dimensional parametric model
for wideband MIMO mobile-to-mobile channels,” Proc. IEEE GLOBE-
COM, Washington, D.C., USA, Nov. 2007.
[6] K. I. Pedersen, P. E. Mogensen, and B. H. Fleury, “A stochastic model
of the temporal and azimuthal dispersion seen at the base station in
(8)
pp. 437–447, Mar. 2000.
 2 e λ e [7] B. H. Fleury, P. Jourdan, and S. Stucki, “High-resolution channel param-
I0 (kT ) 4βT m dT z
1− λ
eter estimation for MIMO applications using the SAGE algorithm,” Proc.
International Zurich Seminar on Broadband Communications, Access,
 Rt2     2π

2 2
2Rt 1 − γ RDt e−j c0 Δf (D+Rt ) Transmission, and Networking’02, pp. 30-1–30-9, Feb. 2002.
I0 xSBT + ySBT 2 − R2 dRt , [8] B. H. Fleury, M. Tschudin, R. Heddergott, D. Dahlhaus, and K. I. Peder-
Rt1 Rt2 t1 sen, “Channel parameter estimation in mobile radio environments using
the SAGE algorithm,” IEEE Journal on Selected Areas in Communica-
p̃q̃ (Δt, Δf ) ≈
SBR
Rpq, (9) tions, vol. 17, pp. 434–450, Mar. 1999.
 2π 
ηR cos λ βRm dRz j 2π (p−p̃)dT x j2πΔtfT max cos γT [9] A. Abdi and M. Kaveh, “A space-time correlation model for multiele-
 2 e λ e ment antenna systems in mobile fading channels,” IEEE J. on Select.
I0 (kR )
1 − 4βRm dRz Areas in Commun., vol. 20, pp. 550–560, Apr. 2002.
λ [10] C. B. Ribeiro, E. Ollila, and V. Koivunen, “Stochastic maximum
 Rr2    2π likelihood method for propagation parameter estimation,” Proc. IEEE
2 2
2Rr 1 − γ RDr e−j c0 Δf (D+Rr ) PIMRC, Sept. 2004.
I0 xSBR + ySBR 2 − R2 dRr , [11] C. B. Ribeiro, E. Ollila, and V. Koivunen, “Cram’er-Rao bound for
Rr1 Rr2 r1 angular propagation parameter estimation in MIMO systems,” Proc.
p̃q̃ (Δt, Δf ) ≈
DB
Rpq, (10) Asilomar Conference on Signals, Systems, and Computers, Nov. 2004.
 Rt2   [12] C. B. Ribeiro, E. Ollila, and V. Koivunen, “Propagation parameter
2π
ADB e−j c0 Δf Rt Rt I0
estimation in MIMO systems using mixture of angular distribution
x2DB + yDB 2 dRt model,” IEEE Int. Conf. Acoust., Speech, Signal Processing, pp. 18–
Rt1
 Rr2    23, Philadelphia, PA, March 2005.
2π Rr [13] A. Abdi, J.A. Barger, and M. Kaveh, ”A parametric model for the
× e−j c0 Δf Rr Rr I0 wDB2 2
+ zDB 1−γ dRr + distribution of the angle of arrival and the associated correlation function
Rr1 D and power spectrum at the mobile station,” IEEE Trans. on Veh. Tech.,
 Rr2   vol. 51, pp. 425–434, May 2002.
−j 2π
c0 Δf Rr 2 2
ADB e Rr I0 wDB + zDB dRr [14] J.D. Parsons and A.M.D. Turkmani, “Characterization of mobile radio
Rr1 signals: model description,” IEE Proc. I, Commun., Speach, and Vision,
 Rt2     vol. 138, pp. 549–556, Dec. 1991.
−j 2π
c0 Δf Rt 2 2 Rt
× e Rt I0 xDB + yDB 1−γ dRt , [15] R. Fletcher, Practical Methods of Optimization - Constrained Optimiza-
Rt1 D tion 2e.. John Wiley and Sons, 1981.
 2π [16] A.G. Zajić, G.L. Stüber, Thomas G. Pratt, and Son Nguyen, “Statistical
Rpq,p̃q̃ (Δt, Δf ) ≈
LoS
Kpq Kp̃q̃ ej λ [(p−p̃)dT x −(q−q̃)dRx ] (11) modeling with experimental verification for wideband MIMO mobile-
√ to-mobile channels in urban environments,” Proc. ICT, St. Petersburg,
j2πΔt[fT max cos γT −fRmax cos γR ]− 2π c0 Δf D 2 +ΔH 2
e , Russia, June 2008.
[17] L. J. Greenstein, D. G. Michelson, and V. Erceg, “Moment-method
where xSBT ≈ j2π[(p − p̃)dT x /λ + ΔtfT max cos γT + estimation of the Ricean K-factor,” IEEE Commun. Letters, vol. 3,
pp. 175–176, June 1999.
Δf Rt /c0 ] + kT cos μT , ySBT ≈ j2π[(p − p̃)dT y /λ + (q −

978-1-4244-2324-8/08/$25.00 © 2008 IEEE. 5

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2008 proceedings.