Sei sulla pagina 1di 5

Channel Prediction in Point–to–Point

MIMO–Systems
Nico Palleit, Tobias Weber
Institute of Communication Engineering, University of Ros tock
Richard–Wagner–Str. 31, 18119 Rostock, Germany
{nico.palleit, tobias.weber}@uni-rostock.de

Abstract—The present paper introduces three sophisticated of the individual discrete propagation paths like the directions
techniques to predict the channel transfer functions in frequency of arrival (DOAs), the directions of departure (DODs), the
division duplex MIMO–systems. In frequency division duplex times of arrival (TOAs), and the complex attenuations. These
systems it is beneficial to have the knowledge of the uplink
channel and the downlink channel at both sides of the point– parameters can be utilized to perform the uplink downlink
to–point transmission link. However, the base station can only transformation. The potential of the proposed techniques
estimate the uplink channel based on the transmission of a– which are based on the propagation path parameters will
priori known training signals from the mobile station. Due to be shown by means of measurement results. One further
the frequency selective behavior of the mobile radio channel, the technique to perform the uplink downlink transformation is
channel state information of the uplink channel and the channel
state information of the downlink channel are different. One based on filtering the previously estimated channel transfer
possibility to obtain the channel state information of the downlink functions (CTFs). In this case it is necessary to estimate the
channel in the base station is to feed back this information from filter coefficients of the predictor filter. Measurement results
the mobile station. Unfortunately, this feedback decreases the show the potential of the proposed technique.
spectral efficiency. In former contributions, different techniques for predicting
The presented techniques to obtain the channel state informa-
tion of the downlink channel in the base station are based on the CTFs in SISO systems were investigated. However, flat
the physical path parameters and are based on filtering the pre- fading was assumed and based on the estimated propagation
viously estimated channel state information. The performance of path parameters, the future channel coefficients were predicted
the investigated techniques is evaluated by means of measurement but not the channel coefficient for another frequency band
results. [2], [3]. Another approach was to model the channel co-
efficients as samples of an autoregressive process and with
I. I NTRODUCTION
the estimated coefficients of the autoregressive process and
In the following, we consider a system where the base with the previously estimated channel coefficients, the future
station (BS) is the receiver and the mobile station (MS) is the channel coefficients were predicted but again not the channel
transmitter. Both sides are equipped with an antenna array. coefficients for another frequency band [4]. In some contri-
The prediction of the channel state information (CSI) is done butions the CTF was also estimated for another frequency
in the BS. However, all techniques presented in this paper can band in SISO systems [5]. The authors draw the conclusion
also be applied for the case that the BS is the transmitter an that the performance of the techniques which are based on
the MS is the receiver, thus the prediction can be also done the autoregressive approach is better than the performance
in the MS. of the techniques based on the propagation path parameters.
The knowledge of the uplink channel and the downlink However, the spatial properties of the MIMO channel can
channel in the BS enables some kind of precoding in the BS in be exploited to improve the performance of the techniques
order to increase the system performance [1]. Unfortunately, which are based on the propagation path parameters. With the
the BS can only estimate the uplink channel based on the assumption that the DOAs, DODs and TOAs do not change
transmission of a–priori known training signals because the significantly over frequency [6] it is reasonable to utilize these
channel reciprocity cannot be exploited in frequency division parameters to perform the uplink downlink transformation.
duplex (FDD) systems. One possibility to obtain the knowl- The remainder of the paper is organized as follows: In Sec-
edge of the downlink channel in the BS is to feed back this tion II we introduce the underlying channel models. Section
information but this decreases the system performance. Thus, III is dedicated to the proposed algorithms. The equivalence
the goal is to obtain the knowledge of the downlink channel between the prediction algorithms is proven in Section IV.
in the BS without any feedback from the MS which is called The measurement results are presented in Section V. Finally,
uplink downlink transformation. Novel techniques to reach this in Section VI we draw our conclusion.
goal in point–to–point MIMO–systems will be investigated in
this paper. II. C HANNEL M ODELS
With an antenna array at the BS and an antenna array at We consider a scenario where the MS is equipped with
the MS it is possible to estimate the physical path parameters N antennas and the BS is equipped with M antennas. The

978-1-4244-6317-6/10/$26.00 © 2010 IEEE 91 ISWCS 2010


holds, thus
(m,d) (m,d)
path D plane wave e−j2π(f +f0 )τRx ≈ e−j2πf0 τRx
(n,d) (n,d)
(n,1)
τTx α(D) , τ (D) e−j2π(f +f0 )τTx ≈ e−j2πf0 τTx (3)
Tx n Rx m
RP (m,1) holds and the CTF in (1) results in
τRx
(n,D) D
(m,D) X
τTx path 1 RP τRx H (m,n) (f, f0 ) =
(d)
H d (f, f0 ) · e−j2πf0 τRx
(m,d)
· e−j2πf0 τTx
(n,d)

MS BS
α(1) , τ (1) d=1
(4)
plane wave with (d)
(d)
H d (f, f0 ) = α(d) · e−j2π(f +f0 )τ . (5)
(d)
Here H d (f, f0 ) is the directional CTF and depends neither
Fig. 1. Directional channel scenario on the DOAs, DODs nor on the geometry of the antenna
(m,d) (n,d)
arrays. The factors e−j2πf0 τRx and e−j2πf0 τTx are the the
steering factors and depend on the geometry of the antenna
(d) (d)
directional scenario is depicted in Figure 1. It is assumed that arrays, the carrier frequency f0 and on the DOAs φRx and θRx
(d) (d)
D electromagnetic waves propagate from the transmitter to the and on the DODs φTx and θTx respectively [7]. The angle
receiver, which interact with other objects. These objects are in φ(d) describes the azimuth and can be in the range [0; 2π[ and
the far filed of the antennas, thus D plane waves impinge at the the angle θ(d) describes the pole–distance and can be in the
receiver antennas. The propagation paths will have different range [0; π[, where θ = π/2 describes the horizon. With the
complex attenuations, DOAs, DODs and TOAs. The spatial knowledge of the DOAs and DODs the path specific delays
distances among the antennas of one antenna array are in (m,d) (n,d)
τRx and τTx can be easily calculated [7].
the range of some wavelength, thus it is assumed that the Sampling the CTF (4) with the sample distance F in the
complex attenuation, the DOAs and the DODs are equal for frequency domain leads to the channel vector h(m,n) (f0 ) with
all antennas. The TOAs are defined as the times which are W ≥ 1 samples of the CTF, where the w-th sample can be
needed for the propagation of the electromagnetic waves from written as
the reference point (RP) at the transmitter side to the RP at D
X (d)
the receiver side. These RPs are the geometrical centers of H (m,n)
w = α(d) · e−j2π(F ·w+f0 )τ
the antenna arrays. With these assumptions the band limited d=1
CTF at the center frequency f0 in the equivalent low–pass (m,d) (n,d)
· e−j2πf0 τRx · e−j2πf0 τTx , (6)
domain between the m-th, m = 1 . . . M , receive antenna and
the n-th, n = 1 . . . N , transmit antenna can be described by a with w = − W2−1 , . . . , W2−1 . The combination of all M · N
superposition of D complex exponential functions channel vectors leads to the W × M · N channel matrix
 
  XD
H(f0 ) = h(1,1) (f0 ) . . . h(M,1) (f0 ) . . . h(M,N ) (f0 ) . (7)
f (d)
H (m,n) (f, f0 ) = rect · α(d) · e−j2π(f +f0 )τ Sampling the directional CTF of (5) leads to the directional
B (d)
d=1
(m,d) (n,d)
channel vector hd (f0 ). The combination of all D directional
· e−j2π(f +f0 )τRx · e−j2π(f +f0 )τTx . (1) channel vectors leads to the W × D directional channel matrix
 
(1) (D)
Hd (f0 ) = hd (f0 ) . . . hd (f0 ) . (8)
Here α(d) is the complex weight of the d-th, d = 1 . . . D,
propagation path and it is assumed that the attenuations of the The combination of the steering factors e−j2πf0 τRx for all
(m,d)

individual propagation paths do not depend on the frequency DOAs for all M antennas results in the M ×D steering matrix
f within the considered bandwidth B. τ (d) is the delay of (n,d)
ARx (f0 ). The combination of the steering factors e−j2πf0 τTx
the d-th propagation path. The difference of the TOA of the
for all DODs for all N antennas results in the N × D steering
d-th propagation path at the RP at the receiver side and at the
(m,d) matrix ATx (f0 ). Merging both steering matrices leads to the
m-th antenna is described by τRx and can be negative. The M · N × D complete steering matrix
difference of the TOA of the d-th propagation path at the RP
at the transmitter side and at the n-th antenna is described A(f0 ) = ATx (f0 ) ⊛ ARx (f0 ), (9)
(n,d)
by τTx and can be negative. For the sake of simplicity
where ⊛ denotes the operator for the Khatri–Rao product.
it is assumed that f is limited to the range [−B/2; B/2],
With these matrices the channel matrix can be written as
thus the rectangular function rect (·) in (1) can be neglected.
Furthermore, it is assumed that H(f0 ) = Hd (f0 )AT (f0 ). (10)

(m,d)

(n,d) In addition to this channel model which is based on the
B · τRx ≪ 1 and B · τTx ≪ 1 (2) physical description of the propagation paths it is also possible

92
to describe an element of the CTF as a linear combination of The second approach to estimate the directional channel matrix
P other known elements of the same CTF. In this case the is to reconstruct the directional channel matrix (5) with the
(w + ∆w)-th element of the (m, n)-th CTF can be written as knowledge of the D TOAs and the D complex attenuations.
(m,n)

(m,n)
 This approach is called reconstruction approach.
H w+∆w = H (m,n)
w . . . H w−(P −1) · pf . (11)
Assuming ∆w = 1, in some contributions this channel model B. Filter based prediction
is called autoregressive channel model but an autoregressive In order to perform the uplink downlink transformation of
process results from filtering white Gaussian noise and under the CTFs based on the filter based channel model (11) it is
ideal conditions there is no noise in this model. In this paper necessary to have the knowledge of P preceding samples for
∆w can be greater than 1. For this reason this channel model each CTF and the knowledge of the P filter coefficients. With
is called filter based channel model. a prediction order of P and W known samples it is possible
At a first glance, these two channel models seem to be to predict (W − P + 1) samples of the CTF
fundamentally different. Nevertheless, both can be used in    (m,n) 
(m,n) (m,n)
order to predict the CTF. H P −1+∆w H P −1 · · · H0
 ..   . .. 
III. P REDICTION   =  .. . ·p ,
AND ESTIMATION OF THE MODEL  .    f
PARAMETERS (m,n) (m,n) (m,n)
H W −1+∆w H W −1 · · · H W −1−(P −1)
A. Propagation path based prediction (16)
With equation (4) and (5) a sample of the (m, n)-th CTF in where pf is the vector with the P filter coefficients. In order
the equivalent low–pass domain at the center frequency f0 + to predict more than (W − P + 1) samples of the CTF this
∆f is calculated as technique can be continued iteratively.
The aim is to estimate the filter coefficients. Based on (16)
D
X
(m,n) (d) (d) it is possible to estimate pf by predicting a–priori known
H w+∆w = α(d) · e−j2π(F ·w+f0 )τ · e−j2πF ∆wτ
samples of the CTF
d=1
(m,d) (m,d)  
· e−j2πf0 τRx · e−j2πF ∆wτRx (m,n)
H P −1+∆w
(n,d) (n,d)  
· e−j2πf0 τTx · e−j2πF ∆wτTx , (12)  .. 
 . 
(m,n)
where ∆f = F ∆w is a multiple of F . The channel matrix H W −1
H(f0 + ∆f ) can be decomposed into the directional channel 
(m,n) (m,n)
 
pf,0

H P −1 ··· H0
matrix and the steering matrix    . 
= .. ..  ·  .  . (17)
H(f0 + ∆f ) = Hd (f0 + ∆f )AT (f0 + ∆f ). (13)  . .   . 
(m,n) (m,n) pf,P −1
H W −1−∆w · · · H W −∆w−P
Thus, it becomes feasible to predict the steering matrix A(f0 ) | {z } | {z }
and the directional channel matrix Hd (f0 ) separately. This H̃
(m,n) pf

leads to
 −j2π∆f τ̂ (1)  For the sake of simplicity in (17) only the estimation of
e 0 a–priori known samples for one CTF is shown. It is easily

Ĥ(f0 + ∆f ) = Ĥd (f0 ) ·  ..  possible to extend equation (17) in such a way that all M · N
. 
(D) CTFs will be taken into account. As long as P + ∆w ≤ W
0 e−j2π∆f τ̂
 T holds, all samples of the CTF in (17) are known, and thus it
· Â(f0 ) ⊙ Â(∆f ) , (14) is feasible to perform a least squares estimation of the vector
pf , where at least as many examples as predictor coefficients
where ⊙ denotes the operator for the element wise multipli- should exist
cation.  
(m,n)
In order to perform the uplink downlink transformation, it  −1 H P −1+∆w
∗T ∗T  .. 
is necessary to firstly estimate the model parameters. This can p̂ = H̃(m,n) · H̃(m,n)
f
· H̃(m,n) · 
 .
.

be done, e.g., by using the SAGE algorithm [8]. (m,n)
Secondly, it is necessary to estimate Hd (f0 ) and A (f0 ). H W −1
(18)
With the knowledge of the D DOAs the steering matrix can be
Setting ∆w = 1 the method in (18) is known in the literature
easily reconstructed for any center frequency [9], as needed
as the covariance method [10].
in (14). In order to estimate the directional channel matrix
we propose two possibilities. The first approach is a least
IV. E QUIVALENCE OF BOTH PREDICTION ALGORITHMS
squares approach. Based on (10) it is possible to perform a
least squares estimation of the directional channel matrix The aforementioned uplink downlink transformation algo-
 −1 rithms (14) and (16) seem to be completely different. However,
Ĥd (f0 ) = H (f0 ) · A∗ (f0 ) · AT (f0 ) · A∗ (f0 ) . (15) it can be shown that both approaches are equivalent.

93
Replacing the elements of the filter based channel model thus the power of the strongest path was approximately 20 dB
(11) by the elements of the directional channel model (12) higher than the power of the weakest path. In order to assess
leads to the proposed algorithms we only used the first 100 MHz of the
D
X (m,d)
measured CTFs for estimating the physical path parameters,
(d) (d)
α(d) · e−j2π(F ·w+f0 )τ · e−j2πF ∆wτ · e−j2πf0 τRx i.e., the uplink channel.
d=1 The next step was to predict the CTFs. In order to predict
(m,d) (n,d) (n,d)
· e−j2πF ∆wτRx · e−j2πf0 τTx · e−j2πF ∆wτTx the CTFs based on the filter based channel model we estimated
XD the filter coefficients according to (18). ∆w was set to 1 and
(d) (m,d) (n,d)
= α(d) · e−j2π(F ·w+f0 )τ · e−j2πf0 τRx · e−j2πf0 τTx the known channel coefficients of the first 100 MHz were used
d=1 for estimating the filter coefficients. We set the filter order P
P
X −1
(d) (m,d) (n,d)
in such a way that P · F = 20 MHz holds. After that, we
· pf,p · ej2πF pτ · ej2πF pτRx · ej2πF pτTx , (19) predicted the CTFs based on (16) iteratively.
p=0 To study the performance of the proposed algorithm,
where pf,p , p = 0 . . . P − 1, are constant. One possibility we limited the measured CTFs at the carrier frequency of
to prove that the equality in (19) holds, is to show that the 2.375 GHz to a bandwidth of 50 MHz by filtering. This led
equality also holds for any path d individually. This means to h(m,n) (f0 ). After that, we limited the predicted CTFs at
that the sum over d in (19) has to be neglected which leads to the carrier frequency of (2.375 GHz + ∆f ) to a bandwidth
(d) (m,d) (n,d)
of 50 MHz by filtering. This led to the predicted CTFs
e−j2πF ∆wτ · e−j2πF ∆wτRx · e−j2πF ∆wτTx ĥ
(m,n)
(f0 + ∆f ). To evaluate the quality of this prediction,
P
X −1
(d) (m,d) (n,d) we also limited the measured CTFs at the carrier frequency of
= pf,p · ej2πF pτ · ej2πF pτRx · ej2πF pτTx . (20)
(2.375 GHz + ∆f ) to a bandwidth of 50 MHz by filtering to
p=0
obtain h(m,n) (f0 + ∆f ) and calculated the normalized mean
If (20) holds for any path d, the equality in (19) is proven. square error between both
The equality in (20) for any path d, leads to a system of D
M P N (m,n) 2
linear equations with P unknowns [11]. This system of linear P
(f0 + ∆f ) − h(m,n) (f0 + ∆f )


equations has no solution in the case of D > P . This means m=1 n=1
ep (∆f ) = .
that the filter order must be equal or higher than the number M P
P N
(m,n)
2

D of propagation paths. Obviously, the filter coefficients pf,p h (f0 + ∆f )
m=1 n=1
depend on the sample distance F , on ∆w, on τ (d) , and on (21)
(m,d) (n,d)
τRx and τTx but not on α(d) . In dependency of the applied approach for predicting the
It is shown that also the filter based channel model is CTFs, the normalized mean square error of (21) is different. In
motivated in a physical sense. the range from ∆f = 0 to ∆f = 50 MHz we did not predict
the CTFs based on the filter based channel model because
V. M EASUREMENT R ESULTS the CTFs are a–priori known, whereas we also predicted the
In order to get an idea of the performance of the pre- CTFs based on the directional channel, although the CTFs
sented techniques, we performed channel measurements in are a–priori known. Thus, even at ∆f = 0 the error of the
an indoor scenario using a vector network analyzer. We used reconstruction approach and the least squares approach is not
one transmit antenna and one receive antenna. Both antennas equal to zero.
could be moved on a circular path using two turntables for The benefits of the prediction over the trivial approach —
forming two virtual uniform circular arrays. In Figure 2, not to perform any prediction at all and to use the M ·N CTFs
the panorama view from the receiving virtual antenna array h(m,n) (f0 ) for all frequencies f0 + ∆f — can be assessed by
towards the transmitting virtual antenna array is shown. The comparing (21) to
antennas were aligned in such a way that the direct line of
M P N 2
sight corresponds to an azimuth angle of 0◦ . The antennas P (m,n)
h (f0 ) − h(m,n) (f0 + ∆f )

were moved in 10◦ steps corresponding to M = N = 36
eu (∆f ) = m=1 n=1M N .
antenna positions. The center frequency was 2.45GHz and P P (m,n) 2

the bandwidth was 200 MHz. The radius of the turntable h (f0 + ∆f )
m=1 n=1
was 25 cm, thus the distance between two adjacent antenna (22)
positions was approximately one third of a wavelength. One The results are shown in the Figures 3 and 4 with different
half of the wavelength would be sufficient, however in order to scalings. It can be seen that it is not appropriate to directly
show the capabilities of the presented techniques we measured use the known uplink CTFs for the downlink. The result
the CTFs at 36 · 36 different positions. of the filter approach is best up to a prediction width of
The first step in order to perform the uplink downlink 100 MHz. For higher ∆f the prediction algorithms which
transformation based on the directional channel model is are based on the directional channel model performs better.
to estimate the propagation path parameters by the SAGE Altogether, the reconstruction approach performs best for a
algorithm. We assumed to have D = 22 propagation paths, high prediction width. Nevertheless, even at ∆f = 0 the mean

94
transmit antenna

Fig. 2. Measurement setup

4
(m,n) ∆f = 0 but incorrect results from the SAGE algorithm lead
ĥ (f0 + ∆f ) = h(m,n) (f0 ) to poorer results at ∆f > 0. Depending on the reason why
no prediction this frequency prediction was done, the results can be used
normalized mean square error

3 either directly or the error can be further reduced by feeding


back information from the MS to the BS with less information
than it would be required for the complete feedback of the
estimated CTFs.
2
VI. C ONCLUSION AND F UTURE W ORK
This paper presents the principles of three sophisticated
uplink downlink transformation techniques in point–to–point
1 reconstruction MIMO–systems based on the estimation of physical propa-
approach least squares filter approach
approach gation path parameters and based on the estimation of filter
coefficients of a predictor. For small prediction widths the
filter approach performs best. For higher prediction widths
0
0 50 100 150 the reconstruction approach performs best. Future work may
∆f /MHz
include further measurement results and the prediction of the
Fig. 3. Normalized mean square error versus ∆f CTFs into time direction.
R EFERENCES
0.5
no prediction [1] G. J. Foschini and M. J. Gans, “On Limits of Wireless Communications
(m,n)
in a Fading Environment When Using Multiple Antennas,” Wireless
ĥ (f0 + ∆f ) = h(m,n) (f0 ) Personal Communications, vol. 6, pp. 311–335, 1998.
normalized mean square error

0.4 [2] Jeng-Kuang Hwang and J. H. Winters, “Sinusoidal Modeling and


Prediction of Fast Fading Processes,” in Proc. IEEE Global Telecom-
munications Conference (GLOBECOM’05), 1998, vol. 2, pp. 892–897.
least squares [3] J. B. Andersen, J. Jensen, S. H. Jensen, and F. Frederiksen, “Prediction
approach
0.3 of Future Fading Based on Past Measurements,” in Proc. IEEE 50th
reconstruction Vehicular Technology Conference (VTC’99-Fall), 1999, pp. 151–155.
approach [4] T. Eyceoz, A. Duel-Hallen, and H. Hallen, “Deterministic Channel
Modeling and Long Range Prediction of Fast Fading Mobile Radio
0.2 Channels,” in IEEE Communications Letters, 1998, pp. 254–256.
[5] S. Semmelrodt and R. Kattenbach, “A 2-D Fading Forecast of Time–
Variant Channels Based on Parametric Modeling Techniques,” in Proc.
of the 13th IEEE International Symposium on Personal, Indoor and
0.1 Mobile Radio Communications (PIMRC ’02), 2002, pp. 1640–1644.
filter approach [6] Klaus Hugl, Kimmo Kalliola, and Juha Laurila, “Spatial Reciprocity of
Uplink and Downlink Radio Channels in FDD Systems,” Tech. Rep.,
COST 273 TD(02)066, 2002.
0 [7] Joseph C. Liberti Jr. and Theodore S. Rappaport, Smart Antennas
0 50 100 150
∆f /MHz For Wireless Communications: IS-95 and Third Generation CDMA
Applications, Prentice Hall, New Jersey, 1999.
Fig. 4. Normalized mean square error versus ∆f [8] Jeffrey A. Fessler and Alfred O. Hero, “Space–Alternating Generalized
Expectation–Maximization Algorithm,” in IEEE Transactions on Signal
Processing, 1994, vol. 42, pp. 2664–2675.
[9] Harry L. Van Trees, Optimum Array Processing Part IV of Detection,
Estimation, and Modulation Theory, John Wiley & Sons, 2002.
square error ep (∆f ) is not equal to zero for the techniques [10] Simon Haykin, Adaptive Filter Theory, Prentice Hall, 4. edition, 2002.
wich are based on (14). This results from the assumption of [11] N. Palleit and T. Weber, “Frequency Prediction of the Channel
D = 22 propagation paths. The real number of propagation Transfer Function in Multiple Antenna Systems,” in Proc. International
ITG/IEEE Workshop on Smart Antennas (WSA’10), Bremen, 2010.
paths is much greater than 22 and additionally there are not
only DOAs and DODs but also some kind of angular spreads,
which are not considered in the directional channel model.
Nevertheless, increasing D would lead to a smaller error at

95

Potrebbero piacerti anche