Sunteți pe pagina 1din 11

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO.

5, MAY 2010

1415

Estimation of Cubic Nonlinear Bandpass Channels in Orthogonal Frequency-Division Multiplexing Systems


Ching-Hsiang Tseng

AbstractModeling and compensation of nonlinear communication channels has long been an important research topic in digital communications. A nonlinear bandpass channel is commonly modeled by a baseband equivalent Volterra series which relates the complex envelopes of the channel input and output. In this paper, we propose a novel method to estimate the frequency-domain baseband equivalent Volterra kernels of cubically nonlinear bandpass channels in orthogonal frequencydivision multiplexing (OFDM) systems. We recognize that the input signal for an OFDM system employing QAM or PSK modulations satises the properties of a kind of random multisine signal. By exploring the higher-order auto-moment spectra of the random multisine signal, a computationally efcient algorithm for determining the frequency-domain baseband equivalent Volterra kernels is derived. The obtained kernel estimates are optimal in the minimum mean square error (MMSE) sense. The proposed method can be used to estimate nonlinear bandpass channels for OFDM systems employing pure QAMs, pure PSKs, or a mixture of QAMs and PSKs. The effectiveness of the proposed method is demonstrated by applying it to estimate the nonlinear bandpass channel of an example OFDM system. Nonlinear channel compensators based on the Volterra model can benet from the proposed method. Index TermsCommunication system nonlinearities, communication channels, estimation, frequency division multiplexing, frequency domain analysis, nonlinear distortion, nonlinear systems, power ampliers, Volterra series.

the various frequency components will exist in the output of the nonlinear channel. This yields out-of-band spurs and inter-subcarrier interference in the OFDM system. In addition, the orthogonality of the subcarriers is essential to an OFDM system. The AM/PM distortion of the power amplier within the OFDM symbol period might cause the OFDM subcarriers to lose their orthogonality. These undesirable effects are often avoided by backing off the operation point of the power amplier to a linear region. This strategy, however, causes the OFDM system to suffer from poor power efciency. The bandpass Volterra series was extensively used to model the nonlinear characteristics of the power amplier in communication channels [1], [4][7]. This is due to the fact that channel impairments like the intermodulation distortion can be readily estimated with the knowledge of the Volterra kernels. In addition, the bandpass Volterra series has also been used to develop methods such as nonlinear equalization [8], nonlinear cancellation [9], and nonlinear predistortion [10][13] for combating channel nonlinearities. In the development of these methods, knowledge of the Volterra kernels of the nonlinear channel may facilitate the design of the compensators. In modeling nonlinear channels using the bandpass Volterra series, the channel model can be simplied to a baseband equivalent representation in which the complex envelopes of the channel input and output are related by a baseband equivalent Volterra series [6], [14]. Besides the fact that the input and output are complex valued, the baseband equivalent Volterra series differs from the real-valued Volterra series in that it has a different mathematical expression. These are the fundamental differences one should bear in mind in developing baseband equivalent Volterra kernel estimation algorithms for bandpass channels. For determining the time-domain baseband equivalent Volterra kernels of nonlinear bandpass channels in PSK and QAM systems, algorithms based on the assumption that the input is an independent, identically-distributed (i.i.d.) signal can be found in the literature [15], [16]. For OFDM systems, a method for time-domain estimation of cubic baseband Volterra channels was developed based on the property that the OFDM signal is asymptotically i.i.d. complex Gaussian [17]. The same asymptotic property of the OFDM signal was also used in [18] to develop a time-frequency domain estimation method for general 2 + 1 order baseband Volterra channels in OFDM systems. Note that methods for estimating the timedomain baseband equivalent Volterra kernels are suitable for nonlinear channels with short memory. For nonlinear channels

I. I NTRODUCTION HARACTERIZING and compensation of channel impairments is an essential task in digital communications. One important channel impairment is the nonlinearity introduced by the AM/AM and AM/PM conversion effects [1] of the high-power amplier in the channel. The power amplier often needs to be driven at or near the saturation point to provide better power efciency. This causes nonlinear distortion, adjacent channel interference, and intersymbol interference in the communication system. This problem is especially important in OFDM systems [2], where the transmitted signal may exhibit a high peak-to-average power ratio (PAR) [3] and can easily saturate the power amplier. Since there are multiple frequency components in an OFDM signal, harmonics of each frequency component and cross modulation products among
Paper approved by N. Benvenuto, the Editor for Modulation and Detection of the IEEE Communications Society. Manuscript received October 29, 2008; revised June 26, 2009, October 7, 2009, and December 12, 2009. The author is with the Department of Electrical Engineering, National Taiwan Ocean University, 2 Pei-Ning Rd., Keelung, Taiwan 202, R.O.C. (email: chtseng@mail.ntou.edu.tw). Digital Object Identier 10.1109/TCOMM.2010.05.080573

c 2010 IEEE 0090-6778/10$25.00

1416

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

with long memory, the computational complexity of these methods may increase dramatically due to the large number of involved time-domain Volterra kernel coefcients. Under the circumstances, a frequency-domain approach might offer certain advantage, because channels with long memory tend to have a narrow bandwidth and hence a small number of signicant frequency-domain Volterra kernel coefcients. Besides, for communication systems which employ frequency multiplexing (e.g., OFDM), a method for directly estimating the frequency-domain baseband equivalent Volterra kernels would be a more natural choice. In this paper, we propose a computationally efcient method for estimating the frequency-domain baseband equivalent Volterra kernels of cubically nonlinear bandpass channels in OFDM systems. We recognize that the input signal to an OFDM system is in fact a random multisine signal. The random multisine signal has been used as the driving signal for studying many science and engineering problems involving nonlinearities [19][21]. In particular, it has recently been used to derive a fast algorithm for identifying these frequencydomain Volterra kernels of real-valued nonlinear systems [22]. In [22], the fast algorithm was obtained by exploring the higher-order auto-moment spectra of the random multisine signal. The concept used in [22] is adopted in this paper to derive a minimum mean square error (MMSE) estimation algorithm for the frequency-domain baseband equivalent Volterra kernels of nonlinear bandpass channels. However, due to the fundamental differences between the baseband equivalent Volterra series and the real-valued Volterra series, distinct higher-order auto-moment spectra of the random multisine signal need to be considered for the case of nonlinear bandpass channels. This leads to a different estimation algorithm which is dedicated to nonlinear bandpass channels. In addition, by allowing the random envelopes used at different tones to be distinct, the proposed method is directly applicable to the problem of nonlinear bandpass channel estimation for OFDM systems employing pure QAMs with alphabet sizes larger than 4. Furthermore, for PSK modulation schemes, certain cubic Volterra kernel coefcients will degenerate to linear kernel coefcients [16]. Therefore, for OFDM systems employing pure PSKs or a mixture of QAMs and PSKs, modied algorithms are derived to take the kernel degeneration problem of the PSK into account. The rest of this paper is organized as follows. In Section II, the baseband equivalent Volterra model for nonlinear bandpass channels is introduced. The symmetry properties and the estimation region of the cubic Volterra kernel are discussed. The random multisine signal is also dened. In Section III, we explore the properties of the higher-order auto-moment spectra of the random multisine signal and derive the algorithm for estimating the frequency-domain baseband equivalent Volterra kernels of nonlinear bandpass channels. In Section IV, the applicability of the proposed algorithm to OFDM systems employing pure QAMs is veried. Modied algorithms for OFDM systems employing pure PSKs and a mixture of QAMs and PSKs are also derived. The goodness of the proposed algorithms for estimating OFDM nonlinear bandpass channels is demonstrated by computer simulation in Section V.

II. N ONLINEAR BANDPASS C HANNEL WITH A R ANDOM M ULTISINE I NPUT S IGNAL A typical example of a nonlinear bandpass channel is a communication link in which the power amplier of the transmitter is nonlinear and operating at the radio frequency (RF) range [1]. Let () and () be the bandpass input and output signals of the channel, respectively. They can be written as () () = () cos(2 + ()) = () cos(2 + ()) = () cos(2 + () + ()) (1) (2)

where () and () are real-valued amplitude functions, (), (), and () are real-valued phase functions, and is the carrier frequency. Note that () and () are both functions of () only and hence can be expressed as () = [()] and () = [()], respectively. The two functions [] and [] characterize the AM/AM and AM/PM distortions caused by the nonlinear power amplier [14]. In general, passing the bandpass signal () dened in (1) through a nonlinear system would normally yield an output signal with spectral components centered at integer multiples of . However, here we assume that the channel is bandpass and hence only the spectral components centered at survive in (2). When () is an OFDM signal, it can be expressed as () = Re{() 2 }, where () = = 2/ . That is, the signal () contains multiple frequency components. In this case, the cross modulation products among the various frequency components in () would contribute to the output signal () in (2). More on the OFDM signal will be addressed in Section IV. The baseband equivalent input and output signals of the channel are dened as () () = ()() = ()
()

(3) (4)

Note that () and () are in general complex-valued signals. They are related by the baseband equivalent Volterra series as follows [14]: () =
=0 +1 =1

2+1 (1 , 2 , , 2+1 )
2 +1 =+2

( )

( )1 2 2+1 (5)

where the baseband equivalent Volterra kernels 2+1 (1 , 2 , , 2+1 ), = 0, 1, . . ., are in general complex valued. They are not equal to (but can be related to) the original real-valued Volterra kernels. However, unless otherwise specied, the baseband equivalent Volterra kernels will simply be referred to as the Volterra kernels in the rest of this paper. Note that only odd-order Volterra kernels are involved in (5). This is due to the fact that even-order Volterra kernels do not generate in-band signals [14]. Let ( ), ( ), and 2+1 (1 , 2 , . . . , 2+1 ) be the Fourier transforms of (), (), and 2+1 (1 , 2 , . . . , 2+1 ),

TSENG et al.: ESTIMATION OF CUBIC NONLINEAR BANDPASS CHANNELS IN ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING SYSTEMS

1417

respectively. The term 2+1 (1 , 2 , . . . , 2+1 ) is referred to as the frequency-domain Volterra kernel. We have 1 ( ) 2 (6) () = [ ( )] =

k
plane b I

where the [] denotes the inverse Fourier transform operator. Using (6), one can easily derive ( ) 2 2 , ( ) = ( ) = 1, 2, . . . , + 1 (7) = ( ) 2 2 , = + 2, . . . , 2 + 1 By substituting (7) and (8) into (5), one obtains () = 2+1 (1 , 2 , , 2+1 )
=0 +1 =1

S
E H D F O

(0,0,M)

plane b : i+j+k = m

(8)
(M,0,0)

(0,M,0)

(m,m,-m)

( )

2 +1 =+2

T
B
(9)

( )

2(1 +2 ++2+1 ) 1 2 2+1

i
Fig. 1. The estimation region for the CVK. The line segments QE and CQ show the intersections of the plane + + = with the planes = and = .

where the fact that 2+1 (1 , 2 , . . . , 2+1 ) is the (2 + 1)dimensional Fourier transform of 2+1 (1 , 2 , . . . , 2+1 ) has been used in deriving (9). By taking the Fourier transform of (9), one can obtain the following frequency-domain representation: 2+1 (1 , 2 , , 2+1 ) ( ) =
=0 +1 =1

In addition, due to the ( ) in the third-order product term of (11), we see that 3 (, , ) = 3 (, , ) = 3 (, , ) (15)

( )

2 +1 =+2

( )

( 1 2 2+1 )1 2 2+1 (10) where ( ) is the delta signal. Based on (10), we see that the discrete frequency-domain Volterra model for a cubically nonlinear bandpass channel can be written as: () = () () + 3 (, , ) () ( ) ( ) + (), (11)
,, (+ +=)

However, since both 3 (, , ) and 3 (, , ) are multiplied by the same third-order product term () ( ) ( ) in (11), we can assume without loss of generality that 3 (, , ) = 3 (, , ). (16)

where () and () are the discrete Fourier transforms (DFTs) of the input () and output (), respectively. The linear Volterra kernel (LVK) and the cubic Volterra kernel (CVK) are denoted by () and 3 (, , ), respectively. The maximum discrete frequency of () with a non-zero value is denoted by , and the modeling error is denoted by (). The estimation region for 3 (, , ) in the threedimensional (, , ) space is a cube, as shown in Fig. 1. Note that, since (), (), and 3 (1 , 2 , 3 ) are complex valued, their Fourier transforms are not conjugate symmetric with respect to zero. Therefore, the usual symmetry properties of the real-valued Volterra series do not apply here. That is, () = () = 3 (, , ) = (), (),
3 (, , ).

By taking (16) into account, the estimation region of the CVK can only be reduced from the cube to the triangular prism GIHAFB in Fig. 1. Bearing the symmetry property (16) in mind, we see that, only those CVK coefcients on the pentagon SEDCT (which is on the plane + + = ) in Fig. 1 need to be estimated in (11). By varying from 3 to 3 in (11), one can cover the CVK coefcients in the entire region of the triangular prism GIHAFB. Now assume that the input signal () of a nonlinear channel is a bandpass random multisine signal in the following form: () = =
= =

() cos(2 ( + 0 ) + ) Re{() 2( +0 ) } (17)

(12) (13) (14)

where the 2 +1 frequency bins are distributed symmetrically with respect to the carrier frequency , and () = () , = , . . . , are mutually independent (but not necessarily identical), zero-mean, complex random variables. The

1418

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

baseband representation of () is () =
=

By substituting (11) into (22) and (23), one can rewrite (22) and (23) as (18) [ () ()] = () [ () ()] 3 (, , ) [ () () ( ) ( )] (24) +
,, (+ +=)

() 20

where () is a complex-valued signal and can be expressed as () () . Using (18), one can rewrite (17) as () = Re{() 2 } = () cos(2 + ()) (19) This suggests that the bandpass random multisine signal dened in (17) (or equivalently, (19)) can be represented by (1) with () = () and () = (). The Fourier transform of (), say, ( ), is equal to ( ) =
=

[ () ( ) ( ) ()] = () [ () ( ) ( ) ()] + 3 ( , , )
, , ( + + =)

[ () ( ) ( ) ( ) ( ) ( )], , , , + + = (25)

() ( 0 )

(20)

Let () be the discrete-frequency counterpart of ( ) at = 0 ( an integer), then we see from (20) that () = (). Note that the spectral resolution of () is determined by 0 .

Dene 2 () = [()2 ], 4 () = [()4 ], and 6 () = [()6 ] for . We see that the term [ () ()] in (24) is equal to 2 () because of () = (). In addition, the term [ () () ( ) ( )] in (24) can be determined as follows: [ () () ( ) ( )] = = = = 4 (), 2 ()2 ( ), = = = , 0, otherwise

(26)

III. E STIMATION A LGORITHM FOR THE F REQUENCY-D OMAIN VOLTERRA K ERNELS OF N ONLINEAR BANDPASS C HANNELS In the following, we will derive the estimation algorithm for the frequency-domain Volterra kernels of nonlinear bandpass channels under the condition that the input signal is the random multisine signal dened in (18). Specically, we want to nd the optimal MMSE estimate of the Volterra kernels in (11) for the nonlinear bandpass channel. That is, to nd the optimal () and 3 (, , ) (where , , and + + = ) which minimize the cost function = = [()2 ] [ () () () 3 (, , ) () ( ) ( )2 ] (21)
,, (+ +=)

where it is understood that and (but not ) are interchangeable. In deriving (26), the constraint of + + = and the fact that (), = , . . . , are zero mean and mutually independent have been used. Using these results, one can rewrite (24) as follows: [ () ()] = 2 () () +4 ()2 (, )3 (, , ) 2 ( )2 (, )3 (, , ) +2 ()
=, =

= 2 () () + (4 () 2 2 ())2 (, ) 3 (, , ) + 2 ()3 () where { 2 (, ) = 3 () = 1, = 2, = 2 ( )2 (, )3 (, , ) (28) (29) (27)

where (11) has been used in deriving (21). According to the orthogonality principle [23], the necessary and sufcient condition for the cost function to attain its minimum value is when the error () is orthogonal to each of the input terms in (11). That is, the MMSE is attained when the following equations are simultaneously satised: [ ()()] = 0 [ () ( ) ( )()] = 0, , , , + + = (22) (23)

Note that the term 2 (, ) is needed in (27) and (29) because of the symmetry property (16). The term [ () ( ) ( ) ()] in (25) is equal to [ () () ( ) ( )] , hence it can also be determined by (26). The term [ () ( ) ( ) ( )

TSENG et al.: ESTIMATION OF CUBIC NONLINEAR BANDPASS CHANNELS IN ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING SYSTEMS

1419

( ) ( )] in (25) can be determined as follows: [ () ( ) ( ) ( ) ( ) ( )] = 4 ()2 ( ), = = = = = , , = ( ) ( ) ( ) , = = = = = , 2 2 2 , = ( ) ( ) , = = , 4 2 = = = = 2 ()2 ( )2 ( ), = = , = = = = (30) 6 (), = = = = = = 4 ()2 ( ), = = = = , = = 4 ()2 ( ), = = = = , = = 0, otherwise where it is understood that and (and also , ) are interchangeable. In deriving (30), the constraints of + + = and + + = and the fact that (), = , . . . , are zero mean and mutually independent have been used. Using (26) and (30), one can see that, in the case of = and = , eq. (25) becomes: [ () ( ) ( ) ()] = 2 ()2 ( )2 ( )2 (, )3 (, , ), = = , , = (31) ( ) ( ) ( , ) ( , , ) , = = , 4 2 2 3 , = In the case of (, , ) = (, , ) and = , eq. (25) becomes: [ () ( ) ( ) ()] = 2 ()2 ( ) () +4 ( )2 ()2 (, )3 (, , ) +4 ()2 ( )2 (, )3 (, , ) 2 ( )2 (, )3 (, , ) +2 ()2 ( )
=, =,

the CVK. It is straightforward to see from (31) that 3 (, , ) = [ () ( ) ( ) ()]/ 2 (, )2 ()2 ( )2 ( ), = = , , = (34) ( ) ( ) ( ) ( )] / ( ) ( ) , [ 4 2 = = , , = for + + = and , = . In Fig. 1, the line segments QE and CQ indicate the intersections of the plane + + = with the planes = and = , respectively. Note that (34) covers the CVK coefcients on the pentagon SEDCT in Fig. 1 except for those coefcients on the line segments QE and CQ. To obtain the optimal estimates of CVK coefcients on the line segment QE other than the point Q (i.e., the region where + + = , = , and = ), one can subtract (27)2( ) from (32) to obtain [ () ( ) ( ) ()] 2 ( ) [ () ()] = (4 ( )2 () 2 ()2 2 ( ))2 (, )3 (, , ), (35) which leads to 3 (, , ) = { [ () ( ) ( ) ()] 2 ( ) [ () ()]}/(4 ( )2 () 2 ()2 2 ( ))2 (, ) (36) For those points on the line segment CQ other than Q (i.e., the region where + + = , = , and = ), they correspond to the CVK coefcients 3 (, , ), = + 1, . . . , . Due to the symmetry property 3 (, , ) = 3 (, , ), we see that these points can also be solved via (36). For the CVK coefcient at the point Q (i.e., the point (, , ) = (, , )), it can be solved by subtracting (27)4() from (33)2 (), which yields 2 () [ () () () ()] 4 () [ () ()] = (6 ()2 () 2 4 ())2 (, )3 (, , ), (37) This leads to 3 (, , ) = {2 () [ () () () ()] 4 () [ () ()]}/(6 ()2 () 2 4 ()) (38) This takes care of the points Q in Fig. 1. By using (34), (36), and (38), one can obtain the optimal CVK coefcients for the entire pentagon SEDCT in Fig. 1. By varying from 3 to 3 , one can cover the entire CVK estimation region (i.e., the triangular prism GIHAFB) shown in Fig. 1. Finally, the LVK can be solved by substituting (38) into (27), which leads to () = {(6 () 4 ()2 ()) [ () ()] (4 () 2 2 ()) [ () () () ()]}

= 2 ()2 ( ) () + (4 ( )2 () 2 ()2 2 ( )) 2 (, )3 (, , ) + (4 ()2 ( ) 2 (, )3 (, , ) + 2 ()2 ( )3 ()

2 2 ()2 ( ))

(32)

In the case of (, , ) = (, , ), eq. (25) becomes: [ () () () ()] = 4 () () +6 ()2 (, )3 (, , ) +4 () 2 ( )2 (, )3 (, , )


=, =

4 () () + (6 () 4 ()2 ()) 2 (, )3 (, , ) + 4 ()3 () (33)

/(6 ()2 () 2 4 ()) 3 (), (39)

The optimal Volterra kernel estimates can be obtained by solving (27) and (31)-(33). First we derive the formulas for

Note that 3 () can be obtained from (29) once (36) and (38) are available. The complete algorithm is summarized in Table I.

1420

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

TABLE I T HE OPTIMAL MMSE ALGORITHM FOR ESTIMATING THE V OLTERRA KERNELS OF NONLINEAR BANDPASS CHANNELS . I T CAN BE USED FOR OFDM SYSTEMS EMPLOYING QAM S WITH ALPHABET SIZES LARGER THAN 4. [ () ( ) ( ) ()] /2 (, )2 ()2 ( )2 ( ), = = , , = 3 (, , ) = , [ () () ( ) ()]/4 ()2 ( ), = = , = 3 3 3 (, , ) = ()] 2 ( ) [ () ()]}/(4 ( )2 () 2 ()2 2 ( )) 2 (, ), , = , 3 (, , ) = {2 () [ () () () ()] 4 () [ () ()]}/(6 ()2 () 2 4 ()), 3 () =
=

where is a complex data symbol taken from the -th symbol alphabet. Each symbol alphabet could be either a QAM or a PSK of a different size. The symbol interval of the OFDM is denoted by . Note that, by choosing the frequency spacing of adjacent subcarriers to be 1/ , the subcarriers are orthogonal over the symbol interval . By up-converting the baseband signal () to the carrier frequency , the transmitted signal for the bandpass channel can be expressed as: () = Re{() 2 } =
=

{ [ () ( ) ( )

Re{ 2( +/ ) } (41)

2 ( )2 (, )3 (, , ),

() = {(6 () 4 ()2 ()) [ () ()] (4 () 2 2 ()) [ () () () ()]} ( )) (), /(6 ()2 () 2 3 4

Note that the algorithm in Table I may look similar to but is fundamentally different from that in [16]. Specically, the algorithm in [16] was derived based on the assumption of the input being i.i.d., zero-mean, and circularly-symmetric complex Gaussian, and is for estimating the time-domain Volterra kernels when the input is QAM or PSK. The algorithm in Table I, on the other hand, is based on the assumption of the input being a random multisine signal with mutually independent zero-mean frequency components, and is applicable to estimate the frequency-domain Volterra kernels when the input is OFDM (to be seen in Section IV). The mutual independence assumption (on the frequency components) works for the algorithm in Table I in the same way as the i.i.d. assumption (on the time-domain input) does for the algorithm in [16]. This causes the two algorithms for different occasions to look similar. IV. E STIMATION OF N ONLINEAR BANDPASS C HANNELS FOR OFDM S YSTEMS In the OFDM modulation scheme, a group of equallyspaced orthogonal subcarriers are used to transmit data. The data to be transmitted are distributed over the subchannels. In each subchannel, the corresponding subcarrier is modulated with a common modulation scheme like QAM or PSK. The summation of all the subchannel signals forms the baseband OFDM signal. This baseband OFDM signal is then further up-converted to a carrier frequency to form the transmitted signal for the bandpass channel. Suppose there are a total of 2 + 1 subcarriers with frequencies / , = , . . . , , then the baseband OFDM signal can be expressed as: () =
=

Comparing (40) and (41) to (18) and (17), respectively, one can see that the OFDM signal is indeed the random multisine signal dened in Section II with = 1/ . Note that, since each subchannel performs the modulation independently, the data symbols at different subcarriers are mutually independent. Furthermore, for the usual QAM and PSK modulation schemes used in the OFDM system, the signal constellation points are arranged symmetrically with respect to the origin on the complex plane. Suppose that the symbols in the alphabet of each subchannel are taken with an equal probability, then one can easily see that the data symbols satisfy [ ] = 0, for = , . . . , . Based on the above analysis, one would expect that the algorithm shown in Table I can be used to estimate nonlinear bandpass channels for OFDM systems. This is indeed true for OFDM systems using only QAMs with alphabet sizes larger than 4 for subcarriers. For OFDM systems using only PSKs or a mixture of QAMs and PSKs, certain CVK coefcients would degenerate to the LVK and thus can not be estimated by the algorithm shown in Table I. Such a degeneration effect can be explained as follows. Let PSK and QAM denote the sets of frequencies in which their symbols are taken from the various PSK and QAM alphabets, respectively. For any frequency PSK , the cubic term 3 (, , ) () ( ) ( ) in (11) becomes 3 (, , ) () ( ) ( ) =
2 3 (, , ) ()( ) ( ) = 3 (, , ) () (42)

where is the constant amplitude of the PSK alphabet for frequency . We see that the cubic term has degenerated into a linear term. On the contrary, for any frequency QAM , such degeneration will not occur because the symbols of a QAM modulation do not have a constant amplitude. One can divide 3 () in (29) into two terms as follows: 3 () where
() = 3 3 () = = 3 () + 3 ()

(43)

QAM

2 ( )2 (, )3 (, , ) (44) 2 ( )2 (, )3 (, , ) (45)

PSK Note that 3 () contains nondegenerative CVK coefcients, while 3 () contains degenerative CVK coefcients. Using

2/ ,

(40)

TSENG et al.: ESTIMATION OF CUBIC NONLINEAR BANDPASS CHANNELS IN ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING SYSTEMS

1421

TABLE II T HE OPTIMAL V OLTERRA KERNEL ESTIMATION ALGORITHM FOR OFDM SYSTEMS EMPLOYING A MIXTURE OF QAM S AND PSK S .

TABLE III T HE OPTIMAL V OLTERRA KERNEL ESTIMATION ALGORITHM FOR OFDM SYSTEMS EMPLOYING ONLY PSK S . [ () ( ) ( ) ()]/22 ()2 ( )2 ( ), = = , , = 3 (, , ) = , [ () () ( ) ()]/4 ()2 ( ), = = , , = 3 3 () = [ () ()]/2 (),

[ () ( ) ( ) ()]/22 ()2 ( )2 ( ), = = , , = 3 (, , ) = , [ () () ( ) ()]/4 ()2 ( ), = = , , = 3 3 3 (, , ) = { [ () ( ) ( ) ()] 2 ( ) [ () ()]}/(4 ( )2 () 2 ()2 2 ( ))2 (, ), QAM , = , 3 (, , ) = {2 () [ () () () ()] 4 () [ () ()]}/(6 ()2 () 2 4 ()), QAM , 3 () = QAM 2 ( )2 (, )3 (, , ), {(6 () 4 ()2 ()) [ () ()] (4 () 2 2 ()) [ () () () ()]} , () = 2 ()) (), ( ) ( ) QAM / ( 6 2 4 3 [ () ()]/2 () PSK 3 (),

degenerative and the quantity 4 () 2 2 () in (27) is equal 4 2 2 to ( ) = 0. Therefore, eq. (27) becomes [ () ()]
= 2 ()() + 2 ()3 () (49)

Hence, the effective LVK coefcient () can be estimated as follows: () =


[ () ()]/2 () 3 ()

(50)

(42), one can calculate the total degeneration to output frequency , say, (), as follows: 2 () = 2 (, )3 (, , ) () =
PSK 3 () ()

This result is also shown in Table II. For OFDM systems employing only PSKs, all CVK coefcients in the form of 3 (, , ) (including = ) are degenerative and do not need to be estimated. Meanwhile, the nondegenerative term 3 () is equal to zero. Therefore, the algorithm in Table II is further simplied to that in Table III. V. C OMPUTER S IMULATION A simulated nonlinear channel was used to demonstrate the effectiveness of the proposed algorithms. The time-domain baseband equivalent Volterra model for the nonlinear channel was as following: () =
=0

(46)

2 where the symmetry property (16) and the fact that 2 ( ) = for PSK have been used in deriving (46). The resulting linear output at frequency , say, (), after considering the total degeneration, becomes

()( ) +

3 (, , ) (51)

() where

= () () + () ()) () = () () (47) = ( () + 3

( )( ) ( ) + ()

=0 =0 =0

() =

() () + 3

(48)

where the values of () and 3 (, , ) were chosen as those used to model the traveling wave tube (TWT) power amplier for satellite communications up to the third order [8]. The kernel coefcients are as follows: 1 (0) = 1 (1) = 1 (2) = 1 (3) = 3 (0, 0, 2) = 3 (3, 3, 0) = 3 (0, 0, 1) = 3 (0, 0, 3) = 3 (1, 1, 0) = 1.22 + 0.646 0.063 0.001 0.024 0.014 0.036 + 0.031 0.039 0.022 0.018 0.018 0.035 0.035 0.040 0.009 0.010 0.017 (52) (53) (54) (55) (56) (57) (58) (59) (60)

is the effective LVK coefcient at frequency after considering the degeneration. Using (43) and (48), one can see that (27), (32), and (33) still hold if we replace () and 3 () therein with () and 3 (), respectively. By performing the replacement and then following the same derivation procedure in Section III, one can obtain the kernel estimation algorithm for OFDM systems employing a mixture of QAMs and PSKs. The derived algorithm is shown in Table II. By comparing Table II to Table I, we see that, In Table II, the estimation of 3 (, , ) is only necessary when QAM (i.e., when 3 (, , ) is nondegenerative), and the estimation of 3 (, , ) is only necessary when QAM (i.e., when 3 (, , ) is nondegenerative). Meanwhile, the term 3 () in Table I has been replaced () in Table II. In addition, by the nondegenerative term 3 when PSK , the CVK coefcient 3 (, , ) is

The SNR was set to 20 dB. The input signal to the channel was an OFDM signal with a baseband equivalent representation of (40). The number of subcarriers in the OFDM signal was 5 (corresponding to = 2 in (40)), and the subcarrier spacing was set to 2 kHz (i.e., 1/ = 2 kHz). This signal was sampled

1422

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

TABLE IV T HE NMSE S OF THE ESTIMATED TOTAL , LINEAR , AND CUBIC V OLTERRA KERNELS ACHIEVED BY VARIOUS METHODS . T HE NOTATION MEANS REALIZATIONS OF DATA WERE USED . NMSE QAM1 1 QAM1 10 QAM1 100 QAM1 1000 QAM2 1 MIX 1 PSK 1 GAU 1000 Total 3.65 100 8.80 102 1.91 102 6.74 103 1.15 103 6.61 107 1.57 107 2.69 100 Linear 3.85 100 9.28 102 2.01 102 7.13 103 1.21 103 6.47 107 1.55 107 2.70 100 Cubic 1.73 101 5.34 103 2.14 103 1.41 104 2.33 104 1.06 104 3.03 105 1.10 100

10 8 6 4 2 0 2 4 6
: f=-2/T : f=-1/T : f=0 : f=1/T : f=2/T

at a frequency of 16 2 = 32 kHz to obtain the input data (). In the rst simulation, the subcarriers of the OFDM signal were modulated with 8QAM schemes of different amplitudes. The constellations of the ve 8QAM schemes used by the ve subcarriers are shown in Fig. 2. Note that any QAM schemes can be used here, we chose these particular ones merely for easy visualization with Fig. 2. In each subchannel, the input data were randomly chosen from the corresponding 8QAM symbols with an equal probability. We collected 1000 realizations of the OFDM input signal, each realization had a length of 32k symbol intervals. This signal was fed to the simulated channel (51) to yield the channel output. The input and output signals were used by the algorithm shown in Table I to estimate the frequency-domain Volterra kernels of the simulation channel. The estimated frequency-domain Volterra kernels were compared to the actual ones calculated from (52)-(60). The normalized mean square errors (NMSEs) of the estimated total, linear, and cubic Volterra kernels obtained by using various numbers of data realizations are shown in Table IV (denoted by QAM1). We see that the NMSEs decrease consistently with the number of data realizations. This agrees with our expectation, because, in theory, the higher-order auto-moment spectral properties described in (26) and (30) hold when the amount of random data approaches innity. With a nite amount of random data, the estimated higher-order auto-moment spectra can only satisfy (26) and (30) approximately. The closer the amount of random data is to innity, the better the estimated higher-order auto-moment spectra satisfy (26) and (30), hence the more accurate the estimated Volterra kernels become. To deal with this problem, one can, instead of choosing input data randomly, use a carefully designed input data sequence which satises (26) and (30) exactly. Note that the OFDM signal contains multiple parallel QAM symbol streams. Each OFDM symbol interval contains a particular combination of parallel QAM symbols. One can choose the input parallel QAM symbols to the subcarriers so that every possible combination of parallel QAM symbols appears in the same number of OFDM symbol intervals in the input data. Since every possible combination appears the same number of times in the input data, the estimated higherorder auto-moment spectra from the data would satisfy (26) and (30) exactly. In this simulation, we had 5 subcarriers, each was modulated with an 8QAM. The number of possible combinations of parallel QAM symbols was 85 = 32k.

8 10 15 10 5 0 5 10 15

Fig. 2. The constellations of the ve 8QAMs used by the subcarriers of the OFDM signal in the rst simulation.

1.6 1.55 1.5 1.45 : Actual LVK : Estimated LVK

|H(m)|

1.4 1.35 1.3 1.25 1.2 2 -1 0 1 Discrete Frequency m 2

Fig. 3.

The actual LVK and the estimated LVK by the QAM2 method.

Therefore, we constructed a realization of OFDM signal which contained 32k symbol intervals. In this realization of data, each possible combination of parallel QAM symbols appeared exactly once. These data were then used by the algorithm shown in Table I to estimate the frequency-domain Volterra kernels. The result is shown in Table IV (denoted by QAM2). We see that the NMSEs achieved by the QAM2 method with only one realization of data are approximately equal to (if not better than) those achieved by the QAM1 method with 1000 realizations of data. In addition, the estimated LVK and CVK by the QAM2 method are compared to the actual LVK and CVK in Figs. 3 and 4, respectively. In Fig. 4, the CVK coefcients are arranged into a vector to conduct the comparison. The close match between the actual and the estimated kernels indicates that excellent estimates were obtained by the QAM2 method. In the second simulation, the simulation setting was basically the same as that in the QAM2 case except that two out of the ve subcarriers were modulated with 8PSKs instead of 8QAMs. The constellations of the PSKs and QAMs used

TSENG et al.: ESTIMATION OF CUBIC NONLINEAR BANDPASS CHANNELS IN ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING SYSTEMS

1423

0.18 0.16 0.14 : Actual CVK : Estimated CVK

10 8 6 4 2 0 2 4
0 10 20 30 40 50 60 70

: f=-2/T : f=-1/T : f=0

: f=1/T : f=2/T

|H3 (i,j,k)|

0.12 0.1 0.08 0.06 0.04 0.02 0

6 8 10 10 8 6 4 2 0 2 4 6 8 10

Index Number
Fig. 4. The actual CVK and the estimated CVK by the QAM2 method.

10 8 6 4 2 0 2 4 6 8 10 10 8 6 4 2 0 2

: f=-2/T : f=-1/T : f=0

: f=1/T : f=2/T

Fig. 6. The constellations of the ve 8PSKs used by the subcarriers of the OFDM signal in the third simulation.

10

Fig. 5. The constellations of the three 8QAMs and two 8PSKs used by the subcarriers of the OFDM signal in the second simulation.

in this simulation are shown in Fig. 5. Since a mixture of QAMs and PSKs was adopted in this case, the algorithm shown in Table II was used to estimated the frequency-domain Volterra kernels. The result is shown in Table IV (denoted by MIX). Note that, in the MIX result, the NMSE of the LVK was calculated by comparing the estimated LVK to the effective LVK dened in (48). In addition, the NMSE of the CVK was calculated by comparing only nondegenerative CVK coefcients. The goodness of the algorithm in Table II is justied by the small values of the NMSEs shown in Table IV under MIX. In the third simulation, the simulation setting was again the same as that in the QAM2 case except that the ve subcarriers were modulated with 8PSKs instead of 8QAMs. The constellations of the PSKs used in this simulation are shown in Fig. 6. Since only PSKs were used in this case, the algorithm shown in Table III was used to estimated the

frequency-domain Volterra kernels. The result is shown in Table IV (denoted by PSK). Similarly to the MIX case, in the PSK result, the NMSE of the LVK was calculated by comparing the estimated LVK to the effective LVK dened in (48), and the NMSE of the CVK was calculated by comparing only nondegenerative CVK coefcients. The small NMSEs suggest that excellent estimates were obtained. This justies the effectiveness of the algorithm in Table III. In the above simulations, an 8-point signal constellation was used by each subcarrier. In fact, signal constellations with a larger number of points can also be used by the subcarriers. In this case, due to the larger number of symbol alphabets, a larger amount of data would be required for the estimated higher-order auto-moment spectra to satisfy (26) and (30). This implies that, to achieve the same NMSE performance, OFDM systems using a higher-level modulation for its subchannels would require a larger amount of data. In addition, for simplicity, we adopted a relatively small number of OFDM subcarriers for the simulations. One should note that the proposed algorithms can also be applied to OFDM systems with a larger number of subcarriers. However, the number of involved frequency-domain Volterra kernel coefcients in (24) and (25) will increase with the number of subcarriers, so will the number of involved higher-order auto-moment spectra in (24) and (25). This implies that, when the number of subcarriers increases, more higher-order auto-moment spectral properties described in (26) and (30) need to be satised to justify the proposed algorithms. This, in turn, suggests that more data would be required by the proposed algorithms to attain good estimates. One may notice that, as the number of subcarriers increases, the baseband OFDM signal will asymptotically approach i.i.d. complex Gaussian. Under the circumstances, methods developed in [17], [18] can be used. To conduct a performance comparison, the method in [18] (referred to as the Gaussian (GAU) method hereafter) was

1424

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 5, MAY 2010

TABLE V T HE NMSE S OF THE ESTIMATED TOTAL , LINEAR , AND CUBIC V OLTERRA KERNELS IN THE IEEE 802.11 A SIMULATION . T HE NUMBER OF SUBCARRIERS IS 52 AND THE SUBCARRIER MODULATION SCHEME IS 16 QAM. NMSE PRO GAU-2 GAU-51 Total 9.93 101 2.13 102 2.80 100 Linear 2.57 102 1.34 102 4.10 101 Cubic 9.99 101 4.03 102 8.60 100

tested in the same experimental setup as the QAM1 in Table IV. The result is also shown in Table IV (denoted by GAU). We see that the NMSEs achieved by the GAU method using 1000 realizations of data is far worse than those achieved by the QAM1 method with the same amount of data. This is probably because that the OFDM signal with this small number of subcarriers did not quite satisfy the i.i.d. complex Gaussianity assumption. We further compared the proposed method with the GAU method using the nonlinear channel in [17] with an OFDM signal composed of 52 subcarriers (like that dened in the IEEE 802.11a standard). The modulation scheme employed in each subcarrier was 16 QAM. The total length of the OFDM signal was 10,000 OFDM symbol intervals. Note that the number of kernel coefcients involved in the GAU method is ( + 1) + ( + 1)2 ( + 2)/2, where is the memory length of the time-domain Volterra model. Since a model with = 2 would accommodate the memory of the channel in [17], only 21 kernel coefcients needed to be estimated in the GAU method. On the other hand, the number of kernel coefcients involved in the proposed method is + 2 ( + 1)/2, where is the length of the supported frequency range in the frequency-domain Volterra model. With the short memory length of the channel in [17], incorporating all the 52 input frequency components in the frequency-domain Volterra model was required. The total number of kernel coefcients to be estimated was therefore 52 + 522 53/2 = 71708. The estimation results by the GAU and the proposed methods are shown in Table V (denoted by GAU-2 and PRO, respectively). We see that, the GAU2 case achieved reasonably small NMSEs for the linear and cubic kernels. The PRO case, however, fell short on the NMSE for the cubic kernel. Given the signicantly larger number of kernel coefcients in the PRO case, one shouldnt be surprised by its less accurate result. To make the comparison fairer, we assumed that the channel memory were expected to be long and hence a time-domain Volterra model with = 51 had to be used by the GAU method. In this case, the GAU method had the same number of kernel coefcients to be estimated as PRO. The estimation result for this case is also shown in Table V (denoted by GAU-51). The larger NMSEs obtained by GAU-51 suggest that GAU-51 performed worse than PRO. One can expect that the NMSEs of GAU-51 and PRO would be improved by increasing the amount of input and output data. Through the above comparisons, we see that, for OFDM systems with a small number of subcarriers, the proposed method outperforms the GAU method. The GAU method is more advantageous for OFDM systems with a large number

of subcarriers and with short memory, because in this case the number of kernel coefcients need to be estimated in the GAU method is signicantly smaller than that for the proposed method. However, this advantage diminishes when the channel memory lengthens. With the same amount of data and the same number of kernel coefcients to be estimated, the proposed method seems to be more accurate than the GAU method, as indicated by the PRO and GAU-51 cases shown in Table V. As far as computational complexity is concerned, the proposed method needs to estimate the cross-moment spectra between the input and output, while the GAU method needs to estimate the cross-cumulants between the input and output. In addition, the algorithm of the proposed method requires ( 3 + 5 2 + 8 )/2 multiplications and 2 2 + additions, while the algorithm of the GAU method requires 2( + 1) multiplications and ( + 1)2 additions [18]. When and are about the same, the algorithm of the GAU method is more computationally efcient. However, estimating the cumulants requires more computations than estimating the moments. When the amount of data increases, estimation of the cumulants (or moments) may dominate the computational complexity. In this case, the proposed method may become more computationally efcient than the GAU method. VI. C ONCLUSION In this paper, we have derived a novel method for estimating the frequency-domain Volterra kernels of nonlinear bandpass channels using random multisine signals. We assumed that the various tones in the multisine signal were modulated by independent (but not necessarily identical), zero-mean, complex random variables. By exploring properties of the higher-order auto-moment spectra of the random multisine signal, an optimal MMSE Volterra kernel estimation algorithm was derived. We have recognized that the OFDM signal ts in the class of random multisine signal described in this paper, hence the proposed method can be used to estimate nonlinear bandpass channels in OFDM systems. Meanwhile, we have presented a technique for designing an OFDM test signal which guarantees the attainment of the optimum MMSE solution. Algorithms suitable for OFDM systems employing pure QAMs, pure PSKs, and a mixture of QAMs and PSKs have been derived, and their performances have been veried. The proposed method can indeed serve as a simple and accurate method for estimating the frequency-domain Volterra kernels of nonlinear bandpass channels. Volterra model based nonlinear channel compensators can benet from the proposed method. ACKNOWLEDGEMENT This work was supported in part by the National Science Council under Contract NSC 96-2221-E-019-005. R EFERENCES
[1] S. C. Cripps, RF Power Ampliers for Wireless Communications. Norwood, MA: Artech House, 1999. [2] B. Umasankar and X. Fernando, OFDM impairment mitigation techniques, in Proc. 2nd Int. Conf. Access Networks Workshops, Aug. 2007, pp. 16.

TSENG et al.: ESTIMATION OF CUBIC NONLINEAR BANDPASS CHANNELS IN ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING SYSTEMS

1425

[3] A. R. S. Bahai, B. R. Saltzberg, and M. Ergen, Multi-Carrier Digital Communications Theory and Applications of OFDM, 2nd edition. New York: Springer, 2004. [4] K. G. Gard, H. M. Gutierrez, and M. Steer, Characterization of spectral regrowth in microwave ampliers based on the nonlinear transformation of a complex Gaussian process, IEEE Trans. Microw. Theory Tech., vol. 47, no. 7, pp. 10591069, July 1999. [5] S.-J. Yi, S. Nam, S.-H. Oh, and J.-H. Han, Prediction of a CDMA output spectrum based on intermodulation products of two-tone test, IEEE Trans. Microw. Theory Tech., vol. 49, no. 5, pp. 938946, May 2001. [6] G. T. Zhou, H. Qian, L. Ding, and R. Raich, On the baseband representation of a bandpass nonlinearity, IEEE Trans. Signal Process., vol. 53, no. 8, pp. 29532957, Aug. 2005. [7] G. T. Zhou and J. S. Kenney, Predicting spectral regrowth of nonlinear power ampliers, IEEE Trans. Commun., vol. 50, no. 5, pp. 718722, May 2002. [8] S. Benedetto and E. Biglieri, Nonlinear equalization of digital satellite channels, IEEE J. Sel. Areas Commun., vol. 1, no. 1, pp. 5762, Jan. 1983. [9] E. Biglieri, A. Gersho, R. D. Gitlin, and T. L. Lim, Adaptive cancellation of nonlinear intersymbol Interference for voiceband data transmission, IEEE J. Sel. Areas Commun., vol. 2, no. 5, pp. 765777, Sep. 1984. [10] E. Biglieri, S. Barberis, and M. Catena, Analysis and compensation of nonlinearities in digital transmission systems, IEEE J. Sel. Areas Commun., vol. 6, no. 1, pp 4251, Jan. 1988. [11] G. Lazzarin, S. Pupolin, and A. Sarti, Nonlinear compensation in digital radio systems, IEEE Trans. Commun., vol. 42, no. 2/3/4, pp. 988999, Feb./Mar./Apr. 1994. [12] L. Ding, G. T. Zhou, D. R. Morgan, Z. Ma, J. S. Kenney, J. Kim, and C. R. Giardina, A robust predistorter constructed using memory polynomials, IEEE Trans. Commun., vol. 52, no. 1, pp. 159165, Jan. 2004. [13] S. P. Stapleton and F. C. Costescu, An adaptive predistortion for a power amplier based on adjacent channel emissions, IEEE Trans. Veh. Technol., vol. 41, no. 1, pp. 4956, Jan. 1992. [14] S. Benedetto and E. Biglieri, Principles of Digital Transmission with Wireless Applications. Norwell, MA: Kluwer, 1999.

[15] C.-H. Tseng and E. J. Powers, Identication of nonlinear channels in digital transmission systems, in Proc. IEEE Signal Process. Workshop Higher-Order Statistics, June 1993, pp. 4245. [16] C.-H. Cheng and E. J. Powers, Optimal Volterra kernel estimation algorithms for a nonlinear communication system for PSK and QAM inputs, IEEE Trans. Signal Process., vol. 49, no. 1, pp. 147163, Jan. 2001. [17] A. J. Redfern and G. T. Zhou, Nonlinear channel identication and equalization for OFDM systems, in Proc. IEEE Int. Conf. Acoust. Speech, Signal Process., May 1998, vol. 6, pp. 35213524. [18] G. Mileounis, P. Koukoulas, and N. Kalouptsidis, Input-output identication of nonlinear channels using PSK, QAM, and OFDM inputs, Signal Process., vol. 89, no. 7, pp. 13591369, July 2009. [19] P. Crama and J. Schoukens, Computing an initial estimate of a WienerHammerstein system with a random phase multisine excitation, IEEE Trans. Instrum. Meas., vol. 54, no. 1, pp. 117122, Feb. 2005. [20] M. Solomou, D. Rees, and N. Chiras, Frequency domain analysis of nonlinear systems driven by multiharmonic signals, IEEE Trans. Instrum. Meas., vol. 53, no. 2, pp. 243250, Apr. 2004. [21] J. C. Pedro and N. B. Carvalho, Designing multisine excitations for nonlinear model testing, IEEE Trans. Microw. Theory Tech., vol. 53, no. 1, pp. 4554, Jan. 2005. [22] C.-H. Tseng, Identication of nonlinear systems driven by random multisine signals, in Proc. IASTED Int. Conf. Signal Image Process., Aug. 2007. [23] S. Haykin, Adaptive Filter Theory, 2nd edition. Englewood Cliffs, NJ: Prentice-Hall, 1991. Ching-Hsiang Tseng received the B.S. and M.S. degrees in electrical engineering from National Taiwan University, Taipei, Taiwan, ROC, in 1987 and 1990, respectively, and the Ph.D. degree in electrical and computer engineering from the University of Texas at Austin, Austin, TX, in 1993. In August 1993, he joined the Electronics Research Center, University of Texas, Austin, TX, as a Post-Doctoral Fellow. Since August 1994, he has been with National Taiwan Ocean University, where he is a Professor of Electrical Engineering. His current research interests include higher-order statistical signal processing, nonlinear signals and systems, signal processing for communications, and underwater acoustic communications. Dr. Tseng is the Executive Editor of the Journal of Marine Science and Technology-Taiwan.

S-ar putea să vă placă și