IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO.

5, MAY 1999

837

An Adaptive Modulation Scheme for Simultaneous Voice and Data Transmission over Fading Channels
Mohamed-Slim Alouini, Member, IEEE, Xiaoyi Tang, and Andrea J. Goldsmith, Member, IEEE
AbstractWe propose a new adaptive modulation technique for simultaneous voice and data transmission over fading channels and study its performance. The proposed scheme takes advantage of the time-varying nature of fading to dynamically allocate the transmitted power between the inphase (I ) and quadrature (Q) channels. It uses xed-rate binary phase shift keying (BPSK) modulation on the Q channel for voice, and variable-rate M ary amplitude modulation (M -AM) on the I channel for data. For favorable channel conditions, most of the power is allocated to high rate data transmission on the I channel. The remaining power is used to support the variable-power voice transmission on the Q channel. As the channel degrades, the modulation gradually reduces its data throughput and reallocates most of its available power to ensure a continuous and satisfactory voice transmission. The scheme is intended to provide a high average spectral efciency for data communications while meeting the stringent delay requirements imposed by voice. We present closed-form expressions as well as numerical and simulation results for the outage probability, average allocated power, achievable spectral efciency, and average bit error rate (BER) for both voice and data transmission over Nakagami-m fading channels. We also discuss the features and advantages of the proposed scheme. For example, in Rayleigh fading with an average signal-to-noise ratio (SNR) of 20 dB, our scheme is able to transmit about 2 Bits/s/Hz of data at an average BER of 1005 while sending about 1 Bit/s/Hz of voice at an average BER of 1002 . Index Terms Adaptive modulation techniques, integrated voice and data systems, Nakagami fading.

I. INTRODUCTION HE RADIO spectrum available for wireless communications is extremely scarce, while demand for mobile and personal communications is growing at a rapid pace. Spectral efciency is therefore of primary concern in the design of future wireless communications systems. Furthermore, these systems will have to support not only voice services but also
Manuscript received January 1997; revised January 1999. The work of M.-S. Alouini was supported in part by a National Semiconductor Graduate Fellowship Award and in part by the Ofce of Naval Research under Grant NAV-5X-N149510861. The work of X. Tang was supported by a Summer Undergraduate Fellowship (SURF) award. This is an expanded version of work which was presented at the IEEE Vehicular Technology Conference (VTC98), Ottawa, Ont., Canada, May 1998. M.-S. Alouini was with the Communications Group, Department of Electrical Engineering, California Institute of Technology, Pasadena, CA 91125 USA. He is now with the Department of Electrical and Computer Engineering, University of Minnesota, Minneapolis, MN 55455 USA (e-mail: alouini@ece.umn.edu). X. Tang is with the Communications Group, Department of Electrical Engineering, California Institute of Technology, Pasadena, CA 91125 USA (e-mail: xiaoyi@systems.caltech.edu. A. J. Goldsmith is with the Department of Electrical Engineering, Stanford University, Stanford, CA 94305 USA (e-mail: andrea@ee.stanford.edu). Publisher Item Identier S 0733-8716(99)03084-X.

data services including facsimile, le transfer, e-mail, and Internet access. The need for spectrally efcient communication has recently led to the development of adaptive transmission techniques. These techniques take advantage of the time-varying nature of wireless channels to vary the transmitted power level [1], symbol rate [2], coding rate/scheme [3], constellation size [4][8], or any combination of these parameters [9][14]. Their goal is to improve the link average spectral efciency ( [Bits/s/Hz]), dened as the average transmitted data rate per unit bandwidth for a specied average carrier-to-noise ratio (CNR) and bit error rate (BER). Good performance of these schemes requires accurate channel estimation at the receiver and a reliable feedback path between the estimator and the transmitter. Buffering of the input data may also be required, since the outage probability of such schemes can be quite high, especially for channels with low average CNR. In general, voice transmission has low data rate requirements with real-time delay constraints, while data transmission demands higher rates with less stringent delay requirements. This suggests that xed-rate transmission combined with power adaptation, where the transmitter adjusts its power to maintain a constant CNR at the receiver, is well suited to voice, while bursty variable-rate transmission, which maximizes average spectral efciency, is best suited to data communication. In addition, voice and data typically have very different BER requirements which must be incorporated into their respective transmission schemes. Considerable research efforts have been devoted in recent years for the integration of voice and data for wireline [15] and wireless communication systems [16][20]. For the latter systems these efforts focused on the development of a variety of media access control (MAC) techniques and protocols such as packet reservation multiple access (PRMA), idle signal multiple access for integrated services (I-ISMA), and dynamic time division multiple access (D-TDMA). In this paper we propose a new hybrid adaptive scheme which supports simultaneous voice and data over fading channels.1 Contrary to the MAC solutions, the proposed scheme offers a link layer solution to the voice and data integration problem by designing the transmitted signal modulation to support their respective delay, data rate, and BER requirements. In particular, the proposed adaptive scheme responds to the fading channel uctuations by
1 More generally the proposed scheme is capable of handling two independent information streams which are inherently different: i.e., they may be generated by different sources and may also differ in their delay and BER requirements.

07338716/99$10.00 1999 IEEE

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

838

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 5, MAY 1999

giving priority voice communication in one of the quadrature channels, while devoting the other quadrature component to variable-rate data communication. For bad channel conditions, most of the transmitted power is allocated to ensure continuous and satisfactory transmission of speech communications. As the channel conditions improve, most of the transmitted power is reallocated to high data rate transmission. Hence the goal of the scheme is to provide a high average spectral efciency for data communications while meeting the stringent delay requirements of speech communications. The power allocation, as well as the constellations selection for the proposed schemes constant-power, will be discussed in more detail in Section III. The remainder of this paper is organized as follows. Section II describes the channel model. Section III presents the details of the proposed scheme. The performance of this scheme, assuming perfect channel estimation and negligible time delay, is analyzed in Section IV. In particular closed-form expressions for the outage probability, average allocated power, achievable spectral efciency, and average BER for both voice and data transmission are derived. Numerical and simulation results that allow discussion of the behavior of the proposed scheme are also presented. Our conclusions are given in Section V. II. CHANNEL MODEL We consider a slowly varying at-fading channel changing at a rate much slower than the symbol data rate, so the channel remains roughly constant over hundreds of symbols. We assume that the multipath fading environment is characterized by the Nakagami- probability density function (PDF). Hence the channel fading amplitude is given by [22, Eq. (11)] (1) is the average received power, is the where , and is the Nakagami fading parameter gamma function dened by [23, p. 942, Eq. (8.310.1)] (2) , Given the channel fading amplitude , a signal power [Hz], and a noise power density a signal bandwidth of [W/Hz], let us dene the CNR . By of using a standard transformation of random variables, it can be shown that the CNR is distributed according to a gamma , given by distribution, (3) where is the average CNR. We use the Nakagami- distribution since it can represent a range of multipath channels via the parameter [22], which can be interpreted as the amount of fading on the channel: as increases the amount of fading on the channel decreases. In particular, the Nakagami- distribution includes the one, which corresponds to sided Gaussian distribution ( ) worst-case fading) and the Rayleigh distribution (

as special cases. Furthermore, the Nakagami- distribution closely approximates the Nakagami- (Hoyt) [22, Eq. (59)] and the Nakagami- (Rice) [22, Eq. (56)] distributions. Finally, and perhaps most importantly, the Nakagami- distribution often gives the best t to land-mobile [24][26], indoor-mobile [27] multipath propagation, as well as scintillating ionospheric satellite radio links [28][32]. III. PROPOSED MODULATION SCHEME The proposed modulation scheme is a generalized and adaptive version of the unbalanced quadrature phase shift keying (UQPSK) [33], [34, p. 622], which also offers the capability of handling two different types of data. For instance, UQPSK was used by the space shuttle and the tracking and data relay satellite system (TDRSS) to communicate scientic data on the inphase ( ) channel and operational/telemetry data on the quadrature ( ) channel. In our case we propose to devote the channel to data communications while transmitting voice over channel. In contrast to UQPSK, where binary phase the shift keying (BPSK) modulation is used on both channels, channel for voice and our scheme uses BPSK on the ary amplitude modulation ( -AM) [34, p. 219], [35, p. 272] [also known as -ary amplitude shift keying ( -ASK)]2 on the channel for data. The proposed scheme suffers a spectral -QAM constellations [5], efciency penalty compared to [8], [10]. However the scheme has the advantage of providing a solution which lends itself to simplicity of design and performance evaluation, as we will see next. In this section we rst introduce the hybrid BPSK/ -AM modulation scheme. We then present the details of the proposed adaptive scheme. A. Hybrid BPSK/ -AM Modulation Scheme

A block diagram of the proposed hybrid BPSK/ -AM modulation scheme is shown in Fig. 2. Following the form of the UQPSK modulation [34, p. 622], the hybrid BPSK/ -AM transmitted signal can be written as (4) is the radian carrier frequency, and and are where , the powers of the (data) and (voice) components of and correspond to the data and respectively. In (4), voice symbol streams, respectively; that is

(5) is a unit power shaping pulse of duration (the where ). In (5) ( , signal bandwidth is hence , , ) are the Gray-mapped data symbols of the (as depicted in Fig. 3) and are the voice data bits bits.
2 We use here a symmetric -AM constellation in which the signal points are symmetrically located about the origin as shown in Fig. 1.

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

ALOUINI et al.: ADAPTIVE MODULATION SCHEME

839

Fig. 1. Gray mapping for the

M -AM constellations.

Fig. 2. Block diagram of the proposed adaptive system.

The channel introduces a multiplicative fading gain , a phase shift , and additive white Gaussian noise (AWGN) with power spectral density [W/Hz]. Hence term the received signal can be written as

where and are independent zero-mean Gaussian noise . For uncoded data samples with the same variance and voice streams and independent hard decisions on the and channels (see Fig. 2), the conditional (conditioned on ) symbol error rate (SER), SER , for data and BER, , for voice are given by [34, p. 631] and BER

(6) SER and ), Assuming perfect channel estimation ( the received signal is rst coherently demodulated, then the (data) signal is passed through an adaptive gain controller signals are passed through matched (AGC). Both and ) to form the lters, then sampled (at times and given by decision variables BER erfc erfc (8) (9)

, and are the data where and voice instantaneous CNR, respectively, and erfc is the complementary error function dened by erfc (10)

(7)

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

840

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 5, MAY 1999

Fig. 3. Bit error rate versus received CNR for

M -AM.
For comparison, the BER upper-bound (12) is plotted in Fig. 3 by star/solid lines for different values of . Note that (12) tightly upper bounds the exact BER expressions (as given in and for BER , the Appendix) for all values of which is the BER range of interest for data transmission. Hence we will use this upper-bound (12) to derive closedform expressions which upper-bound the average data BER. In addition, (12) has the advantage of being invertible in the sense that it provides simple expressions for the data switching thresholds, as shown in Section III-B. B. Proposed Adaptive Scheme We now describe the details of our proposed system shown in Fig. 2. Assuming a perfect channel fading amplitude esti3 (equivalently, a perfect channel CNR estimation mate ) and a peak power constraint of [W], variable[W] is used on the BPSK of the power channel to ensure continuous xed-rate voice transmission at the target voice BER BER i.e., the power allocated to is set to just meet the voice BER requirement voice BER . The remaining available power [W] is dynamically assigned on the channel to support -AM below the target data BER BER . the (adaptive) Specically, based on the channel CNR estimate and on the
3 Accurate channel fading estimation can be obtained via two techniques: transparent tone in band (TTIB) or pilot symbol assisted modulation (PSAM). The usage of these two techniques over fading channels is described in details in [37, Sect. 10.3].

The data symbol estimates are then passed through an ary Gray demapper to obtain an estimate of the source data bits . Using the same procedure described in [36], [37, Ch. 5] to obtain the exact BER of square -ary quadrature amplitude modulation ( -QAM) with two-dimensional Gray coding, we -AM modulation derived exact BER expressions for the with Gray coding as shown in Fig. 1. The procedure as well as the exact BER expressions are given in the Appendix. These exact BER expressions are plotted by the solid lines in Fig. 3 and are in excellent agreement with Monte Carlo simulated BER values which are plotted by o on the same Fig. 3. For large CNR, where the probability of symbol error is dominated by the probability of adjacent symbol error, the BER with Gray encoding can be approximated by [34, p. 210], [35, p. 265] BER SER (11)

For comparison, the dash lines in Fig. 3 show the BER approximation (11) for different values of . Note that (11) lower bounds the exact BER expressions (as given in the and the bound is tighter for Appendix) for all values of and high CNR. Using the Chernoff-bound on the erfc( ) low function in (11), it can be shown that (11) is upper-bounded for large CNR by BER SER (12)

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

ALOUINI et al.: ADAPTIVE MODULATION SCHEME

841

v d Fig. 4. Outage probability for voice Pout and data Pout versus the average CNR .

available power , the decision device at the receiver to be transmitted on selects the signal constellation size the channel, congures the demodulator accordingly, and informs the transmitter about that decision via the feedback path. We now describe the power allocation for voice and data as well as the constellation size assignment for data transmission in more detail. Our proposed modulation scheme uses the channel state information at the transmitter to minimize its average power consumption subject to the peak power constraint. Specically, voice transmission is not attempted when the power required to meet the target voice BER exceeds the peak power constraint , and in this case a voice outage is declared. Furthermore, since the voice has to operate at the target BER , we see from (9) that the power allocated to voice transmission must be set to (or equivalently otherwise (13) erfc BER and erfc denotes the where inverse complementary error function. For data the scheme responds to the instantaneous channel CNR uctuation by varying its constellation size as follows. The data CNR range fading regions, and the constellation is divided into (where is the number of bits per -AM size ). symbol) is assigned to the th region ( When the received data CNR is estimated to be in the th is transmitted. The region region, the constellation size are set to the boundaries (or switching thresholds) )

CNR required to achieve the target BER using M -AM over an AWGN channel. Specically from (12) we have BER (14) If during voice transmission the remaining available power is not able to support BPSK on the channel, then no data is transmitted and a data outage is declared. Hence the power allocated to data transmission can be written as

equivalently otherwise. (15) IV. PERFORMANCE ANALYSIS In this section we analyze the performance of the proposed scheme and we present both numerical and simulation results which are in perfect agreement, as can be seen in Figs. 411. All our numerical and simulation results are plotted as a for different values of the function of the average CNR and for different maximum Nakagami fading parameter constellation sizes (levels). Note that all these numerical and simulation results assume a target uncoded voice BER, BER , of 10 , and a target uncoded data BER, BER of 10 . We used these values to speed up our simulations, however our analytic derivations apply to any set of BER requirements.

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

842

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 5, MAY 1999

Fig. 5. Average power allocation for voice hPv i=P and data hPd i=P versus the average CNR .

We use the MATLAB Communication Toolbox for our computer simulations. The powers allocated for voice and data as well as the constellation size for data transmission are determined at each symbol time according to the fading level, as explained in Section III. We assume perfect channel estimation,4 coherent phase detection at the receiver, and Gray coding for bit mapping on the -AM constellations, as shown in Fig. 1. All our simulations use a 4 level modem which is able to support up to 16-AM modulation for data transmission. A. Outage Probability Since no voice is sent when the required power exceeds , the voice transmission suffers an outage probability of (16) Substituting (3) in (16), then using [23, p. 364, Eq. (3.381.3)], as we can express

Eq. (8.350.2)] (18) For positive integers [23, p. 949, Eq. (8.352.2)], (19) where denotes the degree polynomial dened by (20) Thus if we restrict ourselves to integer values of be expressed as , (17) can

(21) For the special case of the Rayleigh fading channel ( (21) reduces to ), (22)

(17) is the complementary incomplete where gamma function (or Pryms function) dened by [23, p. 949,
4 We do not address in this paper the effect of channel estimation errors. However, the analytical tools used in [8], [38], and [39] to characterize the effect of channel estimation errors and feedback delay on adaptive M -QAM modulations can be used to study the performance of our proposed hybrid scheme under imperfect channel estimate conditions.

is insufSince no data is sent when the available power cient to support BPSK on the channel, data transmission of suffers an outage probability

(23)

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

ALOUINI et al.: ADAPTIVE MODULATION SCHEME

843

Fig. 6. Overall normalized average power hP i=P allocated to both voice and data versus the average CNR .

where is the rst data switching threshold. If we restrict ourselves to integer values of , (23) can be expressed as (24) Hence for the special case of the Rayleigh fading channel ), (24) reduces to ( (25) and for voice Fig. 4 shows the outage probability and data transmission, respectively. In the high average CNR region (i.e., higher than 4 dB for voice and higher than 9 dB for data), the higher the average CNR, the lower the outage probability, as expected. In addition, the scheme meets the more stringent delay requirements of voice since for a xed data suffers a higher outage probability than voice at all average CNRs. Although these outage curves appear simple and intuitive, they will in fact be crucial to explain many of our subsequent performance results. B. Average Power Allocation The normalized average power allocated for voice transis given by mission

If we restrict ourselves to integer values of be expressed as

, (26) can

(27) For we have [23, p. 951, Eq. (8.359.1)] (28) where dened as is the exponential-integral of rst order function

(29) Thus for the special case of the Rayleigh fading channel ), using (28) in (26) we obtain ( (30) The normalized average power allocated for data transmission is given by

(26)

(31)

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

844

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 5, MAY 1999

Fig. 7. Achievable spectral efciency for voice hRv i=W and data hRd i=W versus the average CNR : (a) m = 1, (b) m = 2, and (c) m = 4.

If we restrict ourselves to integer values of be expressed as

, (31) can

(32) For the particular case of the Rayleigh fading channel ( using (28) in (31) we get ),

(33) Fig. 5 shows in dash lines the normalized average power . This gure also allocated for voice transmission displays in solid lines the normalized average power allocated . The overall normalized average for data transmission, is shown in Fig. 6. The power behavior of the curves in Fig. V, which varies in the different regions of average CNR, can be explained by the outage curves in Fig. 4. In particular, we see in Fig. 4 that at low s both voice and data suffer a large outage probability. Hence, since there is no transmission during outage, the corresponding power consumptions in Figs. 5 and 6 are low. Consider now dB). the region of extremely low average CNR (i.e., in this region, as increases Observe that for a xed (i.e., the amount of fading decreases) the power consumption for voice decreases. This can be explained by the following argument. At these extremely low values of note from Fig. 4 that the outage probability for voice is essentially the same for values. However, when voice transmission is possible all

channels with a higher amount of fading will require more power to maintain a constant voice CNR . Thus, in this region power consumption for voice increases relative to the amount of fading. In the medium CNR region (i.e., 2.5 dB dB), we see that a larger value of corresponds to a larger power consumption for voice and a smaller one for data. This can be explained by observing that in Fig. 4 the data outage probability in this region is essentially independent of but the voice outage probability decreases as increases. Thus, as increases, we are transmitting voice more often and therefore we must allocate a larger percentage of our power to voice. In the region of high average CNR (i.e., 12.5 dB ), voice outage probability is small, and since the channel is quite good, a small fraction of the total power is needed for the voice transmission. Thus most of the power is allocated to (i.e., a data transmission. In this favorable region a large small amount of fading) implies that less power is needed for voice transmission and therefore more power can be allocated to high rate -ary data transmission. C. Achievable Spectral Efciency The average link spectral efciency for voice transmission is given by

(34) When is restricted to integer values (34) may be written as (35)

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

ALOUINI et al.: ADAPTIVE MODULATION SCHEME

845

which reduces to (36) ). The average for the special Rayleigh fading case ( is just link spectral efciency for data transmission ) associated with the the sum of the data rates ( regions, weighted by the probability individual

explained by the following argument. Channels with a small exhibit signicant fading and a corresponding wide range will have most of of CNR values. Channels with a large their CNR concentrated around the average CNR which is small in the low average CNR region. Hence channels with a smaller fading parameter will have a slightly higher spectral efciency since the larger CNR uctuation results in a lower probability of outage in this low average CNR region (as can be seen in Fig. 4). D. Average Bit Error Rate

that the data CNR

falls in the th fading region (37)

s can be expressed using [23, p. 364, Eq. where the (3.381.3)] as

Voice transmission is always operating at the target BER, BER . On the other hand, since the choice of the constellation size for data transmission is done in a conservative fashion, data is transmitted at an average BER, BER smaller than BER . This average BER can be computed exactly as the ratio of the average number of bits in error over the total average number of transmitted bits BER BER (41)

(38) in the most general case and may be written as where (39) when is restricted to integer values. For the Rayleigh fading case (39) reduces to (40) The dashed lines in Fig. 7 show the average spectral ef. This gure also ciency for voice transmission shows the average spectral efciency for data transmission for different maximum constellation sizes. For high average CNRs (above 15 dB) the scheme provides a higher spectral efciency for data then for voice and can therefore meet the higher data rate requirements for data transmission. dened as The overall average spectral efciency the sum of the voice and data average spectral efciencies ), is shown in Fig. 8. At (i.e., corresponds to a large overall high average CNR a large average spectral efciency for voice and data. However, at low average CNR (i.e., less than 4 dB for voice and less than 10 corresponds to a low overall average dB for data) a large spectral efciency. This may seem surprising at rst but can be BER BER (42)

It can be shown using (3) and (12) in (42) that BER is upper-bounded as shown in (43) at the bottom of the page where

When BER

is restricted to integer values these bounds become

BER (44)

BER

BER

(43)

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

846

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 5, MAY 1999

Fig. 8. Overall spectral efciency hRi=W versus the average CNR : (a) m = 1, (b) m = 2, and (c) m = 4.

which reduces in the Rayleigh case ( BER

) to

BER

(45)

Fig. 9 shows the average BER for both voice and data, for different maximum constellation sizes or levels. Note that voice transmission is always operating at the target BER, BER . On the other hand, data is transmitted at an average BER BER smaller than the target BER , as expected from our conservative choice of constellation size. Since data transmission uses the largest constellation often when the average CNR is high, the average BER prediction as increases becomes dominated by the BER performance of that constellation. In addition, at high average CNR as increases the average BER for data decreases, as one might expect. dB) the average However, at low average CNR (i.e., BER for data actually increases as increases. This behavior may seem surprising at rst, but can be explained by the fact dB a large implies that only a small amount that for of power is allocated to data transmission, as can be seen in Fig. 5. Hence since data can only use a small fraction of the power, its BER increases. ) We show the simulated BER for Rayleigh fading ( in Figs. 10 and 11, and for Nakagami fading with respectively. The BER simulation results for voice transmission in these gures are in perfect agreement with the analytical calculations. However, the simulated BERs for data

transmission are slightly lower than the analytical calculations since the latter are based on the upper-bound (12) of the BER performance of -AM with Gray coding. The fact that (see Fig. 3) combined this bound is tighter (12) for lower with the fact that the scheme often uses the smallest available constellation at low average CNRs explains why the overall average BER upper-bound for data transmission is tighter at low average CNRs. V. CONCLUSION We have proposed an adaptive modulation scheme which offers a simple and energy-efcient solution to voice and data integration over fading channels. The proposed design is intended to provide the user with a high average spectral efciency for data communications while meeting the stringent delay requirements imposed by voice. For favorable channel conditions, most of the power is allocated to high rate data transmission by using -AM with a large constellation size. As the channel degrades, the modem reduces its data throughput and reallocates most of its available power to ensure a continuous and satisfactory voice transmission. We evaluated the performance of our proposed scheme in terms of outage probability, average allocated power, achievable spectral efciency, and average BER for both voice and data transmission. Although the design and analyses for our proposed scheme is quite simple, this simplicity comes at the expense of a spectral efciency penalty compared to -QAM constellations [5], [8], [10]. We are currently looking at other possibilities of improving the spectral efciency of the proposed scheme. One

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

ALOUINI et al.: ADAPTIVE MODULATION SCHEME

847

Fig. 9. Average BER for voice hBERv i and data hBERd i versus the average CNR : (a) m = 1, (b) m = 2, and (c) m = 4.

Fig. 10.

Average BER for voice hBERv i and data hBERd i versus the average CNR for Rayleigh fading (m = 1).

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

848

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 5, MAY 1999

Fig. 11.

Average BER for voice hBERv i and data hBERd i versus the average CNR for Nakagami fading (m = 2).

possibility is to multiplex the voice and data bit streams and -QAM constellations. Besides to use adaptive symmetric having several parameter choices to optimize, this multiplexing scheme will exhibit some performance and complexity tradeoffs relative to our proposed technique. Another area of further study is unequal error protection codes which can be used in conjunction with adaptive modulation to achieve different levels of error protection while improving the throughput of data communication and further reducing the outage probability of voice. Design and performance evaluation of multiresolution adaptive modulations where the constellation of voice and data are superimposed on the top of one another would also be another interesting future research direction. Finally, the use of a speech activity detector (SAD), which segments a conversation speech into talkspurts and silences [40], can also improve the overall spectral efciency. A SAD can be used in conjunction with an adaptive multimode modem to send adaptive -QAM [8], [10] for data transmission during the channel, silences when voice is not transmitted on the and our proposed scheme for simultaneous voice and data transmission can be used during the talkspurts. APPENDIX EXACT BER EXPRESSIONS FOR -AM OVER AN AWGN CHANNEL In this Appendix we derive the exact BER expression for 4-AM with Gray coding over an AWGN channel, and give the exact BER expressions for 8-AM and 16-AM. For 4-AM the four symbols are symmetrically distributed about zero with equal distance between two adjacent symbols

as shown in Fig. 1. In Fig. 1, is the amplitude level, is is the distance between two the symbol duration, adjacent symbols, and the dashed vertical lines represent the decision boundaries. Since we consider an AWGN channel , the noise is with a noise power spectral density of normally distributed with zero mean and variance . Consider rst the left bit of each 4-AM symbol, as shown in Fig. 1. A bit error occurs when the bit 1, corrupted by noise, falls into the boundaries of bit 0 or vice versa. For example, the left bit of the symbol 10, i.e., 1, will be interpreted 0 when . Hence its probability of error the noise is larger than is given by (46) is the Gaussian -function which is related to the where error complementary function as dened in (10) by erfc (47)

, and . Similarly, Assuming each of the four symbols has equal probability, the is error probability of the left bit

(48) Consider now the right bit of each 4-AM symbol as shown in Fig. 1. Following the same procedure it can be shown that its

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

ALOUINI et al.: ADAPTIVE MODULATION SCHEME

849

probability of error

is given by (49)

ACKNOWLEDGMENT The authors would like to thank Dr. M. K. Simon of the NASA Jet Propulsion Laboratory (JPL), Pasadena, CA, for early discussions regarding unbalanced QPSK and its applications. They would also like to thank the anonymous reviewers for their valuable comments and for the suggested alternative method of multiplexing voice and data bits. REFERENCES (51)
[1] J. F. Hayes, Adaptive feedback communications, IEEE Trans. Commun. Technol., vol. COM-16, pp. 2934, Feb. 1968. [2] J. K. Cavers, Variable-rate transmission for Rayleigh fading channels, IEEE Trans. Commun., vol. COM-20, pp. 1522, Feb. 1972. [3] B. Vucetic, An adaptive coding scheme for time-varying channels, IEEE Trans. Commun., vol. 39, pp. 653663, May 1991. [4] S. Otsuki, S. Sampei, and N. Morinaga, Square-QAM adaptive modulation/TDMA/TDD systems using modulation level estimation with Walsh function, Electron. Lett., vol. 31, pp. 169171, Feb. 1995. [5] W. T. Webb and R. Steele, Variable rate QAM for mobile radio, IEEE Trans. Commun., vol. 43, pp. 22232230, July 1995. [6] Y. Kamio, S. Sampei, H. Sasaoka, and N. Morinaga, Performance of modulation-level-controlled adaptive-modulation under limited transmission delay time for land mobile communications, in Proc. IEEE Veh. Technol. Conf. VTC95, Chicago, IL, July 1995, pp. 221225. [7] J. M. Torrance and L. Hanzo, Upper bound performance of adaptive modulation in a slow Rayleigh fading channel, Electron. Lett., vol. 32, pp. 718719, Apr. 1996. [8] M.-S. Alouini and A. Goldsmith, Adaptive M-QAM modulation over Nakagami fading channels, in Proc. Communication Theory MiniConf. (CTMC-VI) in conjunction with IEEE Global Commuication Conf. GLOBECOM97, Phoenix, AZ, Nov. 1997, pp. 218223. [9] V. O. Hentinen, Error performance for adaptive transmission on fading channels, IEEE Trans. Commun., vol. COM-22, pp. 13311337, Sept. 1974. [10] A. Goldsmith and P. Varaiya, Increasing spectral efciency through power control, in Proc. IEEE Int. Conf. on Commun. ICC93, Geneva, Switzerland, June 1993, pp. 600604. [11] S. M. Alamouti and S. Kallel, Adaptive trellis-coded multiple-phasedshift keying for Rayleigh fading channels, IEEE Trans. Commun., vol. 42, pp. 23052314, June 1994. [12] A. Goldsmith and S. G. Chua, Adaptive coded modulation for fading channels, IEEE Trans. Commun., vol. 46, pp. 595602, May 1998. [13] T. Ue, S. Sampei, and N. Morinaga, Symbol rate and modulation level controlled adaptive modulation/TDMA/TDD for personal communication systems, in IEICE Trans. Commun., vol. E78-B, pp. 11171124, Aug. 1995. [14] H. Matsuoka, S. Sampei, N. Morinaga, and Y. Kamio, Adaptive modulation system with variable coding rate concatenated code for high quality multi-media communication systems,IEICE Trans. Commun., vol. E79-B, pp. 328334, Mar. 1996. [15] G. Bremer and K. D. Ko, Simultaneous voice and data on the general switched telephone network using framed QADM, IEEE Commun. Mag., vol. 34, pp. 5863, Dec. 1996. [16] D. J. Goodman, R. A. Valenzuela, K. T. Gayliard, and B. Ramamurthi, Packet reservation multiple access for local wireless communications, IEEE Trans. Commun., vol. 37, pp. 885890, Aug. 1989. [17] K. Zhang and K. Pahlavan, An integrated voice/data system for mobile indoor radio networks, IEEE Trans. Veh. Technol., vol. 39, pp. 7582, Feb. 1990. [18] S. Nanda, Analysis of packet reservation multiple access: Voice data integration for wireless networks, in Proc. IEEE Global Commun. Conf. GLOBECOM90, San Diego, CA, pp. 19841988. [19] N. Wilson, R. Ganesh, K. Joseph, and D. Raychaudhuri, Packet CDMA versus dynamic TDMA for multiple access in an integrated voice/data PCN, IEEE J. Select. Areas Commun., vol. 11, pp. 870884, Aug. 1993. [20] G. Wu, K. Mukumoto, and A. Fukud, Analysis of an integrated voice and data transmission system using packet reservation multiple access, IEEE Trans. Veh. Technol., vol. 43, pp. 289297, May 1994. [21] B. C. Kim and C. K. Un, An efcient wireless voice/data integrated access algorithm in noisy channel environments, IEICE Trans. Commun., vol. E79-B, pp. 13941403, Sept. 1996. [22] M. Nakagami, The -distributionA general formula of intensity distribution of rapid fading, in Statistical Methods in Radio Wave Propagation. Oxford, UK: Pergamon, 1960, pp. 336.

Hence the average BER for 4-AM is given by BER (50) On the other hand the average power per symbol is

Thus

where is the signal bandwidth. The exact BER of 4-AM can hence be rewritten in terms of average CNR, , as BER (52) The exact BER expressions for 8-AM and 16-AM can be calculated in a similar way and are given by BER -

(53) BER -

(54)

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.

850

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 5, MAY 1999

[23] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products, 5th ed. San Diego, CA: Academic, 1994. [24] H. Suzuki, A statistical model for urban multipath propagation, IEEE Trans. Commun., vol. COM-25, pp. 673680, July 1977. [25] T. Aulin, Characteristics of a digital mobile radio channel, IEEE Trans. Veh. Technol., vol. VT-30, pp. 4553, May 1981. [26] W. R. Braun and U. Dersch, A physical mobile radio channel model, IEEE Trans. Veh. Technol., vol. 40, pp. 472482, May 1991. [27] A. U. Sheikh, M. Handforth, and M. Abdi, Indoor mobile radio channel at 946 MHz: Measurements and modeling, in Proc. IEEE Veh. Technol. Conf. VTC93, Secaucus, NJ, May 1993, pp. 7376. [28] S. Basu, E. M. MacKenzie, S. Basu, E. Costa, P. F. Fougere, H. C. Carlson, and H. E. Whitney, 250 MHz/GHz scintillation parameters in the equatorial, polar, and aural environments, IEEE J. Select. Areas Commun., vol. SAC-5, pp. 102115, Feb. 1987. [29] E. J. Fremouw and H. F. Bates, Worldwide behavior of average VHFUHF scintillation, Radio Sci., vol. 6, pp. 863869, Oct. 1971. [30] H. E. Whitney, J. Aarons, R. S. Allen, and D. R. Seeman, Estimation of the cumulative probability distribution function of ionospheric scintillations, Radio Sci., vol. 7, pp. 10951104, Dec. 1972. [31] E. J. Fremouw, R. C. Livingston, and D. A. Miller, On the statistics of scintillating signals, J. Atmos. Terrest. Phys., vol. 42, pp. 717731, Aug. 1980. [32] P. K. Banerjee, R. S. Dabas, and B. M. Reddy, C-band and L-band transionospheric scintillation experimentSome results for applications to satellite radio systems, Radio Sci., vol. 27, pp. 955969, June 1992. [33] M. K. Simon, Error probability performance of unbalanced QPSK receivers, IEEE Trans. Commun., vol. COM-26, pp. 13901397, Sept. 1978. [34] M. K. Simon, S. M. Hinedi, and W. C. Lindsey, Digital Communication TechniquesSignal Design and Detection. Englewood Cliffs, NJ: Prentice Hall, 1995. [35] J. G. Proakis, Digital Communications, 2nd ed. New York, NY: McGraw Hill, 1989. [36] P. M. Fortune, L. Hanzo, and R. Steele, On the computation of 16-QAM performance in Rayleigh-fading channels, IEICE Trans. Commun., vol. E75-B, pp. 466475, June 1992. [37] W. T. Webb and L. Hanzo, Modern Quadrature Amplitude Modulation. New York: IEEE Press, 1994. [38] A. J. Goldsmith and S. G. Chua, Variable-rate variable-power M-QAM for fading channels, IEEE Trans. Commun., vol. 45, pp. 12181230, Oct. 1997. [39] D. L. Goeckel, Adaptive coding for fading channels using outdated channel estimates, in Proc. IEEE Veh. Technol. Conf., VTC98, Ottawa, Ont. Canada, May 1998, pp. 19251929. [40] P. T. Brady, A model for generating on-off speech patterns in two-way conversation, Bell Syst. Tech. J., vol. 48, pp. 24452472, Sept. 1969.

Xiaoyi Tang will receive the B.S. degree in electrical engineering from the California Institute of Technology (Caltech), Pasadena, in June 1999. Currently, he is an Undergraduate Research Assistant with the Communications group at Caltech.

Andrea J. Goldsmith (S94M95) received the B.S., M.S., and Ph.D. degrees in electrical engineering from the University of California, Berkeley in 1986, 1991, and 1994, respectively. From 1986 to 1990, she was with Maxim Technologies, where she worked on packet radio and satellite communication systems, and from 1991 to 1992, she was with AT&T Bell Laboratories, where she worked on microcell modeling and channel estimation. She was an Assistant Professor of electrical engineering at the California Institute of Technology, Pasadena, from 19941998, and is currently an Assistant Professor of electrical engineering at Stanford University, Stanford, CA. Her research includes work in capacity of wireless channels, wireless communication theory, adaptive modulation and coding, joint source and channel coding, and resource allocation in cellular systems. Dr. Goldsmith is a recipient of the National Science Foundation CAREER Development Award, the ONR Young Investigator Award, two National Semiconductor Faculty Development Awards, an IBM Graduate Fellowship, and the David Griep Memorial Prize from the University of California, Berkeley. She is an Editor for the IEEE TRANSACTIONS ON COMMUNICATIONS and the IEEE PERSONAL COMMUNICATIONS MAGAZINE.

Mohamed-Slim Alouini (S94M99) was born in Tunis, Tunisia. He received the Dipl. Ing. degree from the Ecole Nationale Sup rieure des e T l communications (TELECOM), Paris, France, ee and the Dipl me dEtudes Approfondies (DEA) o degree in electronics from the University of Pierre & Marie Curie (Paris VI), Paris, France, both in 1993. He received the M.S.E.E. degree from the Georgia Institute of Technology (Georgia Tech), Atlanta, in 1995, and the Ph.D. degree in electrical engineering from the California Institute of Technology (Caltech), Pasadena, in 1998. While completing the DEA thesis, he worked with the optical submarine systems research group of the French National Center of Telecommunications (CNET-Paris B) on the development of future transatlantic optical links. While at Georgia Tech, he conducted research in the area of Ka -band satellite channel characterization and modeling. From June to August 1998, he was a postdoctoral fellow with the Communications Group at Caltech, carrying out research on adaptive modulation techniques and on CDMA communications. He joined the Department of Electrical and Computer Engineering, University of Minnesota, Minneapolis, in September 1998, where his current research interests include statistical modeling of multipath fading channels, adaptive modulation techniques, diversity systems, and digital communication over fading channels. Dr. Alouini is the recipient of a National Semiconductor Graduate Fellowship Award.

Authorized licensed use limited to: Stanford University. Downloaded on June 30,2010 at 01:09:03 UTC from IEEE Xplore. Restrictions apply.