A Technique for Orthogonal Frequency Division Multiplexing Frequency Offset Correction
Pad H.M m e , Member, IEEE
is beiig suggested as an efficient modulation for applications ranging from modems [4], to digital audio broadcast [5]. Oneof theprincipaladvantages of OFDM isitsutility for transmission at very nearly optimum performance in unequalized channels and in multipath channels. As described in [3]+5], htersymbol inkxfereace,(ISI)and intercarrig interference (KT) c t be entitely eliminated by the s xn of inserting between symbok a d l time i * n km@a as aguardinterval. The lengthof the guard inbw+,.i$~&de equal to or greater than the time spread of the c m. &+ E into , this guard interval, the symbol signal wavefom, >'@ mtmled periodically, orthogonality o the carriers is m$nfained;qr f thvs sBcGbg9tye the symbol period, eliminatingICI. ~lscx~.&ke symbols do not overlap because of the guard interval, IS1 is e l i i a t e d , too. One of the principal disadvantages of OmlM is sensitivity to frequency offsetin the channel. example, the coded For OFDM system developed by CCElT (Cenhe.Commun. d'Etudes de I. INTRODUCTION Telediffusion et Telecommuications) for digital sound broadcasting to mobile receivers incorporates an AFC (automatic HE TECHNIQUE demibed in this paper has been defrequency control) loop in the receiver to reduce frequency veloped to correct frequency offset errors in digital comoffsetcausedbytuning oscillatorinaccuraciesanddoppler munications systems employing orthogonal frequency division multiplexing (OFDM) the method of modulation. The aim shift [6]. as of the paper is two€ol& show theeffect offset errors have on There are two deleterious effects caused by frequency offset; to in theoutput of the signal-to-noise ratio of the OFDM carriers and to present one is thereductionofsignalamplitude the filters matched to each of the carriers and the second is analgorithmto eswate offset so that it mayberemoved introduction of IC1 from the other carriers which are now no prior to demodulitbn. OFDM, the carriers OFDM is a bandwidth efficient signalling scheme for dig- longer orthogonal to the filter. Because, in ital communicationsthatwasfirstproposedbyChang [l]. are inherently closely spaced in fkquency compared to the offset becomes The main difference between frequency division multiplexing channelbandwidth,thetolerablefrequency (FDM)andOFDM, is that in OFDM thespectrum of the a very small fraction of the chaonel bandwidth. Maintaining individual carriers mutually overlap, giving thereforean opti- sufficient open loop frequency accuracy can become difficult satellite links with multiple frequency translamum spectrum efficiency (asymptotically Q b/Hz for 2Q-ary in links, such as modulation of each carrier). Nevertheless, theOFDM carriers tions or,as mentioned previously, in mobile digitalradio links exhibit orthogonality on asymbol interval if synthesized such that may also introduce significant Doppler shift. The effects that they are spaced infrequencyexactlyatthereciprocal of frequency offset are presented in Section II. of thesymbol interval. Fortunately, this synthesiscan be In Section 1 1 we present an algorithm estimate frequency 1, to accomplished perfectly, in principle, utilizing discrete the offset from the demodulated data signals in the receiver. The d tas Fourier transform ( f ) first described by Darlington [2] and algorithmextendsto OFDM, withimportantdifferences,a later, for data modems, by Weinstein and Ebert [3]. With the method described by Simon and Divsalar for single carrier [7] f recent evolution o integrated circuit digital signal processing MPSK.Thetechniqueinvolvesrepetitionof a data symbol (dsp) chips, OFDM has become practical to implement and and comparison of the phases of each of the carriers between are P p r approved by M. J. Joindot, the Editor for Radio Communications of the successive symbols. Since the modulation phase values ae the IEEECommunications Society. Maauscriptreceived November 27, 1991; not changed, the phase shift of each of the carriers between revised June 29,1992. the frequency offset. successive repeated symbols is due to The. author is with the Electrical and Computer Engineering Department, is The frequency offset estimated using a maximum likelihood Naval Postgmduate School, M m t m y , CA 93943, USA. EEE Log Number 9401947. estimate (MLE) algorithm. Performance of the algorithm as


0090-6778/94$04.00Q 1994 IEEE

a function &/No (individual OFDM carrier energy to onesided spectral density of additive white Gaussian noise) and in freq&cy offset is i u c l u ~ Section III. In the eventthatthefrequencyoffsetexceeds f l / 2 the iis intercarrier spacing, the maximum lmt of the algorithm, a strategy is required for initial acquisition. One such strategy is described in Section N.

n. omhi

synchronization at thisstage of thereceiver)leading to a receivedsequence as givenby (4). It is assumedthatthe the channeldoesnotchange(much) impulseresponseof duringthesymbolplus gnard interval (this correspondsto "slow-fading" in a radio frequency channel). The insertion of guard intervals renders the received carriers orthogonalonthe N pointsymbolinterval.However,the demodulation process, which is implemented with a dft (the dft is equivalent to m t h d filter reception in the absence of ace frequency offset) is affected by frequency offset. That is,
N-1 nO =

An OFDM transmission symbol is givenby the N point complex modulation sequence

xn.= (1/N)

= 0, 1, 2 , . ...N - 1;

N 3 2K

+ 1.


the kth element of the dft sequence, consists of three components; Y = (x&k){(sinm)/Nsin ( m / N ) } k .&re(N-l)/N

It consists of 2K 1 complex sinusoids which have been modulated with 2K + 1 complex modulation values {xk}. we that the h i i u l sinusoids are orthogonal on the dvd a symbol interval, that is


+ + wk. (7)

The first component is the modulation value x k modified by the channel transfer function. This component experiences N-1 an amplitude reduction and phase shift due to the frequency c % k & = (1/N)Ixk12skl (2) offset. Since Nis always much greater than N sin (?rc/N) XE, n=O may be replaced by r e . The second t r is the IC1 caused by the frequency offset em where Xnk = (l/N)XkeZrink/N. and is given by We also note that the N point discrete Fourier transform (dft). of (1) is ,the N point sequence K 4= (XI&){(sin?re)/(Nsin(x(l- k E)/N)))
, a

& t e


1s-K l#k

. ej7re(N-l)/Ne-jr(l-k)/N .


= {XO, XI,.-.XK, , . 0, 0


0,x - K . ..x-2,X

I }



is theinversediscrete

Fourier transform (IDFl') of (3) and defines a practical modulation-carrier synthesis technique for generating OFDM wt perfect o~ogonality. ih
After passingthrough a bandpasschannel,thecomplex edp of the & d sequence can be expressed as noe v e
y,, = <1/N) XJ$ke2xjn(k+c)lN

In order to evaluate the statistical properties of the ICI, some further assumptions m necessary. Specifically, it will be a S S U l l l e d that E[Xk] = 0 and E[XkX;] = IXl26lk, that is, the modulation values have, zero mean and are uncorrelated. W h this provision E[&]= 0, and i t


I=:, 1


/ {Nsin(?r(l- L

-I- ) / N ) } z . (9) ~

n=o, 1 , 2,..., N - 1


The average channel gain, E{IHi12} = lHI2, is constant so it can be separated from the sum and (9) becomes

where & is the transfer function of the channel the at E[1&12] IX121H12(sin.lrc)2 = fresueacy of the kth m e r , t is the relative frequency offset K-k of the channel (h ratio of the actual frequency offset to te . l / { N s i n ( r ( p + ~ ) / ~ ) } ~(10) . the i&xmnier spin&, and w,, is the complex envelope of p=-K-k ##O additive white Gaussian noise (AWGN). Let the actual symbol transmitted be the N + Ns point sequence The sum in (10) can be bounded for E = 0. It consists of 2K positive terms. The intmval of the sum is contained within the { x N - N ~ , . . .,XN--2, XN-I, ~ ~ 5 1 . -,, ~ N - I } . (5) longer interval -2K 5 p <_ 2K,ita loaation dependent on k. with Ng greater than orequal to thetimespread of the Recall that2K 5 N - 1. Also note-thefollowing; the argument channel. The N, point precursor signal allows the received of the sum is periodic with period N , it is an ev& function of symbol sequence to. reach steady state by n = 0 (we assume p, and it even about p = N / 2 . Thus the 2K terms of the sum is




m. WWCY E ~ T I M A ~ O N OFFSET If an OFDM transmission symbol is repeated, one receives, ih the absence of noise, the 2N point sequence

= 0, 1,.. . , 2 N - 1. (16)
The kth element of the N point dft of the first N points of (16) is


I a



01 .5


n z







Fig. 1. S N R versus relative frequency offset.

N terms in the intervals -N/2 5 p 5 -1 and 1 5 p 5 N/2 for every k. Consequently,
are a subset of the

andthekthelement sequence is
2N-I n=N N -1

of the dft of thesecond half of the


l/(Nsin?rp/N)' < 2 E l / ( N s i n ( ~ p / N ) ) ' . (11)

= C T ~ + N ~ - ' ~ ' ~k ~ /0,~ 1,. = ;

.. ,N - 1.(18)

Observe that (sin~pPjN)" ( ~ P / N for Ipl 5 N/2. There2 )~ fore,

N f2

< 2c1/(2p)2 < - c l / p 2


= 7r2/12 = 0.882


upper boundsthesum

for small


0.5947 for

d k mn d that the sum in (10) isboundedby er ie


0.5 so that

5 0 . 5 9 4 7 1 X 1 2 1 H 1 2 ( ~ i n ~ ~ ) 2161 5 0.5 ;


upper bounds the variance of the IC1 for values of frequency offset up to plus or minus one half the carrier spacing. i= (1/27r) tan-1 Equation (13) may be used to give a lower bound for the S N R at the outputof the dft for the OFDM carriers in a channel with AWGN and frequency offset. Thus, This is an intuitively satisfying result since,in the absence of SNR 2 l X l " l H 1 2 { S i n ? r E / ( ~ E ) ) 2 / E noise, the angle of Y 2 k Y i k is ~ A for each k . Fig. 2 shows {0.5947)X121H12'(Sin?r~)2 l w k 1 2 ] } . (14) +E[ simulation results for the estimate of E obtained using (21) It is easilyestablishedthat ~ X ~ 2 ~ H ~ 2 / E= /EJN,,~ 2 versus E for values of EJN, corresponding to 17 and 5 dB. [ W~ ] where E, is the averaged received energy ofthe individual f carriers and N0/2 is the power spectral density of the AWGN A. Statistical Properties o the Estimate in the bandpass transmission channel. Therefore, (14) may be The conditional mean and variance of E^ given E and { R k } more conveniently expressed as can be approximated as follows. Consider the complex products Y2kY:k from which we estimate e. For a given E, subtract S L {WNo){sind(?r~)I2/ m the corresponding phase, 2ae, from each p&ct to obtain the {1+0.5947(E,/N,)(sin?r~)~} (15) tangent of the phase e m with equality atE = 0. Equation (15) is plotted i Fig. 1 versus n k$~[yZky~ke-2T3' frequency offset E for values of EJN, equal to 11, 17, 23, tan [27r(i - E ) ] = and 29 d . Simulation results for N = 256, K = 96, and 8 B FSK modulation in a f a channel are included for validation. lt It can be seen that the bound of (15) is quite tight for small o t vdues of E but is &out 3 dB t o low a E = 0.5.

Observe that between the first and second DFT's, both the are altered inexactlythesameway,by IC1andthesignal a phase shift proportional to frequency offset. Therefore, if offset E is estimated using observations (20) it is possible to o obtain accurate estimates even when the offsetis t o large for satisfactory data demodulation. It is shown in the Appendix that the maximum likelihood estimate (MLE)of E is given by


a1 au a2 a25 a3 au a4 0.6 a5 Fig. & MLE relative frequency offset estimate versus relative frequency













Fig. 3. MLE relative frequency offset estimate error standard deviation versus E,IN,.

e = 0 and 0.45. The theoretical standard deviation from (25) is plotted for comparison. We conclude that (21) will give very accurate estimates ofthe relative frequency offset E. Under i - e M (1/2r) normal conditions for communications signalling, the accuracy is sufficient to correct E to well within tolerances (see (14) and Fig. 1) for negligible losses due any residual error in the offset estimate. The l i i t s of accurate estimation by (21) are I E ~ 50.5, that is, f1/2the intercarrier spacing. As E +OS, E may, due to noise and the discontinuity of the arctangent, jump to -0.5. (23) When this happens the estimate is no longer unbiased and, in practice, it becomesuseless. Thus, for frequencyoffsets At high signal-tcmoise ratios, a condition compatible with sucan initial acquisition exceeding one half the carrier spacing, cessful communikations signaling, (23) may be approximated strategy must be prescribed. One such strategy is discussed in bY Section IV. z - E w (1/274

Por 1 - E ( << 1/27r,the tangent can be approximated by its 2 argument so that

B. Frequency ofsset Estimution in a Multipath Channel

It is evident from (25) and (26) that the mean and variance of the offset estimate do not depend on the actual received frequency coefficients { R k } . Furthermore, if the symbol pair has been received through an unknown multipath channel, and as described in Section I it has been preceded by a periodic from which we lind that precursor of length NB2 N , the time spread of the channel E [ i - t I E , {Rk}] = 0. (25) then the carriers remain at their Steady state values throughout the duration of both symbols because the modulation values Therefore, for small errors, the estimate is conditionally un- are repeated. Thus, as no guard interval is required between biased. the symbol pair, the algorithm of (21) can be used without conditional variance the estimate easily determined modification. of is for (24). Fig. 4 shows six amplitude responses of a channel with five been multipaths whose arrival times have uniformly distributed I € 9 {&)I = 1/{(2~)2(Es/No>) (26) over an interval T, = T/16.The paths have equal weight and random phases so that the overall channel exhibits frequency where rm selective Rayleigh fadingas is evident f o the figure. Fig. 5 N-1 showsestimatesof E from (21) with PIg = N/16 for the same conditions as Fig. 2. It can be seen that the estimate is n=O unaffected by the multipath. is the total symbol energy. Since the total energy is the sum o the energies o the 2K 1 carriers, the error variance f f IV. ACQUISITTON of the offset estimate will in practice. be verylow.Fig. 3 shows the sample standard deviation of theerror in the relative In the event that the frequency offset is greater than f 1 / 2 fresuency offset estimate for 100 simulationtrials of (21) the carrier spacing, a strategy for initial acquisition to bring versus Es/No for 2K 1 =193 carriers and for offsets of the offset within limits of the algorithmmust be developed. the
/ K
\ \





14 OCTfBJ3R 1994



Ii \I
I ,





10 0

I 150

Fig.4. M I I ~ I ~ @ I






. a45 -



a4 0.3-


&fNo=l7 dB
0=5 dB

an -

/i :
aos oo F 5


,' . ,


,-. . a2 azs a3 a s a4 0.4s as MLE dative fresuency offset estimate versus relative frequency 0 . a multipath c. k b 1


,.., ..

repeated symbols by a repetition of the first full length data symboi or by the use of an AFC loop as shown in [6]. To illustrate, consider the following example for a digital audio broadcasting service. Assume an intercarrier spacing for the data stream of 1 kHz, and a frequency offset uncertainty in thesystemdominatedbythelong t r accuracy of the em oscillators in the receiver that heterodynereceived signal to te h IF and quadrature demodulate to obtain the complex envelope. t Assume VHF radio transmission a 150 MHz andoverall of 1 part in lo5. oscillator uncertainty (long term stability) Thus,, , S =1500 H and Afinit;al must be greater than3000 , z Hz. At regular intervals in the datastream of 1 ms (plus guard interval) symbols insert a short symbol of length 250 p s (4 lcHz carrier spacing) and repeat it once. From this xpeated shortened symbol estimateE. Assume E,/N, is 11 dB for the data symbols and that there are 200 carriers, so that E./N, is 34 dB. The shortened symbols have 1 4 henergy so E,/N,, is /t only 28 dB for the initial estimate E . From (25) we find that of DEinitial = 0.0063 SO that Dcdata = 0.025. From (15) we S e e errors of this magnitude causeIC1 resulting that residual offset in a signal-to-interference ratio of 24.3 dB in the data symbols for only 0.2 dB loss in overall S N R at EJN, of 11 dB. This exampleillustratesasituation for whichthe initial offset acquisition estimate with shortened data symbols is of sufficient accuracy that refinementof the estimate with longer symbols is not necessary.





v. DISCUSSION CONCLUSIONS AND We have seen that, as expected, frequency offset in OFDM causes serious loss of SNR of the dft outputs due primarily to We envision that,if continuous, theOFDM symbol stream will ICI. A lower bound for S N R has been derived and simulation be punctuated at appropriate intervals with qpeated symbols. results show that the bound is quite accurate for small offsets, A continuous symbol stream occurs in applications such as but about 3 dB too pessimistic as the offset approaches 112 di&d audio broadcast[5]. A second possibility is that OFDM the carrier spacing. modulation is used in session oriented digital data or voice An algorithm for maximum likelihood estimate communications such as digital radio [8]. Here, we envision of frequency offset using the dft values of a repeated that the session hitiation interval will include one or more symbol has been presented. It has been shown that for small lqeated symbols. error inthe estimate, the estimate is conditionally unbiased %e basic strategy for initial frequency offset acquisition, and is consistent in the sense that the variance is inversely in &her case, is to shorten the dft's and use larger carrier in the OFDM signal. proportional to the number of carriers spacings such that the phase shift does hot exceed f ~The Furthermore,boththesignalvaluesand . the IC1 contribute i i q k n c y offset in Hz is S = t / T = eAf Where Af is the coherently to the estimate so that it is possible to obtain very intercarrier spacing and T is the symbol interval. Let us assume ht accurate estimates even when the offset is too great, t a is that initial the frequency offset is no greater Then than there is too much ICI, to demodulate the data values. Since the estimation error depends onlyon total symbol energy, the algorithm works equally well in multipath spread channels. as However, it is required that the frequency offset as well determinesthe initial canier spacing,and corre- the channel impulse response be constant for a period of two spoutling dft lengths. If the average power of the shortened symbols. the kept m e , the varianceof the estimate of einitial Theaccuracyrequiredoffrequencyoffsetcorrectiondewill ' e larger than for the longer data symbols since there b be tolerated. Offset pends on how much residual offset can is less symbol energy. Also, the offset estimate error for the inducedIC1 can be treated quite satisfactorily as additional shortened symbols, since it estimates the fraction of carrier AWGN since its s o m e is the multiplicity of other OFDM spacing, corresponds to aproportionatelylargerfractional carriers thatare zero mean and uncomlated random processes. MLE estimate Note that the SNR defined in Section II [see (14)and (15)] offset f a the longer data symbols. However, the is .sa accurate t a in practice the initial estimate still may is just EJN, in the absence of offset. Thus we may interpret ht be a e u t . If not, it is refined by following the shortened (15) as theeffective E,/N, of thecarriers, o if divided dqae r


u m





by the number of bits encoded in each of the carriers, their ReSUired Eb/N, O f COurSe depends u p the effectiire &/No. moddation constellation. the fading statistics of the chaanel

To find the conditional density function in (A.8), note that

Yz = (Y1-W , ) H ( @+ wz )


if any,employed m the so that theforwarderrorcontrolcoding, OPDM system and the desired BER (see, for example, [ 5 , Yz = Y l H ( 8 ) Wz - W l H ( 8 ) . (A.10) Figs. 11-13]). The aquisiion range of the algorithm presented here is If W1 and W2 are Gaussian, zero mean whiterandom f / the intercarrier spacing of the repeated symbol, It is 12 vectors with variancea’,then the conditional density function independent o the modulation constellations chosen for the f mean valuevector carriers and whether the symbols are coherently or differen- in (A.6) is multivariateGaussianwith tially e c dd The AFC loop shown in [6] does not require a Y l H ( 6 ) and 2M x 2M covariance matrix no e . repeated symbol. However its acquisition range is f1/2m only K = E[(W2- W1H(@))t(W2 wlH(e))] 2 2 1 . = of the intercarrier spacing m-ary PSK. The initial frequency for (A.11) t of of offset a the time the initiation the communicationsession We note t a K is independent of 8, therefore, ht may be greater than 1/2 theintercarrierspacingandthus 6 = r n ~ [ f ( 18, Y , ) ]= %[~(8)1 (A.12) ~ z even outside the range of the MLE algorithm. In this event, a strategy is required for initial acquisition. We propose to use a pair of shortened data symbols whose carrier spacing with is sufficiently large to insure that the algorithm will operate J ( 6 ) = (Yz YlH(Q))(Yz YlH(8))’. (A.13) w t i its range. Dueto the low variance the initial estimate, ihn of further refinement will normally not be required. It may be Using the fact that advantageous to use shortened repeated symbols for tracking H(O)[dH(Q)/dO]’ [ d H ( @ ) / d @ ] H t ( 6= 0 (A.14) ) an AFC loop, because this offsetvariationstoo,insteadof reduces the time during which the channel must be stable. we canfind that






Let if complex values b

(Zk}be representedbyalength

2M row vector
= [ Z l R Z2R ’ . .Z M R = [ZR 211.
’* ’Z M I ]

d J ( Q ) / d Q = -Y2[dH(Q)/dQItYE Yl(dH(@)/d@]Yi. (A.15) Using (A.4), it follows directly that (A.15) is identically zero when 0 = Q such that
sin ( ~ ) [ Y Z R Y : R y z r % I l =



cos (

~ ) [ Y ~ ~ Y E zRY,tll. -y R


(A. 1)
(-4.2) (A.3)



Consider the random vectors

Y1=R1+W1 Y2 = RlH(6)

+ w2



E g],

(A. 17)

C = c o s ( e ) ~&

s = sin(e)I

is the maximum likelihood estimate (MLE) of 0.

(A.4) is a 2M X 2M rotationmatrix.Themaximumlikelihood estimate of the parameter 8, given the observations Y1 and Y2 (see, example, Sage and Melsa, p. 1961) is the value for [9, of 8 that maximizes the conditional joint density function of the observations. That is

8 = m$f(Y1, Y I e)] z
which can be written as


The authorwould l i e to thank Dr. S . Pupolin of University of Padova for his helpful suggestions to the original manuscript. The author also indebted to the is anonymous reviewers whose constructive remarks have improved thequality of this PaperREFERENCES

e = me!$f(Y2
W l

18, Y M Y l

I @)I.


But 9 gives no information about Y 1 ,that is

I Q) = f w d


so that

6 = m z [ f ( Y 2I 8, l ) ] . Y


[I] R. W. Chang, “Synthesis of band-limited orthogonal signals for multichannel data transmission,” Bell Sysr. Tech J., vol. 45, pp. 1775-1796, Dec. 1966. 121 S. Darling, “On digital single-sideband modulators,”IEEE Trans. Circuit Theory, vol. CT-17, pp. 409414, Aug. 1970. [3] S . B. Weinstein and P. M. Fibert, D t transmission by frequency’a ‘ a division multiplexing using the discrete Fourier transform,” IEEE Trans. Comnutn. Technol., vol. COM-19, pp. 628-634, Oct. 1971. [4] J. A. C.Bingham, ‘“ulticarriw modulation for data transmission: An idea whose time has come,” IEEE Cornnurn. Mug., vol. 28, pp. 17-25,

Mar. 1990.










~ ~~








NO. IO, O C m B W 1994

M. N r and R Halbert. “Principlesof modulation and ad channel coding for digital broadcasting for mobile receivers,” EBU Rev., no. 224. pp. 3-25, Aug. 1985’. [6] B. LeFlocb, R. Halbert-Lassalle, and D. Ca~telain.“Digital sound brosdcasting to mbde receivers,” IEEE Ifm.C o m m Ekc.. voL 35, n. 3, pp. 493-503, Aug. 1989. o [ 7 l M. K S i and D. Divsalar, “Doppler-ted differential detection of MPSIC’IEEE Tram. Commun.. vol. 37, no. 2, pp. 99-109. Feb. 1989. [8] D.C. Cox, W.S. G 0 and H.Sherry, “Low-power d g t l 1adi9asa 8d iia ubiquhm subscriber loop,” IEEE Commm Mag., vol. 29, pp. 92-95,

Mar. 1991.
[9] A. P. Sage and 1. L. Melsa, Estimation Theory with Applicarions to Commwicolionr and Control. New Y r :McGraw-Hill, 1971. ok
fir M~~

ppulM.Moose(M’79)wasbornin0maha,NE, on July 22. 1938. He received the B.S., M.S., and e in elecnical engineering in 1960, w Ph.D. d 1966, and 1970, respectively, f o thc University rm of Washington. Seattle. Since 1977 he has b e with the Department Of en Electrical andComputer Engineering, Naval Past,~ graduate School, Monterey, CA, wbere he is an Associstt Rofess~r Blectrical and COmpMer Enof gineoring. His research interests 82e in digital communications. He is a -founder of and consultant ~igital o m m u ~h., c c ~ ~m ~CA. ~ o w ~


Sign up to vote on this title
UsefulNot useful