OFDM Lecture 2

What is OFDM?

Orthogonal frequency division multiplexing is a modulation technique that transmits several constellation symbols simultaneously on dierent carriers. The baseband OFDM symbol takes the form: x(t) = 0
N 1 j2nt/(N T ) n=1 ck e

if 0 t N T + otherwise

where N is the number of carriers, T is the minimum sampling period, is the size of the cyclic extension to the OFDM symbol. x(t) is the inverse Fourier transform of a set of discrete constellation symbols.

Why OFDM?
1. Equalization made easier. 2. Can adjust transmission for inconvenient spots in the allotted band. 3. In some cases (Cimini, 85) reduce the negative eects of at-fading without coding (assuming fast at-fading.)

Who uses OFDM?

7. VDSL (DMT) 8. Flarion (Flash! OFDM) (802.20?) 9. etc.

Some history of OFDM

where B is the bandwidth of the lter such that F 2 (f )+F 2 (B/2 f ) = 1. So at this time, the OFDM symbol was made up of analog overlapping Nyquist lters. The OFDM symbol was

x(t) = odd n=

anm f (t nT ) cos((2t)(fc + mB/2)) + m bnm f (t nT T /2) sin((2t)(fc + mB/2)) + anm f (t nT T /2) cos((2t)(fc + mB/2)) + m even n= bnm f (t nT ) sin((2t)(fc + mB/2))

where T = 2/B. Consider one OFDM baseband signal (historic OFDM that is) (ignoring factors of 2 for now...) x(t) = m )t) + T m odd m b0m f (t T /2) sin((2)(fc + )t) + T m a0m f (t T /2) cos((2)(fc + )t) + T m even a0m f (t) cos((2)(fc + 2

b0m f (t) sin((2)(fc + xbb (t) = m odd m even mt = ej2 T f (t)(c0m ) +

m )t) T mt ej2 T (a0m f (t) jb0m f (t T /2)) + ej2 T (a0m f (t T /2) jb0m f (t)
mt

ej2 T f (t T /2)(d0m )
m

F (f

mt )(c0m + ejf T d0m ) T

a0m if m odd a0m if m even and d0m = jb0m if m even jb0m if m odd This means that the in-phase component and quadrature component are modulated by dierent lters. The quadrature lter is out-of-phase and has minimal interference with the in-phase lter. where c0m = (b) Weinstein and Ebert 1971. Paper describing digital implementation using FFTs. Looks more like the modern OFDM symbol described above. i. Avoid expensive, clunky analog lters by using the FFT. ii. Use of cyclic prex mentioned but not highlighted. iii. Windowing of the time-domain to mitigate high-frequency. (c) Peled and Ruiz, Frequency domain data transmission using reduced computational complexity algorithms Cyclic extension. 1980 (d) Keasler 1982 OFDM over the voice telephone band. (e) Cimini, 1985. Analysis and Simulation of a Digital Mobile Channel using Orthogonal Frequency Division Multiplexing i. Flat and fast fading channel. ii. Desires fast fading plus a frequency domain equalizer to achieve frequency diversity. 3

This idea is somewhat similar to the dumb antennas proposed by Tse et al. By increasing the amount of fading, the diversity increases. The higher the diversity, the lower the chance that the symbol is stuck in a fade. iii. Inserted pilot symbols to estimate the channel. (f) Rault, Castelain, LeFloch, 1989. The Coded Orthogonal Frequency Division Multiplexing (COFDM) Technique and its Application to Digital Radio Broadcasting towards Mobile Receivers Proposed European DAB system using slow-fading (channel static over 1 symbol) and coding for diversity gain. i. 16 stereo programs ii. 288 kbits/sec each iii. 4 PSK with rate 1/2 convolutional code and an outer (cyclotomatically shortened ) Reed-Solomon code with rate (288, 346). 448 carriers each with a bandwidth of 15.635 kHZ. 16 s cyclic prex. (g) DSL (a wireline sidetrip) In the early 90s discrete multitone modulation proposed by Cio et al. for DSL applications. i. Relatively static channel periodic handshaking allows knowledge of the channel at the transmitter. ii. Energy distribution and bit allocation (variable size constellations) on each tone. (Water-pouring) algorithm P. Chow, Cio and Bingham. iii. Time-domain equalizer to shorten the length of the cyclic prex J. Chow, Cio and Bingham. iv. ADSL asymmetric Digital subscriber loop. Most trac in the downlink. FDD, 256 carriers. 0.1 MHz BW uplink, 1 MHz BW downlink. v. TEQ time-domain equalizer shortens the length of the channel to allow a shorter cyclic prex. Late 90s VDSL (very high rate digital subscriber loops Zipper FDD) Mikael Isaksson and Frank Sjberg. o 4

(h) Now DAB was broadcast. One problem as we will see with OFDM is the requirement for synchronization within an OFDM symbol. This means that TDMA is the simplest way to do point-to-point. This leads to wireless LANs and HiperLan2 and 802.11a chipsets being built around 2000. a (i) In 1996, Telia, Sony and Lule University of Technology worked on an OFDM proposal for the 3G standard. The OFDM symbol was shared in both uplink and downlink by several users. Issues with this system included Peak-to-average power ratio, uplink synchronization, overhead, etc. (j) HDTV (DVB-T (terrestial ) Coded OFDM around 1997. (k) Wireless MAN (802.16) (metropolitan area network) OFDM standardization started 2001. Contains a TDD and FDD (similar to the 3G OFDM proposal) OFDM scheme. (l) Mobile Wireless MAN (802.20) OFDM with multi-user diversity (Flarion Technologies). Standardization started in 2002.

OFDM basics
1. AWGN channel
1 NT
kt N 1 j2 N T k=0 ck e

x(t) = Suppose we receive

t [0, N T ] otherwise

y(t) = x(t) + n(t) = n(t)

+ n(t) t [0, N T ] otherwise

The matched ltered signal is 1 NT

y(t)ej2 N T dt = 5

nt

N 1

ck (n k) +
k=0

1 NT

NT 0

n(t)dt

= cn + n NT NT 1 n(t)n(s) dtds] V ar( ) = E[ n NT 0 0 NT NT N 1 0 = (t s)dtds NT 0 2 0 N0 = 2 So the matched ltered output is ck + n with SNR
Ex N0 2

Now suppose we have lter at the receiver that band-limits the signal, 1 f [ K , K ] T T H(f ) = for some large K. 0 otherwise Then consider the ltered signal yf (t) = xf (t) + nf (t). If we ignore the fact that the frequencies outside the lter of our signal were cut o, then E[nf (t)nf (s) ] = K2 K K E[ n(t )sinc( )n(s ) sinc( )d d] 2 T T T K 2 N0 K K = (t s + )sinc( )sinc( )d d 2 2 T T T K (s t + )K K 2 N0 sinc( )sinc( )d = 2 2 T T T N0 K (s t)K = sinc( ) 2 T T
N0 K . 2 T

So the variance of the ltered noise is

If we sample the ltered received signal every T seconds (and ignore any truncation in the spectrum), then we have y(n) = 1 NT
N 1 k=0 N 1 n=0 N 1 k=0

ck ej2 N + nf (nT ) nf (nT )ej2nl/N nf (kT ) ej2km/N ]

kn

1 1 DF T (y(n)) = cl + T N 1 R(l m) = E[ N
N 1

1 nf (nT )ej2nl/N N n=0 6

= = =

1 N 1 N

N 1 N 1 n=0 k=0 N 1 N 1

N0 K (n k)T K j2(kmln)/N sinc( )e 2 T T N0 K (n k)ej2(kmln)/N 2 T

1 e N n=0 2 T N0 K = (m l) 2 T
Ex N K 20

n=0 k=0 N 1 N0 K j2n(ml)/N

So the sampled noise after the DFT is white! And the SNR is SN R = 2. Now lets make life more interesting and suppose you dont have an AWGN channel, but a dispersive channel, p(t). The dispersion will cause intersymbol interference. But the OFDM signal is much longer (N times) the single carrier system. So lets add a guard period to remove intersymbol interference. i.e. lets put zeros between successive OFDM symbols so that the OFDM symbol has a guard period of length Then what happens??
min t, N 1 k=0 ck 0
2(t)k N 1 j T k=0 ck e

h x(t) = =

h( )ej
min t, 0

2(t )k T

irrelevant!

d t [, (N T + )] otherwise
2 )k T

h( )ej

d otherwise

irrelevant!

which almost gives us the Fourier transform of h except for this term min t, in the limit of the integral. So, we have avoided intersymbol interference, but we now have intercarrier interference due to the channel distortion! Whats the x? A cyclic extension. So now let x(t) = 0
1 N T +
j2tk N 1 NT k=0 ck e

t [0, N T + ] otherwise

no cyclic extension

cyclic extension

Figure 1: Convolution with and without the cyclic prex Now what happens when we convolve with the linear channel? h x(t) = = =
N 1 k=0 ck 0 N 1 k=0 ck e N 1 k=0

h( )ej
0

2(t )k T

d otherwise d otherwise

irrelevant! irrelevant!

j2 kt T

h( )e
j2 tk T

j 2 k T

H(k/T )ck e irrelevant!

otherwise

Is the cyclic prex part really irrelevant? Probably not by the Theorem of Irrelevance its irrelevant in the sense that you want to use the ease and convenience of the DFT and the fact that this multi-tap time domain channel is now a 1-tap channel in the frequency domain. Another way of interpreting the cyclic prex is: convolution in the time domain is multiplication in the Fourier domain for continuous systems. But with the Discrete Fourier samples, multiplication in the frequency domain is cyclic convolution in the time domain. So by adding a cyclic extension, weve made linear convolution look cyclic. By adding the cyclic prex, we ensure orthogonality in the tones. 3. Practical issues.

1.5

1 Modulated OFDM signal

0.5

0.5

1.5 20

15

10

0 frequency index

10

15

20

Figure 2: OFDM modulated on carriers -7 through 8 (a) The OFDM symbol as written has innite spectrum. Its the sum of modulated rectangular signals. That means its Fourier transform is the sum of shifted sinc functions. Solutions? (a) Not using tones at the edge of the band see Figure 4. Loss of data rate symbol size is still N T + even though the data rate N0 is N T + where N0 < N . (b) Windowing If we multiply the OFDM signal by a window (as in Weinstein and Ebert) then we can reduce the spectral lobes. So you transmit x(t) w(t). What kind of window can you use? i. Window type 1 w(t) = + (1 ) cos( (t(Ts )/2) ) t [0, Ts ] (Ts )/2)

10

0 OFDM spectrum (dB)

10

15

20 20

15

10

0 frequency

10

15

20

Figure 3: spectrum of OFDM modulated on carriers -7 through 8 where Ts = N T + . Notice that w(t) + w(t Ts /2) = 1 when = 0.5 (Hanning window) w(t) + w(t Ts /2) = + (1 ) cos(

(t Ts ) (t Ts ) ) + + (1 ) cos( ) Ts /2 Ts (t Ts /2) = 2 + (1 ) cos( ) + (1 ) Ts /2) (t Ts /2) (Ts /2) cos( ) cos( )+ Ts /2) Ts /2 (Ts /2) (t Ts /2) ) sin( ) sin( Ts /2) Ts /2 (t Ts /2) = 2 + (1 ) cos( ) Ts /2) (t Ts /2) (1 ) cos( ) Ts /2) 10

10

0 OFDM spectrum (dB)

10

15

20 20

15

10

0 frequency

10

15

20

Figure 4: spectrum of OFDM modulated on carriers -5 through 6 with original spectrum with tones from -7 through 8

11

0.5

amplitude

alpha = 0.25 Hanning window Hamming window 0

0.5

10 time

12

14

16

18

20

Figure 5: General window with = 0.25, Hanning window = 0.5 and Hamming window = 0.54 = 2 In Figure 6 3 windows with dierent choices of are shown. The Hanning window has the good property that it goes to zero at the end of the symbol allowing for good frequency reduction. The Fourier transform of the Hanning window is sinc(f Ts ) (1 )/2 sinc((f 1 1 )Ts ) + sinc((f + )Ts ) Ts /2 Ts /2

The problem with this is the subcarriers are no longer orthogonal. Solutions? You can live with the intercarrier interference or you can extend the OFDM signal as in Weinstein and Ebert. t [Ts , 0] 1 t [0, Ts ] w(t) = 1/2(1 + cos(2 (tTs ) ) t [Ts , (1 + )Ts ] Ts 12
1/2(1 + cos(2 t ) Ts

10 Rectangular pulse Hanning Window

0 OFDM spectrum (dB)

10

15

20 20

15

10

0 frequency

10

15

20

Figure 6: Spectrum of OFDM signal with rectangular and Hanning window This preserves orthogonality while bringing down the sidelobes. But there are 2 problems with this window: A. Loss of data rate on the order of Ts extra overhead. (Dont need 2Ts extra time as adjacent symbols can overlap.) B. At the receiver you do something non-linear you cut o the cyclic prex. Suppose you have another OFDM signal transmitting simultaneously (cross-talk from an adjacent channel or far-away). Then when you remove the cyclic prex, if the other signal is synchronized, then the highfrequency components can leak back into your OFDM signal as youve also truncated the other guys signal.

A little bit on water-lling

If you know the channel at the transmitter, then you can adjust the power on each channel. 13

Now in many wireless systems, they have power control. The mobile sends back control information to the basestation, telling the BS, among other things, its power. The BS adjusts the transmit power to each mobile turning down the power on one mobile if its SNR is too high, turning up the power on another mobile if its SNR is too low. Power control is a politeness protocol users who have too much signal, turn down their power so that others may have less interference. Power control acts like an equalizer across the dierent users. Where the SNR is low, the power is turned up; where the SNR is high, the power is turned down. For a single-user, this strategy of transmitting the power so that the received power on all tones is equal is not necessarily the way to go. Waterlling is a way to assign power where it can do the most good. It comes from the denition of capacity. That is, suppose you have a vector channel with receive SNR (SN R1 , . . . , SN RN ) when the transmit energy on each element of the vector is 1, then the capacity of the vector channel is 1 2 log2 (1 + Pn SN Rn )

where Pn is the transmit power. We dont control the channel SNR but we do control the transmit power. The goal is to nd the power distribution that maximizes the capacity while keeping the sum n Pn xed to some constant (our budget power). Calculus and Lagrange multipliers to the rescue! d1 2
n

log2 (1 + Pn SN Rn ) + ( n Pn C) = 0 dPn 1 SN Rn + = 0 2 log(2) 1 + Pn SN Rn (1 + Pn SN Rn ) = SN Rn 1 Pn = SN Rn max( SN1Rn , 0) = C.

1 where = 2 log(2) .

Because Pn 0, the goal is to nd such that Then you nd Pn = max( SN1Rn , 0). 14

Notice that when the SNR is very high, Pn and when the SNR is very low, Pn 0. Also when C the total power allowed is very high Pn for all Pn When the total allowed power is very low, (with frequency selectivity) many of the subcarriers may have zero power. Note that waterlling is based on capacity which implies a code lurking in the background. But its also a way of saying transmit the power where it will do the most good. Now if you have a frequency selective channel and you transmit equal power on each tone then the tone with the lowest SNR will bring down the average probability of error. But if you focus on the good tones, then they can save the day. In a single-carrier system, the symbol is spread across the whole spectrum, the high SNR parts and the low SNR parts. The equalizer picks out the good parts of the spectrum in some sense. So a major failing in OFDM is that without some sort of correction at the transmitter, water-pouring alone, water-pouring + a code, or a code only (given no knowledge of the transmitter) the low SNR subcarriers will increase the probability of error of the bits to a point where the system can break down. So with OFDM you either need a code, some sort of distribution of power or both.

15