The Cyclic Prefix of OFDM/DMT – An Analysis

Werner Henkel, Georg Tauböck
Telecommunications Research Center (ftw.)
A-1220 Vienna
Henkel,Tauboeck
@ftw.at
Per Ödling, Per Ola Börjesson, Niklas Petersson
Lund Institute of Technology
SE-221 00 Lund
Per.Odling,POB,Niklas.Petersson

@tde.lth.se

Albin Johansson
Ericsson Telecom AB
SE-126 25 Stockholm
Albin.Johansson@etx.ericsson.se

Abstract— We address the impact of a too short cyclic prefix on multi-
carrier systems such as Orthogonal Frequency Division Multiplex (OFDM)
and Discrete MultiTone (DMT). The main result is that the intersymbol in-
terference (ISI) and intercarrier interference (ICI) may be spectrally con-
centrated and analytical expressions showing this are given. A practical
implication is, e.g., that the cyclic prefix in some xDSL systems can be sur-
prisingly short, as shown in one example of ADSL transmission.
Keywords— Intersymbol interference (ISI), intercarrier interference
(ICI), multicarrier transmission, cyclic prefix, time-domain equalization.
We give a mathematical analysis of the ISI and ICI for a sys-
tem with insufficient length of the cyclic prefix for the case when
all the tones are used to transmit data in one direction, for in-
stance in simplex communication or for a time-division duplex
(TDD) system. This is done in Section II. In Section III, we
present an example where the interference is strikingly concen-
trated in frequency.

 "#%$'&
I. I NTRODUCTION   ! ( 
*

)




,+-
N this paper, we address the impact of a too short cyclic pre- DFT
fix in multicarrier systems [1]. The cyclic prefix removes the

intersymbol interference (ISI) and intercarrier interference (ICI) Fig. 2. The system under analysis: the receiving end of a multicarrier system
[2]. The introduction of the cyclic prefix of length  , see Fig. 1,
gives a constant capacity loss, since the channel does no longer
carry data for short periods of time. As such, one would like to II. S IGNAL MODEL AND I NTERFERENCE C ALCULATION
minimize the length of the cyclic prefix, preferably maintaining
performance. Common wisdom is to choose the cyclic prefix The interference that we are studying consists of two parts:
to be of roughly the same length as the channel (or system) im- the intersymbol interference (ISI) and the intercarrier interfer-
pulse response, thus eliminating ISI and ICI. It is also well es- ence (ICI). We start with analyzing the ISI, which can be de-
tablished that, if the tail of the impulse response contains only rived in a more intuitive fashion. We will also see that the ICI
very little energy, it has little impact and can be considered zero is of very similar structure, and can be described with the same
mathematics under our present assumptions.
allowing a shorter cyclic prefix. We show that, furthermore, the
 / through a chan-
ISI and ICI can be spectrally concentrated and sometimes have Consider the transmission of symbols  .
nel with the impulse response of length -0 . We ex-
limited or even zero impact on performance. Although this is 
not completely unknown among designers of DMT digital sub- tend our notations with an index 1 on the input sequence . 
scriber loop (DSL) systems, it has rarely been given a thorough as we need to distinguish / between the input corresponding to
analysis. Our attempt here is to provide an intuitive and imme- the present symbol 2  / , which gives rise to the ICI, and
 
diate understanding of the mechanisms involved. /previous data .
the which causes the ISI1 . The signal
.  is assumed
98  to<8>=?be  mean with a variance of 3564 and
=>=>8 zero
.  7 .  ;: .  0 @ are assumed to be pairwise
 
 
 uncorrelated. The received signal, which is to be processed by
  the FFT,
 

  BACD/E /BAGFIH 9NO /   NO<8
         
 2 . . (1)
JLKIM
Fig. 1. Symbol structures
P
We do not consider the case where the evaluation frame at the receiver is
This work was partially financed by the Telecommunications program of the positioned such that it has post- and precursors from both neighboring frames. In
Swedish National Board for Industrial and Technical Development (NUTEK) a real system, this could be the case. However, this would make the math lengthy
and by the Austrian Kplus program. It was initiated by Per Ödling while being and difficult to follow. Thus, in order to outline the fundamental properties, we
with ftw. decided to restrict this presentation to only postcursors from a preceding frame.

0-7803-7257-3/02/$10.00 2002 IEEE. 22-1

  / _
where . denotes the concatenation of all .  up to the -0 G   , i.e., “’ -0   C . This is a reasonable
present symbol. A part of this signal will then be ISI from the assumption. One would certainly not design a DMT (OFDM)
previous symbol. Figure 3 illustrates the tails of the impulse system with a frame length shorter than the channel impulse re-

response that are not covered by the cyclic prefix. Note that in sponse. While is increased, the upper limit of the inner sum

contrast to Fig. 1 not only the guard interval but also its counter- -0 ”  is decreased–A accordingly. The lower limit  @ is
„A
part at the end of the frame is highlighted (hatched) and labeled reached -0 •
 I
    when
;
 ˆ0   —: . If we now
ƒ
with G. The ISI that affects a symbol 1 from a symbol 1 @ is take the sum over  as outer sum, we ƒ obtain the upper limit for
   — ^BA ˆ0  ƒ `
U the inner sum over as ˆ0   ˜:U
/TA F H 9NO    NO98XWZY\[^] `_  _  : =
-0   a
and thus,
2RQS?Q . U  ‚ U  
JLK   %

 V
<qrTA   ƒ„ U  ƒ
 Ov ™x z V}| =
F (2) p Q/S?Q
F H
I
.
F H  <u { (6)
In (2), we actually used the concatenated symbols in order to ‚K    V K M
I
F
simplify the description. ƒ  „A›š
Substituting yields
impulse responses
U U
9qrœA  7ƒ }FIH D /š9u wvyx^{Ÿz ž   ‚  |
F H
I
d p Q/S>Q .
h ih ih ih
i
h ih ih ih
f g
g f g
f g
f gf 
f Gg
b c
f gf c
b c
bc c
b cb
b Gcb cb d ded ded ded
e ‚K 
'K ‚  
i G g eeGee F U
A  ƒ79u  v x^{ z ‚ |¢¡
FIH
. 
evaluation frame

symbol jIk;l symbol j ‚K 

U F
Fig. 3. ISI from symbol m5no to symbol m
¡ F H D/š9u wv x^{ z  | =
(7)
 < qr , then becomes 'K ‚
The signal after the DFT, p
U K£ ¤¥ ž |  
p 
9qrsA t
2  /9u wvyx^{ z }V | Note that the expression for ¦ ‚
<qr
is actually the DFT of the
V KIM 9qr
U   tail of the impulse-response. The power spectral density  Q/S>Q
A t F H 9NO    NO<u Ov ^x { z }V | = after the FFT due to the intersymbol interference follows to be
.
 QS?Q <qr§A©¨«ª pRQ/S>Q 9qr p—Q/S>¬ Q 9qr™­
(3)
V KIM JLKIM (8)
With a cyclic prefix of length  , the residual ISI is determined U U
as A F H
I FIH <qr <qr
  U ¦ ‚ ® ¦ ‚¬ x
‚Š®¯K    ‚ xK 
 
<qrsA t 9NO /   ON <u Ov x^{ z }V |
F H F F
p~Q/S?Q .  ƒ <­ˆu v ^x  { z ž ‚Š®  ‚ x   |
(4) ¨«ª   ƒ 
V KUM JK €-V . . ¬  4
T (9)
  FUwV
A t FIH D/ƒ     „ ƒ <u Ov x^{ z V}| =  
V KUM ‚K  . (5)
A F H <qr 4
F 3 64 ¦ ‚ (10)
‚‘K 
 5
 ° °
F
…ˆ‡ n`o
impulse responses
Ž
µ ¶•
¶• µ ²• ± ¶µ ²• ± ²• ± ²• ± ²± · ¸•
¸• · ¸· • ³´ ´• ³ ´• ³ ´³
¶µ•¶• µ ²• ± ¶µ ²• ± ²• ± ²• ± ²± ¸·•¸• · ¸· • ³´ ´• ³ ´• ³ ´³
…`† o ¶µ•¶• µ H²• ± ¶µ ²• ± ²• ± G²• ± ²± ¸·•¸• · H ¸· • ³´ ´• ³ G´• ³ ´³
… ‡ nŠ‰ evaluation frame
‹ oŒ 
Fig. 5. Hypothetic extension ”H” of the guard interval ”G”
…ˆ‡ n … n‰
Œ
no for computing the ICI

Fig. 4. Summation area in (5) Remains now to calculate the ICI term. ICI would not be an
issue if the guard interval would be big enough to fake the chan-
 Next, we interchange the two sums. Due to the dependence of nel convolution as a cyclic convolution within the frame that is
in the inner sum, we have to investigate the summation over evaluated by the receiver. Thus we just assume an extension of
‘8<ƒ„
the pair in some more detail. Figure 4 shows the effec- the guard interval that would null the ICI and compute its im-
tive pairs that are used in the two sums. There, we assume that pact on the evaluation frame. Its negative would then be the ICI

22-2
signal. The hypothetic frame extension is shown in Fig. 5. The h(n)
0.45
negated ICI signal is then similar to (2) and can be written as
0.4
U
/BA  FIH D<NO  ^  ON º8 0.35

2RQ¹OQ . mod  (11)

JLK   % V 0.3

F
_  _ =
0.25

-0   a: 0.2

0.15
Note that this time, (11) refers to only one input signal block
.  from which the hypothetical guard interval is taken. Corre- 0.1

spondingly, we count modulo  to stay within this frame. In 0.05

DFT domain, we obtain
0
0 50 100 150 200 250
9qr‘A sample no.
pRQ/¹wQ
 U U Fig. 6. Impulse response of a 4 km, 0.4 mm loop

 I
F H F F I H 9NO    ON 5u Ovyx™{ z }V |
. mod 
V KUM JLK   % V 2
|H(q)| /Hmax [dB]
F (12)
0

−5
which corresponds to (4). If we further follow the steps in the
derivation of the ISI result down to (10), we see that the only −10
differences are the minus sign and that we count modulo  ,
−15
staying within the same frame. The minus sign disappears with
the squaring when computing the power spectral density. If we
consider .  instead of . , does not change the result, either, so
−20

−25
that we obtain
 Q¹OQ <qrBA  Q/S>Q 9qr (13) −30

−35
and U
¡
 Q¹OQ  Q/S>Q A : 3 64 I F H <qr 4 = −40

° ¦ ‚ ° (14)
0 64 128 192 256

‚K    carrier no.

F Fig. 7. ISI and ICI power spectral densityP according to (10) and (13) normalized
ISI and ICI have the same power spectral density. This is an im- to its maximum ( ¾ ¿—ÀÁÃÂ%¾ÅÄˆÆ Ç HyÊ È P ¾ ¿ É ÀÁþÅÄ )
ÉT
portant first result that has relevance for practical time-domain Ç}Ë
equalizer algorithms (see, e.g., [3]).
R EFERENCES
III. E XAMPLE [1] Bingham, J.A.C., ‘Multicarrier modulation for data transmission: An idea
whose time has come’, IEEE Communications Magazine, vol. 28, no. 5,
As a practical example we choose an ADSL transmission with
A¼» : A¼½ : pp. 5–14, May 1990.
an FFT length of  and a guard interval of  [2] Peled, A. and Ruiz, A., ‘Frequency domain data transmission using reduced
computational complexity algorithms’, Proc. IEEE ICASSP, pp. 964–967,
samples. We select the impulse response of a 4 km long loop of Denver, Colorado, 1980.
German 0.4 mm cables, which is shown in Fig. 6, cutting non- [3] Henkel, W., Kessler, Th., ‘Maximizing the Channel Capacity of Multicar-
causal precursors that are due to the underlying cable model. We rier Transmission by Suitable Adaptation of the Time-Domain Equalizer’,
IEEE Transactions on Communications, Vol. 48, No. 12, pp. 2000-2004,
see that it certainly has a portion exceeding the guard interval. If Dec. 2000.
we compute the ICI and ISI components according to (10) and
(13), we see in Fig. 7 that the noise is predominantly disturbing
low-frequency components. It has long been realized that the
signal-to-noise ratio decreases near DC, which was considered
partly to be due to the leakage effect of the DFT which folds
high-frequency noise components into the low-frequency range.
The results in here now show that noise due to ISI and ICI is
concentrated around DC as well.

IV. C ONCLUSIONS
We derived closed formulas for the intersymbol and inter-
channel interference power spectral density and found that they
are actually the same. From a typical example we concluded
that this noise will be concentrated around DC.

22-3