Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

1882 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 8, NO. 12.

DECEMBER 1990

Amplifier Induced Crosstalk in Multichannel Optical


Networks

Abstrucf-We study the effect of crosstalk introduced due to gain duce the amount of link budget improvement actually ob-
saturation in a n optical amplifier when it is used for amplifying mul- tained.
tiple channels in a wavelength division multiplexed network employing
ON-OFF keying with direct detection. The system power penalty is quan-
Two mechanisms introduce undesirable interactions be-
tified as a function of the amplifier input power, the number of chan- tween different channels: four-wave mixing or inter-mod-
nels, and the extinction ratio. We consider two cases: the low-data-rate ulation distortion [6], [7] and gain saturation [8]-[ 131.
case, where the product of the data rate a n d the gain response time of Their impact on system performance is a strong function
the amplifier is small, typical of semiconductor-laser amplifiers, a n d of the lifetime in the amplifier. The carrier recombination
the high-data-rate case, where the product is large, typical of doped-
fiber amplifiers.
lifetime in SLA’s is of the order of 1 ns, while the fluo-
We find that the system penalty due to the presence of other inter- rescence lifetime of the Erbium ions in DFA’s is around
fering channels can be decomposed into two components: a term aris- 10 ms.
ing from the steady-state reduction in the amplifier gain due to the Intermodulation distortion is significant only if the op-
larger average input power, a n d a crosstalk component because of the tical carriers are spaced so closely that the carrier density
variation in the gain due to the randomness of the total input power
around the mean. Results indicate that when the number of channels
or the ion density can respond at the beat frequency. This
is small, the crosstalk component dominates, but when the number of is never the case with DFA’s for any realistic channel
channels is large the component due to the steady-state gain reduction spacing. In filter-based WDM networks employing direct
is more significant. For the high-data-rate case, the crosstalk penalty detection, channel separations are usually much greater
is no longer present, a n d only the component due to the steady-state than a few gigahertz, and hence, even with SLA’s, this
gain reduction, is retained.
We study the performance of frequency-shift keying as a n alterna- effect is not expected to present a problem, and we do not
tive to ON-OFF keying a n d suggest techniques for reducing the crosstalk consider this further.
penalty. Finally, we show that using amplifiers in a passive star net- Saturation induced crosstalk is known to be a potential
work allows a significantly larger number of stations to be supported. problem in WDM networks. It occurs because of the de-
pendence of the carrier or ion density in the gain region
of the amplifier on the input intensity. As the input inten-
I. INTRODUCTION sity is increased, carriers or ions are depleted from the
AVELENGTH division multi-access (WDM) is active region, resulting in a reduction in the gain. Thus
seen as an effective technique to make use of the the gain seen by a particular channel is affected by the
vast bandwidth offered by the optical fiber in order to build intensity levels on the other channels.
high-speed computer networks. The number of stations We determine the error rate performance of a WDM
that such a network can support is limited by the link bud- system using ON-OFF keying and direct detection as a
get, which can be improved by using optical amplifiers function of the number of channels, input power per chan-
[I]. nel, and the extinction ratio. We define the system power
Wide-band traveling wave semiconductor laser ampli- penalty as the difference in power required per channel to
fiers (SLA’s) [2]-[4] and doped-fiber amplifiers (DFA’s) obtain the same error rate ( lop9typically) when all the
[5] are suitable for multichannel networking applications interfering channels are present, and when only the data
because they are capable of amplifying several channels channel is present. We consider the case when the input
at different wavelengths at the same time. The improve- data rate is much smaller than the reciprocal of the life-
ment in the link budget obtained using these amplifiers time and also the case when it is much larger. The system
can be used to support more stations, higher bit rates per penalty due to the presence of other interfering channels
station, and to increase the distance between stations. can be decomposed into two components. The first com-
However, amplifier induced crosstalk and noise can re- ponent arises from the steady-state reduction in the am-
plifier gain due to the increase in the average input power.
We shall refer to this component as the saturation com-
Manuscript received July 18, 1989; revised March 28, 1990.
R. Ramaswami is with IBM T. J . Watson Research Center, P.O. Box ponent. The second component is the component arising
704, Yorktown Heights, NY 10598. from the variation in the gain due to the randomness of
P. A. Humblet is with the Department of Electrical Engineering and the total input power around the mean. We shall refer to
Computer Science, and Laboratory for Information and Decision Systems,
Massachusetts Institute of Technology, Cambridge MA 02 139. this component as the crosstalk component. Our results
IEEE Log Number 9038818. indicate that when the number of channels is small, the
0733-8724/90/1200-1882$01.OO O 1990 IEEE
~

RAMASWAMI AND HUMBLET: AMPLIFIER INDUCED CROSSTALK IN MULTICHANNEL OPTICAL NETWORKS 1883

crosstalk component dominates, but when the number of the number of stations that can be supported in a WDM
channels is large, the component due to the steady-state network using amplifiers. Section VI summarizes our re-
gain reduction becomes significant. For the high-data-rate sults.
case, the crosstalk penalty is no longer present and only
the component due to the steady-state gain reduction is 11. LOW-DATARATES
retained. With DFA’s, because of the long fluorescence The system under consideration is shown in Fig. 1. It
lifetime, we will be operating in the high-data-rate regime consists of the amplifier followed by a tunable optical fil-
in most cases of interest, implying that the crosstalk pen- ter before the photodetector and subsequent receiver pro-
alty will be smaller when compared to SLA’s. We then cessing. The input to the amplifier consists of several ON-
apply these results to determine the number of stations OFF modulated signals, each at its unique wavelength. The
that can be supported in a passive star WDM network tunable filter picks out only the desired channel for recep-
using amplifiers. tion and rejects all the other channels. It also serves to
Experimental results on crosstalk for two channel sys- band limit the amplifier noise. The filter is characterized
tems have been reported extensively for SLA’s [ 111, [ 121, by its bandwidth and its loss, but is otherwise assumed to
[14]. Typically two channels, one modulated and the other be ideal.
unmodulated, are amplified and the amount of modulation Given the relationship between the amplifier gain and
appearing in the unmodulated channel is used as a mea- the input power, using the input statistics, we can com-
sure of the crosstalk. In one multichannel direct detection pute the statistics of the gain, and consequently, those of
experiment [15] 20 channels were amplified using an SLA. the output. Using these results along with the amplifier
Of these, 19 were modulated at 50 MHz, and the desired noise statistics, the system bit error rate can be evaluated.
data channel at 1 Gb/s. No crosstalk penalty was ob- We present illustrative results assuming values for the
served because the input power levels on each channel parameters as shown in Table 1. The receiver electrical
were kept sufficiently low so that amplifier was operated bandwidth of 100 MHz corresponds to an operating data
well below saturation. While a small penalty in receiver rate of approximately 200 Mb/s, which is much smaller
sensitivity (0.5 dB) was incurred due to the extra ampli- than the reciprocal of the carrier lifetime in SLA’s but
fier noise, the net amplifier gains obtained (6-8 dB) en- much larger than the reciprocal of the fluorescence life-
abled the system to support four times as many stations time in DFA’s, allowing us to compare the performance
in a multiwavelength network based on a passive optical of the two types of devices. We assume a loss-free optical
star at the hub. filter with a bandwidth of 10 GHz ( ~ 0 . nm), 1 unless
Recent investigations of the gain saturation phenomena otherwise stated. The thermal noise current value of 3.8
in DFA’s [ 161-[ 181 have confirmed that the crosstalk pen- PA/& corresponds to a receiver sensitivity of -39
alties with these amplifiers are small because of the long dBm at 200 Mb/s for error rate, typical of direct-
lifetime. detection receivers [ 191.
Previous analyses of crosstalk [8], (91, with the excep-
tion of [lo] have dealt with two-channel systems and have A. Amplijier model
not attempted to determine the impact of the crosstalk on
system performance. In [IO], the system penalty was de- Consider a traveling wave amplifier with negligible re-
termined for the multiple-channel case using certain ap- sidual facet reflectivities. When the input data rate is much
proximations. While evaluating the penalty, the decision smaller than the reciprocal of the lifetime, the carrier (or
threshold was set corresponding to the median received ion) density attains steady state within a small fraction of
signal levels, which is not optimal, as will be seen in Sec- the bit period. For this case, the rate equation reduces to
tion 11-E. Moreover, for comparison purposes, the noise a steady-state charge conservation equation, and can be
in the two cases (when a single channel is amplified, and solved to yield the amplifier gain G as a function of the
when multiple channels are amplified), was assumed to input power P [lo], [20]
be the same, which is not the case. If the power per chan-
nel is kept constant, then the total power is proportional
p=- psat
In -.Go
to the number of channels. Thus, the gain is smaller in G-1 G
the multiple channel case, and hence the amplifier noise
is also smaller. Finally, the system penalties evaluated at Here, Psatis the internal-saturation optical power, and
different error rates are different, and hence, complete er- Go is the maximum amplifier gain, corresponding to the
ror rate curves are required in order to be able to deter- case when the input power is zero. It can be shown easily
mine the penalty at any operating point. that G is a monotonically decreasing function of P , and
The remainder of the paper is organized as follows. In that the output power Po,, = GP is a monotonically in-
Section 11, we derive the theory for the low data-rate case, creasing function of P . Fig. 2 shows the amplifier gain
based on using a simple amplifier model, and present il- plotted against the input power for a typical 1.5-pm am-
lustrative results. In Section 111, we consider the high- plifier (PSat= -6 dBm, Go = 1800). Even at low-input
data-rate case. Section IV discusses schemes that can be powers the gain dependence on the input signal is signif-
used to reduce the crosstalk. In Section V , we determine icant.
1884 JOURNAL OF LIGHTWAVE TECHNOLOGY. VOL. 8. NO. 12, DECEMBER 1990

E Amplifier Filt

Fig. 1. Block diagram a multichannel system using an amplifier.


Receiver
of asynchronous users will be discussed later). Let Pi, be
the average input power per channel and r the extinction
ratio defined as
r = Po/Pl
where Po is the power transmitted for a ZERO bit and PI
is the power transmitted for a ONE bit. Then
(2)

2Pin
PI = ~
(3)
30- l + r

25- and
2 rPin
g20- Po = - (4)
v

-c 15- 1 r' +
The input power P is a binomial random variable. It takes
1 a-
on N +1 possible values with probabilities given by

5-
Pr { P = kP, + (N - k)Po}
0 I 1 I I I I I

Fig. 2. Amplifier gain versus input power for a typical 1.5-pm traveling
Using (1) and ( 5 ) , the distribution of the amplifier gain
wave semiconductor amplifier. Saturation power P,,,= -6 dBm, max- can be computed. Fig. 3 shows the probability mass func-
imum gain Go = 1800. tion of the gain for Pin= -50 dBm, N = 10, and r = 0.
C. Noise Statistics
The amplifier induces spontaneous emission noise. The
Parameter 1symb01 Value detection problem consists of detecting an optical signal
field in an additive Gaussian noise field using a photode-
Optical bandwidth B. 10 GHz tector that responds to the intensity of the optical signal
Electronic charge e 1.6 x 10-19 coui. 1211. The noise field beats against the signal and against
Optical gain G itself, giving rise to noise components referred to as the
Gain with zero input G" 1800
signal-spontaneous beat noise and spontaneous-sponta-
Photon energy hv 0.55 eV
neous beat noise. In addition, shot noise and thermal noise
Thermal noise current I la 13.8pAIZ
components are also present.
Received simal level IL.andL, I The thermal noise, shot noise, signal-spontaneous
No. of interferingchannels IM noise, and spontaneous-spontaneous noise average pow-
Spontaneous emission factor IN,. 11.4 ers at the receiver are, respectively (see Table I, for a
Shot noise power IN,,, description of the symbols and their chosen values):
Simal-spontaneous beat noise wwer IN.-.. I
Spontaneous-spontaneous beat noise power IN, -,
Amplifier saturation wwer I P.", 1-6 dRm
IAverage amplifier input power per channel IP," I -1
Amplifier input power per channel for ONE
]PI I
Amplifier input power per channel for ZERO IPo
Power penalty 1 PP
Extinction ratio lr. I O or 0.05 ^ ^

Decision threshold T

Detector quantum efficiency v 1

Output coupling efficiency Tar, 0.26


Total noise s t a n k d deviation a.. These statistics are derived in Appendix A in a simpler
and more general manner than has been found in the lit-
erature so far 113, [19], 1221.
B. Gain Statistics The signal power is given by
The amplifier input is assumed to consist of the sum of
N ON-OFF modulated signals of equal input power. We
assume that all the signals are bit synchronous (the effect
RAMASWAMI AND HUMBLET: AMPLIFIER INDUCED CROSSTALK IN MULTICHANNEL OPTICAL NETWORKS 1885

0.250.

!
0.200-

E 154 ' 3
E
g o . 150- E
=
n
n
--
0
m
10-

~o.,oo. -
f 0-
?
f -5-
0.050. i-10-

0.000-
500
-J 700
900
1100
Gain
1300 1500 1700 1900
E-15-
Z
-20 1
-55 -50
I I

-45
I
-40
I

-35
I

-30
Amplifier input rlgnal power (dBrn)
I

-25
1
-20

Fig. 4 . Noise components at the receiver versus input signal power. Curve
Fig. 3. Probability mass function of the amplifier gain. Number of input
1: Thermal noise. Curve 2: Signal-spontaneous beat noise. Curve 3:
channels, N = 10, input power per channel P , , = -50 dBm. extinction
Spontaneous-spontaneous beat noise. Curve 4: Shot noise.
ratio r = 0.

We assume the noise statistics to be Gaussian. This is 0.250-1

partially justified because at low gains and low-input


power levels (low signal-to-noise ratios) the receiver ther- 0.200
mal noise is the dominant source of noise, while at high
gains or high signal powers (high signal-to-noise ratios)
the amplifier induced signal-spontaneous beat noise dom- --n
t0.150
inates, and both these components are Gaussian. In Fig. n
4,we plot the different noise component average powers to.100
(see (6)-( 10)) against the input signal level for parameter
values shown in Table I, when a single unmodulated
channel is being amplified. The amplifier gain is large 0.05C
enough so that the signal-spontaneous beat noise is the
dominant component, although the spontaneous-sponta-
neous beat noise can become significant if the optical .OO .10 2 0 .30 .40 S O .60 .70 .EO .90 1 .OO
Elrctrlcal signal Irvrl (normallzrd to maximum Iwel)
bandwidth is large. Note that the spontaneous-sponta-
neous beat noise is not strictly signal independent, be- Fig. 5 . Signal levels after photodetection along with associated probabil-
ities. Number of interfering channels, M = 10, r = 0.05.
cause it depends on the gain, which depends on the input
signal.
From Fig. 5 , one might think that under certain con-
D. Error-Rate Calculations ditions, the maximum ZERO and minimum ONE levels may
Let us now consider a single channel and the effect of overlap, leading to bit-error-rate floors. The maximum
M other interfering channels. Because of the fluctuations ZERO level is obtained when the data channel is OFF (send-
in the amplifier gain, the output of the amplifier can be ing out power P o ) and all the interfering channels are also
one of M +1 possible levels with the probability distri- OFF. The minimum ONE level is obtained when the data
bution corresponding to the distribution of the gain. The channel is ON (sending out power P , ) and so are all the
possible signal levels after photodetection ( M = 10, r = interfering channels. Thus the condition for overlap is that
0.05, Pi, = -45 dBm) are shown in Fig. 5 along with P o G [ ( M + l ) P O ]L P I G [ ( M 1)P,] where G [- 1 is +
their associated probabilities. This figure is obtained by the function expressing the gain as a function of the input
multiplying the gain levels shown in Fig. 3 by the input power, the inverse of (1). Multiplying this inequality on
powers. There are M +
1 possible levels each for a ZERO both sides by M + 1 yields a comparison between the
and a ONE. Let the ONE levels be denoted as Lo,I through total output powers when all channels OFF and ON. Since
LM, where Lo,I is the maximum received ONE level (when the output power is a monotonically increasing function
1 9

all interferers are OFF) and LM, is the minimum received of the input power, this inequality is never true. Thus, the
ONE level (when all interferers are ON). Similarly, let Lo,o maximum ZERO level is always smaller than the minimum
through LM,odenote the received ZERO levels, with Lo,o ONE level irrespective of the number of channels, input
now being the minimum received ZERO level (when all in- power per channel, and the extinction ratio.
terferers are ON) and being the maximum received Using Fig. 5 , the system bit error rate (BER) can be
ONE level (when all interferers are OFF). evaluated by conditioning on the event that a particular
1886 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 8, NO. 12. DECEMBER 1990

signal level Li,oor Li,is received. Thus 1E-01


----- -----
M 1E-02

1E-04
\
1E-05
1E-06
1E-07 ‘\ \ \

where
- OE-09
, l E1 Bi \;. \ \
\

1E-11
1 E-11
nm . 1E-12
1E-13
1E-14 \3 ‘\ I

and oi,oand oi,I are the total noise standard deviations for
1E-15
1E-16
\
i ,\;, \
., I\[,
the ith received signal level for a ZERO and ONE,respec- -54 -52 -50 -48 -4 -44 -42 -40 -38 -36
Avrragr Input powor prr channel (dBrn)
tively. Here, T i s the decision threshold.
Fig. 6. BER versus received signal power. Curve 1: M = 0, r = 0. Curve
E. Setting the Decision Threshold 2: M = 1, r = 0. Curve 3: M = 10, r = 0. Curve 4: M = 100, r = 0.
Curve 5 : M = 500, r = 0. Curve 6: Thermal noise limited detection
Finding the optimal threshold is analytically intractable without the amplifier.
but several heuristics can be tried by setting the threshold
to iances) appears to be no higher than the case of binary
signals in additive Gaussian noise with equal noise vari-
T= LIDO + LoGI
(13) ances for the ZERO and the ONE.
01 + 00

and varying the choice of L, and Lo (and correspondingly F. Bit Error Rate Versus Number of Channels
also o1 and oo). Observe from Fig. 5 that the spread in Fig. 6 plots the BER versus the average received signal
received levels for a ZERO is much smaller than that for a power per channel. Also shown for reference is the curve
ONE.This implies that the threshold setting is relatively without the amplifier. With the amplifier, a 12.5-dB sys-
independent of the actual ZERO level chosen for reference. tem gain is obtained when no interfering channels are
A simple receiver simply sets the threshold at the average present (at lop9 BER). With one other channel, almost
received signal level, but this is not optimal in our case no penalty is paid. The penalty (measured relative to the
because of the spread in the received signal levels. The no interference case at lop9 BER) paid for 10, 100, and
bit error rate obtained by using the minimum ONE level 500 interfering channels is 0.4, 2.4, and 6 dB, respec-
and the maximum ZERO level ( L1 = L M , ,and Lo = LM,o ) tively. A significantly higher penalty is paid at higher sig-
is found to be very close to the optimal. Intuitively one nal powers because of the gain saturation characteristic of
can see that at high signal-to-noise ratios the bit error rate the amplifier. However, it is seen that using the amplifier
is essentially determined by these levels provided that provides a net system gain in all the cases.
these levels occur with a significant probability. Although
in a practical case, the robability of receiving these lev- G. Components of the System Penalty
els is very small ( 1/2 R ), the neighboring levels occur If one looks at the amplifier gain curve (Fig. 2) and the
with significant probabilities. Therefore, not much of an input power distribution, two components to the penalty
error is incurred by setting the threshold according to LM,I can be identified: a component attributed to the gain re-
and LM,o.At low signal-to-noise ratios, the choice of level duction due to the average input power level, and a com-
barely affects the BER, but at high signal-to-noise ratios, ponent due to the randomness in the input signal around
the BER is strongly influenced by the reference levels its mean value. We refer to the former component as the
chosen. This is because of the behavior of the Q function. saturation component and to the latter as the crosstalk
The threshold obtained by using LM, and LM.owas within component. The saturation component can be evaluated
10%of the optimal value in several test cases considered. by computing the error rate with all the interfering chan-
For all our future results, we will use this threshold set- nels sending out a constant average power. The crosstalk
ting. penalty is then the difference between the total penalty
The relative sensitivity of the bit error rate to the and the saturation penalty. The results are shown in Fig.
threshold can be defined as the ratio of the change in 7. The curve showing only the saturation component is
threshold to the difference between a ONE and a ZERO level. computed with the threshold set optimally between the
For the case of binary signals in Gaussian noise, it is easy only two possible received signal levels. The curve show-
to show that the sensitivity of the error rate to the thresh- ing the total penalty is computed with the threshold set as
old setting is minimized when the noise powers for a ONE described in Section 11-E. It is seen that when the number
and ZERO are identical. In the multichannel case, the sen- of channels is small, the saturation component is small
sitivity (measured with reference to the minimum ONE and compared to the crosstalk component (Fig. 7 for 100
maximum ZERO levels, which have comparable noise var- channels). Despite the 20-dB increase in the average input
__

RAMASWAMI AND HUMBLET: AMPLIFIER INDUCED CROSSTALK IN MULTICHANNEL OPTICAL NETWORKS 1887

4.00-
,
3.50-
,I..,

c5
3.00-
c
.
.- ,
m
52.50-
r ,
;
n
2.00-
/
I
c

lE-11-
1E-12-
1E-13-
1E-14-
1E-15-

-54 -52 -50 -48 -46 -44 -42 -40 -38 -36 .OOO .010 ,020 .030 .040 .050 .060 ,070 .OB0 .090 ,100
Avorage input powor por channol (dBrn) Extlnctlon ratio
Fig. 7. Saturation and crosstalk components of the system penalty. Curve Fig. 8. Power penalty versus extinction ratio (f,"= -48 dBm, M = 0 ) .
1: M = 0, r = 0. Curve 2: M = 100, r = 0. All interferers O N at average Curve 1: Thermal or spontaneous-spontaneous noise limited. Curve 2:
power. Curve 3: M = 100, r = 0. Curve 4: M = 500, r = 0. All inter- Signal-spontaneous noise limited. Curve 3: Calculated penalty.
ferers ON at average power. Curve 5 : M = 500, r = 0.

power with 100 channels, resulting in a reduction in the Fig. 8 plots the power penalty as a function of the ex-
gain by several decibels, only a 0.5-dB penalty is paid. tinction ratio, calculated using (14) and (15), as well as
The reason is that in this regime the detection is domi- the exact calculated power penalty. The amplifier input
nated almost entirely by amplifier noise, resulting in an power is Pi, = -48 dBm, which gives a BER of
error rate that is almost independent of the amplifier gain. with r = 0. In this range, the actual penalty curve is be-
For a larger number of channels the saturation component tween the curves predicted by (14) and (15), because, the
is significant (Fig. 7 for 500 channels) because the ther- spontaneous-spontaneous beat noise is not negligible
mal noise is no longer negligible compared to the ampli- compared to the signal-spontaneous beat noise.
fier noise. The effect of a nonzero extinction ratio r = 0.05 on
The important consequence of these results is that since crosstalk is shown in Fig. 9. There are two different ef-
the crosstalk reduction schemes to be proposed in Section fects to be considered.
IV can only eliminate the crosstalk component, they will 1. A detectability penalty is paid due to the decrease in
lose their effectiveness when the number of channels is the difference between a ZERO and ONE level. Thus the
large. curve showing the saturation component for the nonzero
extinction ratio will be to the right of the curve for the
H. Nonzero Extinction Ratios zero extinction ratio.
The effect of a nonzero extinction ratio r for single- 2. On the other hand, with r = 0.05, the signal varia-
channel amplification is to introduce a power penalty PP tion in the other channels is also reduced, which tends to
(defined in decibels as the degradation in the noise margin' decrease the bit error rate. Thus the crosstalk penalty for
for the same input power level), given by [23] the r = 0.05 case is smaller than the crosstalk penalty for

PP = 10 log (E) the r = 0 case.


In Fig. 9, with an optical bandwidth Bo = 10 GHz, the
first effect is dominant, and hence the r = 0.05 case per-
when the detection is thermal noise limited, and

PP = 10 log (E s)
when the detection is limited by the amplifier signal-spon-
(15)
forms poorer than the r = 0 case. However, if the optical
bandwidth is increased to 100 GHz, effectively letting in
more spontaneous-spontaneous beat noise, the penalty due
to the decrease in the difference between a ZERO and ONE
level is smaller, being closer to the value predicted by
(14) than (15). Fig. 10 plots the error rate versus input
taneous beat noise. Equation (14) also holds for any sig- signal power for this case. With 100 and 500 interfering
nal independent noise, including the spontaneous-spon- channels, the Y = 0.05 case crosses the r = 0 curve and
taneous beat noise (if we ignore the gain dependence due performs better at higher input powers. As the input power
to the small variation in input signal power while com- per channel increases, the crosstalk penalty begins to in-
puting the penalty). Thus, with r = 0.05, PP = 0.4 dB crease because the fluctuations in the gain become larger.
in the former case, and 2.4 dB in the latter case. After a certain point, the crosstalk penalty for the Y = 0
case becomes large enough compared to the r = 0.05 case
'The noise margin is the argument of the Q( . ) function that determines so as to offset the difference between the saturation pen-
the bit error rate and is given by ( L , - & , ) / ( U , + U,,), where L , , i = 0,
1 are the received signal levels for a ZERO and O N E , respectively, and U, are alties for the two cases (which is smaller than the Bo =
the corresponding noise variances. 10 GHz case). This effect is not seen in the 10 channel
1888 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 8, NO. 12. DECEMBER 1990

by considering the small signal transfer function between


the gain and the input intensity, which is low pass in na-
ture. The penalty in this case consists only of the satura-
tion component. The crosstalk component is negligible
because the gain variation with time is negligibly small.
We first present results for a 100-MHz receiver so that
1 E-09- we can compare the results with the results for the low
1E-10- data rate case in Section 11-F. With DFA’s, this band-
lE-11- width will fall into the high-data-rate regime, while with
1E-12- SLA’s, this will fall into the low-data-rate regime. Fig.
1E-13-
1E-14-
11 shows the bit-error-rate curves.
1E-15-
When a single channel is being amplified, a careful
1E-16 I l I . . . . . . comparison between Fig. 11 and Fig. 6 (the low-data-rate
-54 -52 -50 -48 -4 -44 -42 -40 -38 -36 case) shows that the high-data-rate case performs margin-
Average Input powor per channel ( d h )
ally better. The reason is that in the low-data-rate case, a
Fig. 9. Effect of extinction ratio on the BER Bo = 10 GHz. Curve 1: M signal ZERO sees a higher gain than a ONE. In the high-
= 0, r = 0. Curve 2: M = 0, r = 0.05. Curve 3 : M = 10, r = 0. Curve
4:M = 10, r = 0.05. Curve 5 : M = 100, r = 0. Curve 6 : M = 100, r data-rate case, the gain remains the same for both a ZERO
= 0.05. Curve 7 : M = 500, r = 0. Curve 8: M = 500, r = 0.05. and a ONE. The result is that the distance between the ZERO
and ONE levels is slightly larger for the high-data-rate case.
The penalty with 1 and 10 interfering channels is neg-
ligible. With 100 channels, the penalty is 0.7 dB and with
500 channels, it is 5.3 dB. These numbers are lower than
the corresponding penalties for the low-data-rate case
(Fig. 6).
We next present results for a 2.5-GHz receiver corre-
sponding to a data rate of approximately 5 Gb/s. Fig. 12
shows the bit-error-rate curves for a different number of
interfering channels. Firstly, observe that the net system
gain obtained on using an amplifier in the absence of any
interference is only 8.8 dB, as compared to 12.5 dB for
the 100-MHz receiver. This is because at the higher bit
rate, higher powers are required to obtain the same error
rate even in the thermal noise case. Thus on using an am-
-54 -52 -50 -48 -4 -44 -42 -40 -38 -36
Average input power per channel ( d h ) plifier, the effective gain seen is smaller because of satu-
Fig. 10. Effect of extinction ratio on the BER B , = 100 GHz. Curve 1: M ration. The penalties with 10, 100, and 500 channels are
= 0, r = 0. Curve 2 : M = 0, r = 0.05. Curve 3: M = 10, r = 0. Curve 0.6, 5 , and 17.2 dB, respectively, significantly larger
4:M = 10, r = 0.05. Curve 5 : M = 100, r = 0. Curve 6 : M = 100, r when compared to the 100-MHz case (Fig. 11) because
= 0.05. Curve 7: M = 500, r = 0. Curve 8: M = 500, r = 0.05.
of the higher degree of gain saturation.

case because the difference in crosstalk penalties between IV. CROSSTALK REDUCTIONSCHEMES
the two cases is small. This crossover is seen only with We study two schemes as a means of reducing the
small extinction ratios. crosstalk: ( I ) using frequency-shift keying and (2) elec-
The results above indicate that under certain circum- tronic compensation.
stances, a nonzero extinction ratio can actually help in
improving the system performance, depending on the op- A . Frequency-Shift Keying (FSK)
tical filter bandwidth. This has other beneficial conse- In FSK, the drive current of the laser is controlled so
quences, such as reduced laser chirp if the laser is directly that the laser emits one of two wavelengths, Xo and X I
modulated. depending on whether a ZERO or ONE bit is being sent.
Two different detection approaches are possible. One
111. HIGH-DATARATES approach uses a single narrow-band optical filter whose
We now consider the case when the input data rate is passband is centered at A, [24]. Thus, the ZERO wave-
much higher than the reciprocal of the lifetime. In this length is rejected and effectively, the incoming FSK sig-
case, the carriers or ions do not get sufficient time to re- nal is converted to OOK. We shall refer to this technique
spond to variations in the input signal and the carrier or as FSK/AM.
ion density in the active region remains almost constant. Another approach uses a wavelength demultiplexor to
The amplifier gain is also constant and its value can be spatially separate the two wavelengths Xo and A l . The two
obtained by using (1) with the input power set to its av- parts are then detected separately and the decision is made
erage value. The validity of this claim can be established depending on the difference between the detected signals
RAMASWAMl AND HUMBLET: AMPLIFIER INDUCED CROSSTALK I N MULTICHANNEL OPTICAL NETWORKS 1889

without interfering channels. For optical communication,


1E-04 the results are different from those of electrical commu-
1E-05 nication because in the former, the electrical signal am-
1E-06 plitude is proportional to the received optical power.
1E-07 We divide the total noise variance into a signal-inde-
1E-OB \
\ pendent component (spontaneous-spontaneous beat noise
5 1E-09 \
\ and thermal noise) that is the same in all directions in the
1E-10 \
\ signal space, and a signal-dependent component (signal-
1E-11 \
1E-12
\
\
spontaneous beat noise) whose variance is proportional to
1E-13
\
\
the signal power and which is present only in the same
1E-14
5 \
\
direction as the signal.
1E-154 \ \ In Appendix B, we derive the error-r+ateperformance of
. . . . . . . . . the three schemes. It is shown that FSK/AM is always 3
-54 -52 -50 -40 -46 -44 -42 -40 -30 -36 dB poorer than OOK. When the dominating noise com-
Average input power per channel (dBm)
ponent in the detection is signal independent, FSK re-
Fig. 11. BER versus received signal power for the case when the data rate quires 1.5 dB more power than OOK to achieve the same
is much larger than the reciprocal of the carrier lifetime ( B , = 100 MHz).
Curve 1: M = 0, r = 0. Curve 2: M = 1, r = 0. Curve 3: M = IO. r error rate. If the signal-spontaneous beat noise is the main
= 0 . Curve 4:M = 100, r = 0 . Curve 5 : M = 500, r = 0. noise component, then FSK requires 3 dB more power
than OOK in order to achieve the same error-rate perfor-
mance.
1E-01-
1E-02-
When both types of noise are present, one might expect
1E-03- that the penalty for FSK with respect to OOK should be
1E-04- between 1.5 and 3 dB. However, we show in Appendix
1E-05- B that the penalty decreases down to 1.25 dB when the
1E-06-
1E-07-
ratio of the signal-dependent noise variance to the signal-
6 1E-08- independent noise variance for FSK is two, and remains
1E-09- in that vicinity over a wide range of ratios.
1E-10-
In the multichannel case, the saturation penalties for
lE-11-
1E-12- FSK, FSK/AM, and OOK are the same since the average
1E-13- I II power is the same in all the cases. Thus, with signal-spon-
1E-14- \ I taneous noise limited detection, FSK is preferable to OOK
1E-15- I I only if the crosstalk penalty for OOK is more than 3 dB.
1E-16 I I I I I 1 I I 1 I
-60 -55 -50 -45 -40 -35 -30 -25 -20 -15 This is only an approximate comparison since the differ-
Average Input power per channel (dBrn) ent penalties are not strictly additive.
Fig. 12. BER versus received signal power for the case when the data rate Fig. 13 plots the calculated bit error rates for the dif-
is much larger than the reciprocal of the carrier lifetime ( B e = 2.5 GHz). ferent schemes, taking into account all the noise compo-
Curve 1: M = 0 , r = 0. Curve 2: M = 1, r = 0. Curve 3: M = 10, r
= 0. Curve 4: M = 100, r = 0. Curve 5 : M = 500, r = 0. Curve 6:
nents as well as gain saturation. For OOK, the error rates
Without amplifier. are calculated as outlined in Section 11-D. For the FSK
schemes, the calculation is simpler because only a single
level is received when either a ZERO or a ONE is sent. The
in the two arms. We shall refer to this system as FSK, error rates are obtained by computing the actual margins
since both the ONE and ZERO bit are detected. With these and then using the Q ( ) function. With no interfering
schemes, the crosstalk penalty can be eliminated because channels, one can see that FSK is poorer than OOK by
the total amplifier input power remains constant.* How- 1.25 dB (at BER). This corresponds to the case
ever, the saturation component is still present. where the penalty for FSK relative to OOK is at a mini-
Let us compare the error-rate performance of FSK and mum, and here the signal-spontaneous-beat-noise vari-
FSK/AM with that of OOK in terms of the average trans- ance for FSK is approximately twice the sum of the spon-
mitted signal power required to achieve the same error taneous-spontaneous and thermal-noise variances.
rates for all the schemes. The comparison is first done FSK/AM is poorer than OOK by 2.9 dB, slightly better
using noise margins assuming there are no interfering than the expected 3 dB. This is because the amplifier gain
channels and the same amplifier gain for all the schemes, for a ZERO bit is smaller in FSK/AM than in OOK and
and we then consider calculated error-rate curves with and hence the spontaneous-spontaneous beat noise variance is
smaller for FSK/AM.
*An additional source of degradation with the FSK schemes is crosstalk With 100 interfering channels, at low signal powers,
due to the nonzero facet reflectivities of the amplifier (in this paper, we OOK is better than FSK, while FSK performs better at
have ignored the facet reflectivities), which result in a frequency discrim- high signal powers (and low error rates). This is because
inator type of effect (FM to AM conversion) 191. This amplitude modula-
tion changes the refractive index of the medium, which in turn degrades the crosstalk penalty with OOK increases with the signal
the FSK signal. However, this effect has been shown to be small [9]. power. FSK/AM is poorer than both OOK and FSK.
1890 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 8. NO. 12. DECEMBER 1990

1E-02- value (when half the channels are ON and the others are
1E-03-
OFF). Then, the crosstalk penalty is eliminated, and only
1E-04-
1E-05- the saturation penalty is retained.
1E-06-
1E-07- V . WDM NETWORK DESIGN EXAMPLE
1E-OB-
1E-09- In this section, we explore the use of semiconductor
E 1E-10-
#lE-11- laser amplifiers in a WDM network based on a star to-
1E-12- pology [26] shown in Fig. 14. The network employs a
1E-13-
1E-14- fixed tuned laser with a unique wavelength for each trans-
1E-15- mitter, that is ON-OFF modulated and a tunable optical
1E-16-
1E-17- filter at each receiver. The passive star coupler combines
1E-18- the signals at its input ports and delivers an equal portion
1E-19-
1E-20-- of the total power to each one of its output ports.
In an N station network, N amplifiers are required. The
problem is to determine where to place these amplifiers.
They can be placed between the transmitter and the star
coupler at the hub, or between the coupler and the re-
ceiver, either before or after the tunable filter.
In practice it is not useful to place the amplifier close
Let us next compare the three modulation/detection to the transmitter because of the gain saturation effect. It
schemes assuming the same peak transmitted power per is always better to place the amplifier just before the star
bit in all the cases. In this case, the average transmitted coupler than just after it, because the total input power to
power for both FSK and FSK/AM is twice that of OOK, the amplifier is the same in both cases, but in the first case
resulting in a 3-dB improvement for both FSK and no crosstalk is introduced. Whether or not this is a good
FSK/AM relative to OOK. The performance of FSK/AM place to use the amplifier will depend on the link atten-
and OOK in the single channel case are identical. In the uation between the transmitter and the hub.
multichannel case, the average power for FSK and Consider the two cases when the amplifier is placed im-
FSK/AM is twice that of OOK, resulting in a larger sat- mediately before the filter, and just after the filter. Before
uration penalty for these schemes. Thus, in the signal- the filter, all the N channels are present resulting in a deg-
spontaneous noise limited case, FSK/AM is preferable radation in performance. However, the relatively narrow
to OOK if its excess saturation penalty over OOK is filter passband (compared to the spontaneous emission
smaller than the crosstalk penalty for OOK. bandwidth) limits the amount of noise that reaches the
The exact error-rate curves for FSK and FSK/AM are receiver. When the amplifier is placed after the filter, there
the same as in Fig. 13 because the average and peak pow- is no penalty due to the presence of other channels since
ers for FSK and FSK/AM are the same, but the curves only the desired channel is amplified. However, the noise
for OOK will be moved toward the right by 3 dB. For the over the entire spontaneous emission bandwidth now en-
100 channel case, both FSK and FSK/AM clearly out- ters the receiver. In addition, typically the filter has a loss.
perform OOK. This is because the excess saturation pen- The effect of this loss is to attenuate the noise when the
alty of FSK (or FSK/AM) over OOK is smaller than the amplifier is before the filter, while this does not occur
sum of the crosstalk penalty with OOK and the perfor- when the amplifier is after the filter.
mance advantage of FSK (or FSK/AM) over OOK. Of course, in the ideal situation when two filters are
These results indicate that using FSK or FSK/AM can available, the first filter selects the desired channel which
result in an improved performance in certain cases. An is then amplified, and the second filter placed after the
advantage of FSK over both OOK and FSK/AM, is that amplifier filters out the excess noise. However, this re-
since only a single level is received for a ZERO and a ONE quires two tunable filters, leading to increased cost along
with equal noise powers in both cases, the decision can with the complexity associated with their control. There
be made simply by taking the difference of the signals in is also the additional loss introduced by the second filter.
the two arms of the receiver matched to the ONE and the This option is not considered further.
ZERO signal and comparing it to zero. Let us consider a particular design point for such a net-
work [26]. Assume that the average transmitted power is
B. Electronic Compensation 0 dBm with an extinction ratio of 0.05, the network radius
In this scheme [ 2 5 ] ,a part of the input to the amplifier is 50 km, corresponding to a station to station link atten-
is tapped off and photodetected. The detected signal is uation of 20 dB. The splitting loss at the central star cou-
then used to control the bias current of the amplifier so as pler is 10 log N dB, where N is the number of stations.
to keep the amplifier gain constant. The effect of this to Therefore, the power per channel at the receiver just be-
maintain the gain constant and independent of the input fore the tunable filter is -20 - 10 log N dBm. Let the
signal level. Suppose we linearize the gain curve in Fig. required bit error rate be at an operating bit rate of
2 so that the gain is maintained constant at its median 200 Mb/s. The amplifier spontaneous emission band-
RAMASWAMI AND HUMBLET: AMPLIFIER INDUCED CROSSTALK IN MULTICHANNEL OPTICAL NETWORKS 1891

TRANSMITTER PORTION HUB RECEIVER PORTION


OF EACH T E R M I N A L OF EACH T E R M I N A L
1 , 2, ...N 1 , 2, ... N

TERMINAL FIXED-TUNED TUNABLE PHOTO- TERMINAL


TRANSMIT L A S E R S , WAVELENGTHS OPTICAL DETECTORS RECEIVE
ELECTRONICS L1, L2, ... LN F ILTERS ELECTRONICS

. 1 PASSIVE +I
Fig. 14. A wavelength division multi-access network using fixed wavelength transmitters and tunable filter receivers.

width is assumed to be 3 THz ( -25 nm). The objective


is to determine the maximum number of stations that can
be supported in this network. In particular, we study this
. .-J: .
as a function of two parameters, the bandwidth of the tun-
able filter, and its loss.
Suppose no amplifiers are used. Then using a baseline
thermal noise limited receiver sensitivity of -39 dBm,
and assuming the filter to be lossless, the link budget al-
lows a maximum splitting loss of 19 dB, implying that
about 79 stations can be supported. With a'filter loss of 3
dB, this number drops to 39, and with a loss of 6 dB, it
becomes 19. We will see that the use of amplifiers can + .. ... ...........1. . .. ....... . .. ... ........... ................................................................. .....
significantly improve these numbers. " I I

The signal power at an input of the star coupler is - 10 1E+01 1E+02 1 E+O3
Filter bandwidth ( G H t )
dBm. Placing the amplifier just before the star does not
help because of gain saturation. Using the crosstalk and Fig. 15. Number of stations versus filter bandwidth. Curve 1: 0 dB filter
loss without amplifier. Curve 2 : 0 dB loss with amplifier before filter.
noise results discussed in the earlier sections of this pa- Curve 3: 0 dB loss with amplifier after filter. Curve 4: 3 dB loss with
per, we have calculated the maximum number of stations amplifier before filter. Curve 5: 3 dB loss with amplifier after filter. Curve
that can be supported with the amplifier being placed just 6 : 6 dB loss with amplifier before filter. Curve 7: 6 dB loss with amplifier
after filter.
before or after the tunable filter at a receiver. Fig. 15 plots
this against the filter bandwidth for filter losses of 0, 3,
and 6 dB. Also shown for reference is the number of sta- Another limit is imposed by the bandwidth of the ampli-
tions that can be supported without amplifiers. It is seen fier itself ( 2 3 0 nm). The channel spacing is limited by
that in almost all cases, it is better to place the amplifier the filter passband and all the channels must fit within the
before the tunable filter and incur the penalty due to the amplifier gain bandwidth.
other channels, rather than place it after the filter and in- We have assumed in this case that all the channels are
cur the extra-noise penalty. For a reasonable filter band- at the same power at the amplifier input. In practice, the
width of 0.1 nm ( = 10 GHz), and loss of 3 dB, about 220 stations will be distributed at varying distances from the
stations can be supported. With a loss-free filter of the hub. Thus the desired channel may be received at low
same bandwidth, about 410 stations can be supported. power while an undesired channel may be received at high
Additional limits on the number of channels are im- power resulting in a degradation in performance.
posed by the tuning range of the filter, and the crosstalk
introduced by the filter because part of the adjacent chan- VI. SUMMARY AND CONCLUSIONS
nels are also let through [27]. In this case, the effect of In this paper, we have quantified the effect of crosstalk
filter crosstalk must be considered in conjunction with the on system performance. A simple worst-case design
amplifier crosstalk to evaluate the system bit error rate. guideline can be obtained by considering only the mini-
1892 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 8, NO. 12, DECEMBER 1990

mum received ONE level and the maximum received ZERO cross-sectional area. In a multichannel system, this trade-
level, though in practice these occur with low probability. off will have to be evaluated. For example, will an am-
Despite the penalty due to the presence of other channels plifier with Go = 30 dB, p,, = 0 dBm perform better or
and amplifier noise, the use of the amplifier provides a net worse than an amplifier with Go = 20 dB, P,,, = 10 dBm?
improvement in the power budget, as seen from a network Extension of these results to coherent multichannel sys-
design example. tems is straightforward. In this case, the dominant noise
The penalty due to the presence of other channels con- component is the beat noise between the amplifier spon-
sists of a saturation component arising from the steady- taneous emission noise and the local oscillator. Future
gain saturation because of a larger average input power work will attempt to use the results derived in this paper
and a crosstalk component because of the statistical vari- to further study the use of amplifiers in networks.
ation of the input powers around the mean. At low-data
rates, for typical signal powers, the saturation component APPENDIX A
is small when the number of channels is small, but i s sig- OF NOISESTATISTICS
DERIVATION
nificant when the number of channels is large. A{ high-
The noise introduced by the optical amplifier studied
data rates, the crosstalk penalty is no longer present and previously 111, [19], [22] but a simple derivation of the
only the saturation component is retained. We found that noise statistics at the receiver has not been found in the
frequency shift keying schemes (FSK, FSK/AM) elimi-
literature. Personick [ 2 2 ] derived the moment generating
nate the crosstalk component and are viable alternatives function for amplifier noise and also the moment gener-
to ON-OFF keying. ating function of the resulting shot-noise current after av-
We saw that at higher bit rates, using the amplifier pro-
alanche detection. We derive the noise statistics for a p-i-n
vides a lower overall improvement in the link budget. This
diode receiver in a simpler manner. The derivation below
is because higher powers are required to obtain the same
relies on the shot-noise statistics derived in [29] and the
error rates, resulting in more gain saturation in the am-
square-law detection results in [30] and [311.
plifier.
The electric field at the input to the receiver may be
We have determined the system penalty under two re- represented as
gimes of operation: when the data rate is either much
smaller or much larger than the reciprocal of the lifetime. E(t) = f i cos (27rLt + 9) + N ( t ) . (Al)
In the intermediate regime, the penalty will be between
Here, P is the signal power, f, is the optical carrier fre-
the saturation component and the sum of the saturation
quency, and 9 is the random carrier phase uniformly dis-
and the crosstalk components. A large signal amplifier
tributed in [0, 27rI. N ( t ) is the amplifier spontaneous
model proposed in [ 2 8 ] could be used in this regime to
emission noise assumed to be zero mean, Gaussian, with
evaluate exact penalties.
autocorrelation R N ( 7 ) = E [ N ( t ) N ( t + T)] and two-
In this paper, we have assumed that all the channels are
sided spectral density
bit synchronized. The asynchronous case was considered
in [lo], where an asynchronous channel that was ON at m

the beginning of the bit period and OFF at the end of the SN(f) = RN(T)e-j2+dT.
-m
bit period was assumed to contribute half its bit energy.
However, this is an approximation. If the amplifier gain The optical power incident on the detector is a random
responds extremely rapidly to changes in the input signal, process
then a more exact calculation can be performed by com-
puting the penalty for a given relative phase offset be-
tween the signals in the different channels and then aver-
~ ( t=) I E ( t ) 1’ +
~ ( 2 7 r ~ t (g)
= [ 2 cos2

aging over the distribution of the relative phase vector. In + 2 4% N ( t ) cos (27rf,t + @) + N 2 (t ) ] .
this case, the penalties for the asynchronous case can be
shown to be smaller than those for the synchronous case.
However, in a practical case, the gain may not be able to The mean power is
respond to variations in the input signal that occur at rates
much higher than the bit rate per channel. Our copjecture
is that the penalty will be still be smaller than the syn- Since N ( t ) is Gaussian, we have, by the moment formula
chronous case, but exact evaluation of this peaalty ap-
pears to be difficult.
+
[ 3 1 ] , E [ N 2 ( t ) N 2 ( t + 7 ) ] = R i ( 0 ) 2 R i ( 7 ) . Using
this, the autocovariance of P ( t ) defined as L p ( 7 ) =
At the device level, the penalty due to gain saturation E [ P ( t ) P ( t + 7 ) 1 - { E [ P ( t ) ] ) *can be computed to
can be decreased by increasing either the unsaturated gain b&
Goof the amplifier, or increasing its saturation power P,,,.
However, there is a trade-off between Go and P,,,. Giveq Lp(7) = 2Ri(7) + 4P&(7) COS (27rf,7)
that the net electrical pump current in an SLA or the op-
tical pump power in a DFA is held constant, one can in- P2
crease P,,, at the expense of reducing Goby increasing the
+-
2
cos (47rL7)

-1

I
RAMASWAMI AND HUMBLET: AMPLIFIER INDUCED CROSSTALK IN MULTICHANNEL OPTICAL NETWORKS 1893

and the corresponding spectral density

SP(f) = 2 M f ) c3 M f ) + 2P[S,(f-f,)
P2 P ( T )h ( t , - 7) d7
+ M f + f ) ]+ 4[ W -
2fc)

+ S ( f + 2f)l. (A5)
P ( T ) h(t2 - 7) d7. WO)
At this point, we shall consider the case of an ideal de-
Here, the 8 denotes convolution. After detection, the last tector, i.e., h ( t ) = 6 ( t ) . The following results may be
term in (A4) and (A5) above can be neglected because the easily extended to the general case, because an actual de-
2f, components will be filtered out. Since N ( t ) is pass- tector may be viewed as an ideal detector followed by a
band noise, centered at some frequency, say fN, the sec- filter with impulse response h ( t ) . For the ideal detector,
ond term above has a baseband component and also pass- the mean and the autocorrelation are, respectively
band components at +( f, +
f,). The passband
components will be filtered out.
The shot-noise current after detection can be expressed
as [29]
m

I(t) = e c
k = -m
h(t - fk) (Ab)

where e is the electronic charge, the { t k } are the photon


arrival times, and h ( t ) is the current pulse generated by
a single electron with j Y m h ( t )dt = 1 . The generation of
electrons may be described by an inhomogeneous Poisson
process with rate parameter Removing the conditioning over P ( ) yields -
X ( t ) = -P(t).
17 E[Z(t)] = re
hv
E[P(t)]
hv
Here 7 is the quantum efficiency ( 2:1 for a p-i-n diode). and
Our approach will be to determine the statistics of the shot ~ [ z ( t ,I () t 2 ) ]
noise current conditioned on a knowledge of the incident
power P ( t ) and then to remove the conditioning.
Following [29], we divide the time axis into small in-
w2
= - E [ P ( t , ) ]6 ( t 2 - t l )
hv
+ (;)2E[Ph) P(t2)I.
tervals 6t with the kth interval being [ (k - 1 / 2 ) 6 t , (k +
1 / 2 ) st]. Let the random variable Nk denote the number (A14)
of arrivals in the kth interval. The shot-noise current may Therefore, the autocovariance of I ( t ) is
then be expressed as L,(t,,t2) = E [ W W 2 ) ] - E[WI)l E[W2)1
m

Z(t) = lim e
6t-0 k=-
c m
N k h ( t - k6t). (A8) 17eL
= -E [ P ( t , ) ] w
hv 2 - 4)
Since the intervals are nonoverlapping, the Nk are inde-
pendent Poisson random variables with rate parameter (A15 1
X ( k 6 t ) 6t. The mean and autocorrelation of I ( t ) can now
be determined under the limit 6t -, 0 as For the case at hand, with 7 = t2 - t l , substituting for
E [ P ( t ) ]and Lp(7)from (A3) and (A4), respectively, we

= e2 1-m
~ ( 7 h) ( t l - 7) h(t2 - 7) d7 The first term above represents the shot-noise terms due
to the optical signal and noise. The second term is the
+ E[Z(rl lP(.))] E [ ' ( ' 2 l P ( - ) ) ] signal-spontaneous beat noise, and the last term above is
1894 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 8, NO. 12, DECEMBER 1990

t
power spectral density
the spontaneous-spontaneous beat noise. The spectral
density of the photocurrent, which is the Fourier trans-
form of L,( 7 )is then
/ \ 2 : l i s 3

+ SN(f + fc)] + ($2,2SN(f) 8SN(f)I.

(A17) -O.So -8, B, 0.58, BO

The results so far are valid for Gaussian noise N ( t ) with frequency

an arbitrary spectral shape SN( f ). As a special case, let Fig. 16. Noise power spectral densities at the receiver. Curve 1 : Shot noise,
us consider the amplifier noise to be centered at fc, and S, = lhvl.' [ P + N O B , ] . Curve 2: Signal-spontaneous beat noise. S, =
white over an optical bandwidth of Bo < 2fc, with
[f = fN,(G - l)k?outa, If*$ Bo
5 -
(E)* 2 P N 0 . Curve 3: Spontaneous-spontaneous beat noise, S, =

2
SN(f) =
otherwise.
('418) K
The signal power can be related to the amplifier input
power as P = GPinqinqouta. For this case, S I ( f ) is shown
plotted in Fig. 1 6 . The shot-noise component is white,
the signal-spontaneous component is rectangular, and the
spontaneous-spontaneous noise has a triangular shape.
The photocurrent Z ( t ) is passed through a low-pass filter 0 2par
L pm 0 pm=
with bandwidth B e . In order to obtain the powers of the OOK FSK =/U
different noise components we integrate their power spec-
Fig. 17. Signal-space representation for OOK, FSK, and FSK/AM. POOK,
tral densities between -Be and Be resulting in PFSK, and P,,,,,, represent average signal powers for OOK, FSK, and
FSK/AM, respectively.

Nshot =
w2 2[p +
of interest. The signal-dependent noise variance for a ONE
l?e bit is proportional to the signal power and is given by
= 2 hv a[GPinqinqout + Nsp(G - l)hvBo~lout]Be
k,2Poo~for OOK and k,PFsK for FSK (and FSK/AM)
~

where k, is a constant whose value is obtained from (8).


(A19) The total noise variance for a ONE bit for OOK is ( T : , ~ ~ ~

where we have substituted for P and No. Similarly


\ 2
= af +
ks2~ooK and the variance for a ZERO bit is
/
U: = a :, since there is no signal-spontaneous beat noise
for a ZERO.Here U: is the sum of the variances of the
thermal and spontaneous-spontaneous beat noises. The
q2e2 optimal threshold setting for OOK is approximately
= 4 -l ? i n q : u t Q ! 2 ~ i n ~ (G - ~ ) N , B , ( '420 )
hv T = ~ ~ P o oOOK
K~I,
and 031)
OO + uI,OOK

Thus

= q2e2q:uta2[Nsp(G - 1)]*(2BoB,- B:). (A21)


BEROOK= ( k2PooK
+ uI,OOK
). (B2)

APPENDIXB Similarly
ERROR-RATE PERFORMANCE FOR OOK, FSK, AND
FSK/AM
The signal-space representation of the three schemes is
shown in Fig. 17, assuming a zero extinction ratio for where u:,FSK/AM = a? + ksPFSK/AM. Therefore,
OOK. P O O K , PFSK, and PFSK/AM represent average signal FSK/AM is always 3 dB poorer than OOK. The error-
powers for OOK, FSK, and FSK/AM, respectively. From rate calculation for FSK is more complicated because the
(lo), the electrical signal amplitude for a ONE bit in OOK noise variances in different directions of the signal space
is k2PooKwhere k is a constant. For OOK and FSK/AM, are different. The signal-independent component is the
only the noise components along the horizontal axis are same in all directions, but the signal-spontaneous beat
RAMASWAMI AND HUMBLET: AMPLIFIER INDUCED CROSSTALK IN MULTICHANNEL OPTICAL NETWORKS 1895

xo
0.7501/\
0.700

0.650

4T?
X
,
,, 0.600
, “
, x1 , x1
(kP. 0) 0.550

0.500 I I I I I 1 I I I l

Fig. 18. Signal-space transformation for FSK to compute error rates. P is 5 10 15 20 25 30 35 40 45 50


bda
the average signal power. (a) Signal-space with different noise variances
in different directions. The decision boundary is the x,, = x I line. (b) Fig. 19. Power penalty for FSK relative to OOK (absolute units) x plotted
Transformed signal-space with unit noise variance in all directions. The against the ratio of signal-spontaneous to signal-independent noise var-
new decision boundary is the line xo. = uoxI / U , .
I iances for FSK, 0.

noise is only in the direction of the signal (Fig. 18(a)). angle with respect to the decision boundary; hence, FSK
For a ONE bit, the noise variance in the Xo direction is is less affected than OOK. At the same time, however,
U ; = U’, and the variance in the XIdirection is u:,FSK =
the boundary for OOK shifts toward the ZERO to compen-
U,’ + k,PFsK. The optimal decision rule is as shown in the
sate for the increased signal-dependent noise, while the
boundary for FSK remains unchanged. Ultimately this ef-
figure. In order to simplify the computation of error rates,
we make a linear transformation of the signal-space from fect compensates for the former effect and the relative per-
(XI, Xo) to (Xi,Xb), where Xi = Xl/ul,FSKand Xb = formance of FSK with respect to OOK becomes poorer.
Xo/uo, so that in the new coordinate system, the noise ACKNOWLEDGMENT
variance is equal to unity in all directions. The new de-
cision boundary is xl, = uox;/ ( T I , FSK (see Fig. 18(b)). The The authors would like to thank P. Green for his en-
minimum distance from this line to the signal point is couragement and support, and for his comments on this
k P F s K / J U i 4- U:,FSK. Thus
paper. They would also like to thank W. Hamdy and all
members of the optical networking group at IBM Re-
search for many insightful discussions.
REFERENCES
[ l ] N . A. Olsson, “Lightwave systems with optical amplifiers,” J .
(B4) Lightwave Technol., vol. 7 , no. 7, pp. 1071-1082, July 1989.
Let us now compare the performance of FSK with OOK. [2] M. J . O’Mahony, “Semiconductor laser optical amplifiers for future
fiber systems,” J . Lightwave Technol., vol. 6 , no. 4 , pp. 531-544,
In order to obtain the same error rates, from (B4) and (B2), Apr. 1988.
we want [3] T. Saitoh and T. Mukai, “Recent progress in semiconductor laser
amplifiers,” J . Lightwave Technol., vol. 6, no. 11, pp. 1656-1664,
- Nov. 1988.
PFSK - 2pOOK
. 035) [4]J. C. Simon, “GaInAsP semiconductor laser amplifiers for single-
J~u+
’ , k,PFsK 0, 4- do’,4- k,2Poo~ mode fiber communications,” J . Lightwave Technol., vol. LT-5, no.
9 , pp. 1286-1295, 1987.
With p = kSPFsK/u and ’, PooK= XPFSK, solving the above [5] P. Urquhart, “Review of rare-earth doped fiber lasers and ampli-
fiers,” IEEProc. PI. J . , vol. 135, no. 6 , pp. 385-407, 1988.
yields [6] T. E. Darcie, R . M. Jopson, and R . W. Tkach, “Intermodulation

x =
p + 2Jp + 2 distortion in optical amplifiers from camer-density modulation,”
Electron. L e f t . , vol. 23, no. 25, pp. 1392-1394, Dec. 1987.
2(P + 2) . [7] T. G. Hodgkinson and R. P. Webb, “Application of communication
theory to analyze c a m e r density modulation effects in traveling-wave
semiconductor laser amplifiers,” Electron. Lett., vol. 24, no. 25, pp.
This relationship is plotted in Fig. 19. Observe that 1550-1552, Dec. 1988.
when = 0 (no signal-dependent noise), x = 1 /&, i.e., [8] M. G . Oberg and N. A. Olsson, “Crosstalk between intensity mod-
FSK is 1.5 dB poorer than OOK. When /3 -, 00 (only ulated wavelength division multiplexed signals in a semiconductor
laser amplifier,” IEEEJ. Quantum Electron., vol. 24, no. 1, pp. 52-
signal-dependent noise), x converges slowly to 1/2, i.e., 59, Jan. 1988.
asymptotically, FSK is 3 dB poorer than OOK. The min- 191 H. E. Lassen, P. B. Hansen, and K. Stubkjaer, “Crosstalk in 1.5 pm
imum difference in performance occurs when x is maxi- optical amplifiers,” J . Lightwave Technol., vol. 6 , no. I , pp. 1559-
1565, Oct. 1988.
mized with respect to p. This occurs at p = 2, which (IO] K. Inoue, “Crosstalk and its power penalty in multi-channel trans-
yields x = 0.75, or a 1.25 dB difference in performance. mission due to gain saturation in a semiconductor laser amplifier,” J .
The reason why the penalty decreases at first when /3 Lightwave Technol., vol. 7 , no. 7 , pp. 11 18-1 124, 1989.
[ I l l G . Grosskopf, R. Ludwig, and H. G . Weber, “Crosstalk in optical
increases is that the signal-dependent noise is in the di- amplifiers for two-channel transmission,” Electron. Lett., vol. 22,
rection of the signal, which in the case of FSK is at a 45” no. 17, pp. 900-902, Aug. 1986.
1896 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 8, NO. 12, DECEMBER 1990

[I21 G . Eisenstein, K. L. Hall, R. M. Jopson, G. Raybon, and M. S . [27] P. A. Humblet and W. M. Hamdy, “Crosstalk analysis and filter op-
Whalen, “Two-color gain saturation in an InGaAsP near-traveling timization of single- and double-cavity Fabry Perot filters,” IEEE J .
wave optical amplifier,” in Proc. OFCIIOOC, 1987. Select. Areas Cornmun., Aug. 1990.
[I31 T. Mukai, Y. Yamamoto, and T . Kimura, “S/N and error rate per- [28] A. A. Saleh, “Nonlinear models of traveling-wave optical ampli-
formance in AlGaAs semiconductor laser preamplifier and linear re- fiers,” Electron. Lett., vol. 24, no. 14, pp. 835-837, July 1988.
peater systems,” IEEE Trans. Microwave Theory Technol. , vol. [29] E. A. Lee and D. G. Messerschmitt, Digital Communicarion. Klu-
MTT-30, no. 10, pp. 1548-1554, 1982. wer Academic, 1988, chap. 3 .
1141 T. Mukai, K . Inoue, and T . Saitoh, “Signal gain saturation in two [30] W. B. Davenport and W. L. Root, An Introduction to the Theory of
channel common amplification using a 1.5 pm InGaAsP traveling Random Signals and Noise. New York: McGraw-Hill, 1958, chap.
wave laser amplifier,” Electron. Letr., vol. 23, pp. 396-397, Apr. 13.
1987. [3 11 A. Papoulis, Probability, Random Variables,and Srochasric Proceses,
[15] C. Coquin er a l . , “Simultaneous amplification of 20-channels cen- 2nd ed. New York: McGraw-Hill, 1984, chap. 9 .
tered at 1.54 pm in a multiwavelength distributed system,” ECOC’88,
1988, postdeadline paper.
[I61 M. J. Pettitt, A. Hadjifotiou, and R. A. Baker, “Crosstalk in Erbium
*
doped fiber amplifiers,” Electron. Lett., vol. 25, no. 6, pp.
.. 416-417,
Mar. 1989. . Rajiv Ramaswami, received the B.Tech. degree
R. I . Laming, L. Reekie, P. R. Morkel, and D. N. Payne, “Multi- in electrical engineering (electronics) from the In-
channel crosstalk and pump noise characterization of Er doped fiber
dian Institute of Technology, Madras, in 1986,
amplifier pumped at 980 nm,” Electron. Lett., vol. 25, no. 7, pp.
and the M.S. and Ph.D. degrees in electrical en-
455-456, Mar. 1989. gineering and computer science from the Univer-
E. Desurvire, C. R. Giles, and J. R. Simpson, “Gain saturation ef-
sity of California, Berkeley in 1988 and 1990 re-
fects in high-speed multichannel Erbium-doped fiber amplifiers at X spectively. He spent 3 mo. in 1988 as a visiting
= 1.53 pm,” J . Lightwave Technol., vol. 7 , no. 12, pp. 2095-2104,
student at the IBM Zurich Research Laboratory.
Dec. 1989.
He joined the optical networking group at the
Y. Yamamoto, “Noise and error-rate performance of semiconductor IBM T.J. Watson Research Center, Yorktown
laser amplifiers in PCM-IM transmission systems,’’ IEEE J . Quantum
Electron., vol. 16, pp. 1073-1081, 1980. .,
Heights. NY in 1989, where he is currently a re-
search staff member. His research has focused on applying communication
M. J. Adams, J. V. Collins, and I. D. Henning, “Analysis of semi-
theory to problems in optical communication and in developing protocols
conductor laser optical amplifiers,’’ IEE Proceedings Part J . , vol.
and algorithms for networks.
132, no. 1, pp. 58-63, Feb. 1985.
Dr. Ramaswami was a recipient of the Smith and Regents Fellowships
P. S . Henry, “Error rate performance of optical amplifiers,” in Tech.
at Berkeley and the National Talent Search Scholarship in India.
Dip. OFC’89. Feb. 1989. U . 170.
[22] S . D. Personick, “Applications for qudntum amplifiers i n simple dig-
ital optical communication systems,” Bell Sjsr Tech J , vol 52, no *
1, pp 117-133, Jan 1973
(231 A J. McDonald, R S Fyath, and J J O’Reilly, “Influence of ex-
tinction ratio on performance of optical receivers incorporating laser Pierre A. Humblet (S’72-M’77), received the
preamplifiers,” Electron Lett , vol 25, no 4 , pp 249-250, 1989 degrees of Electrical Engineer from the Univer-
1241 I P Kaminow, P P Ianonne, J Stone, and L W Stulz, “FDMA- sity of Louvain (Belgium), and the MSEE and
FSK star network with a tunble optical filter demultiplexer,” J Lighr- Ph.D. degrees from the Massachusetts Institute of
wave Technol , vol 6, no 9 , pp 1406-1414, Sept 1988 Technology, Cambridge, MA.
(251 T E. Darcie, R M Jopson, dnd A A Saleh, “Electronic compen- After graduating in 1978 he has remained at
sation of saturation-induced crosstalk in optical amplifiers,” Elec- MIT where he is now an Associate Professor of
tron Lett , vol 24, no 18, pp 1154-1 155, Sept 1988 Electrical Engineering. His teaching and research
[26] N R Dono, P E Green J r , K Liu. R Rama\wami, and F F interests are in the area of communication sys-
Tong, “A wavelength divicion multiple access network for computer tems, particularly wideband optical networks and
communication,” IEEE J Selecr Arras Cornmun , vol 8 , no 6, pp distributed algorithms. He is a consultant with a
983-994, Aug 1990 number of companies, most recently IBM and Codex Corporation.

You might also like