power up to 55mW.

Beyond 55mW, no further gain increase minimal trellis structure of a block code can be used to
nor overshoot can be detected. At this latter operating point, perform low-complexity maximum likelihood soft decision
when the input signal is modulated, the ASE spectrum decoding. On the theoretical side, these trellises can be used
remains unchanged and only the laser output is modulated. for constructing other codes [3]. Recently [4], generalised
More detailed analysis is required if part of the modulation array codes (GACs) and their trellis structure have been intro-
spectrum lies around the relaxation oscillation frequency of duced. The technique allows the design of array codes with
the laser (here around 20 kHz). In addition, if the pump power the same code length, n, and d,,, = 4, but with an increased
is increased, only the laser output power is affected. number of information digits.
The pump powers required for controlled gain operation In this Letter we propose a new encoding technique for
are quite large due to the unoptimised system. An improve- Hamming codes, based on generalised array codes (GACs)
ment in the launched efficiency, fibre length and the pump and their trellis structure. The trellis structure of such codes
wavelength, can lead to a strong reduction of the pump provides lower complexity soft maximum likelihood decoding
powers required. (SMLD) compared with conventional techniques.

Conclusion: We have written two Bragg reflectors at 1480nm Encoding procedure: A generalised (n, k , d,,,) array code is an
at both fibre ends of an EDFA. When lasing at 1480nm, the array code in which the column and row subcodes may have
gain is clamped. This is a simple implementation of a passive different numbers of information and parity check symbols,
all-optical gain control which induces no intrinsic extra loss the code length n = nin2, and the total number of information
for the signal. The pump power required for controlled oper-
ation is somewhat larger than that predicted by theory. We
digits k = k, k ,+ + + .. . k,,, where n, and n2 are the
number of rows and columns, respectively, and k, is the
believe this is due to inhomogeneous broadening. number of information digits in the pth row [4]. The pro-
cedure for designing a Hamming code, using GACs, is as
Acknowledgments: It is a pleasure to thank J. C. Simon (from follows, and is illustrated for the (7, 4, 3) code:
CNET Lannion LAB/OCM/FOG) for fruitful discussions.
We thank E. Taumieb and P. Simonpietri (from Photonetics (i) Design the basic (8, 3, 4) array code, C,, and an additional
S A . ) for technical assistance. code, C , , as follows:

IEE 1993 4th May 1993

E. Delevaque, T. Georges, J. F. Bayon and M. Monerie (France
Telecom-Centre National d’Etudes des TQlicornmunications, LAB/
OCMIFOG-BP40, 22301 Lannion Cedex, France)
P. Niay and P. Bernage (Laboratoire de Dynamique Moleculaire et c5 c4 0x4
Phofonique, Unit; de recherche Associee au C N R S no. 779, Uniuersite
des Sciences et Technologies de Lille, U F R de Physique, Bit. P5,59655 where x i , i = 1, , .., 4, represent information digits; c j , j = I,
Villeneuue d’Ascq Cedex, France) . . . , 5, represent parity check symbols.

References (ii) Add the two codes, C, and C,, and delete the parity-check
symbol, which is located in the n , th row and n,th column, as
ZIRNGIBL, M.:‘Gain control in erbium doped fibre amplifier by an follows :
all-optical feedback loop’, Electron. Lett., 1991, 27, pp. 560-561
GEORGES, T., and DELEVAQUE, E.: ‘Analytic modelling of high-gain
erbium doped fiber amplifiers’, Opt. Lett., 1992, 17, pp. 1113-1115
MELTZ, G., MOREY, w. w., and GLENN, w. H . : ‘Formation of Bragg
gratings in optical fibers by a transverse holographic method‘, Opt.
Left., 1989, 14, pp. 823-825
LEGOUBIN, S., FERTEIN, E., W U A Y , M., BERNAGE, P., NIAY, P., BAYON, J. c5
F., and GEORGES, T.: ‘Formation of moire gratings in core of ger-
manosilicate fibre by transverse holographic double exposure = (XI,~4 0 cis ~ 2 x4% 0 ~ 3 x4
, 0 C J
method’, Electron. Lett., 1991, 27, pp. 1945-1947
GILES, c. R., DESURVIRE, E., and SIMPSON, J. R.: ‘Transient gain and
(2)
cross talk in erbium doped fiber amplifiers’, Opt. Lett., 1989, 14,
pp. 880-882 The designed code is a linear nonsystematic code with the
parameters: n = 7, k = 4, d m i , = 3.

Example I : A binary vector of information digits is given as

follows: X = (xi, x2. x3. x4) = (1101). Following the pro-
cedure outlined above, C,, C , and C codewords can be
written as follows:
LOW-COMPLEXITY TRELLIS DECODING OF
H A M M I N G CODES 1 1 0 1 1 0
B. H o n a r y a n d G. M a r k a r i a n 1 1
c - C - O C= o=(loloolo)
‘-0 0 ,-o 1 0 1
Indexinq terms: Codes and coding, Information theory 0 0 0 1 0

A new encodingJdecoding technique for Hamming codes, The code table of the designed (7, 4, 3) code is given in Table
based on generalised array codes (GACs) is proposed. The 1, and the weight distribution function for the designed code is
proposed technique allows the design of Hamming codes shown in Fig. 1. This is similar to that given in Reference 5 for
with minimal trellises. An example is given for the (7, 4, 3) the (7, 4, 3) Hamming code. Following Reference 6, the
Hamming code, hut the proposed technique is applicable to
all existing Hamming codes. The trellis structure of such designed code is a Hamming code.
codes provides low complexity soft maximum likelihood
decoding. Trellis design f o r Hamming c o d e : The trellis design procedure
for GACs [4] can be used to design the soft maximum likeli-
hood decoder of Hamming code. The procedure is as follows:
Introduction: The trellis decoding technique for linear block
codes was first introduced by Wolf [l], Recent publications (i) Choose the trellis depth (number of columns) N , , and
by Forney [2] and Muder [3] have increased the interest paid number of states N , , as
to the low-complexity trellis decoding of block codes for both
practical and theoretical reasons. On the practical side, a N , = n, +1 N , = 2-Ik. p = 1, 2, .. ., n2 (4)

1114 E L E C T R O N I C S LETTERS 10th June 1993 Vol. 29 N o . 12

(ii) Identify each state of depth p by the max{k,}-tuple binary tively. The trellis branches are labelled with X,/C, and are
vector S,(4 = S,(a,, a2, ..., obtained according to eqn. 5. Following the above procedure,

Table 1 CODE TABLE O F DESIGNED (7,4,3) CODE

Information vector Encoded codeword Information vector Encoded codeword
m 0000000 lo00 1 1 m 1
OOO1 0101010 1001 1001011
0010 oooo111 1010 1100110
0011 0101101 1011 1001100
0100 0011001 1100 1111OOO
0101 01 10011 1101 1010010
0110 0011110 1110 111I111
0111 0110100 1111 1010101

7-

2
0
6-

5 5-

:
0
4-

k 3-
E, 2-

1-

weight
197(12/
. 1 w .
1911111 I( 7 , 4 , 3 )GAC
Fig. 2 Trellis diagram o f ( 7 , 4 , 3 )Hamming code [ I ]
Fig. 1 Weight distribution o f ( 7 , 4 , 3 )GAC
(iii) The trellis branches start and finish at depth p = 0 and the trellis diagram for the (7.4, 3) Hamming code is presented
p = N , - 1, which are labelled as So(OO, . .., 0) and S,,(00, _ . _ , in Fig. 3, and Table 2 compares the complexity of the two
O), respectively. trellis diagrams. It is clear from Table 2 that the designated
Hamming code has a low-complexity trellis diagram, which is
(iv) The trellis branches at depth P are labelled X&,, where similar to that given by Forney [2] for the codes with minimal
x,and C, are k,- and n,-twle binary vectors of information trellises. The proposed technique can be implemented for the
digits and encoded codewords for the pth row code, and trellis design of all known Hamming codes: for example the
(15, 11, 3) Hamming code can be derived from the (16, 11, 4)
C, = Cj . G , + C: (5) GAC, introduced in Reference 4, etc.
where G , is a generator matrix for the pth row code, and Cb
and C i are codewords from the pth row of C , and C, codes, Conclusion: A new encoding/decoding technique for
respectively. Hamming codes, based on generalised array codes and their
trellis structure, is proposed. The technique allows the design
(v) There are 2’, branches starting from each state, S,(A), at of Hamming codes with minimal trellises. An example is given
depth p , and each branch is connected with state S , , ,(A) at for the simplest (7, 4, 3) Hamming code, however the tech-
depth p +
1, which can be defined as follows: nique can be implemented for the design of all known
Hamming codes.
S,+,(A,) = S,(A,) + c, (6)
oo/oo o/oo o/oo
There are N - 2kl+k2+...+k n 2 distinct paths through this
0-
trellis diagram and each path corresponds to a unique code-
word from the Hamming code.

Example 2. The aim is to design the trellis diagram of the (7,4,

3) Hamming code. Using the technique proposed by Wolf [l],
the trellis diagram for the (7,4,3) Hamming code has N , = 8,
N , = 8, and is shown in Fig. 2. Following the technique out-
lined above, the new trellis diagram of the (7,4,3) code will
have N , = 4 + 1 = 5, N , = 2’ = 4, and the code has the fol-
lowing generator matrices: G, = [l 11, i = 1, . .. , 4 .
We identify each state of depth p (p = 1, 2, 3, 4) by 2-tuple
binary vector S,(a,, a2), and at depth p = 0 and p = 4, the O/Ol O/Ol
trellis has only one state, namely So(OO)and S,(00), respec- Fig. 3 Trellis diagram of proposed ( 7 , 4 , 3 )Hamming code

Table 2 COMPLEXITY O F TWO TRELLIS DIAGRAMS

Code States Columns Branches Additions Comparison
Conventional [l] 8 8 36 34 IO
Proposed 4 5 24 40 9

ELECTRONICS L E T E R S 10th June 1993 Vol. 29 No. 12 1115

Acknowledgments: The authors wish to acknowledge the gate width in the n-structure is easier to fabricate than the
financial support given by the Royal Society, UK to G. Mar- long gate width in the T-structure because, as the gate length
karian for his visit to Lancaster University.

0IEE 1993 6th M a y 1993

B. Honary and G . Markarian (Hull-Lancaster Communication
Research Group, Engineering Department, Lancaster University,
Lancaster LA1 4YR, United Kingdom)

References
1 WOLF, J. K.: ‘Eficient maximum likelihood decoding of linear
block codes using a trellis’ IEEE Trans., 1978, IT-24, (1). pp.
76-80
2 FORNEY, G. D.: ‘Coset codes-Part 2: Binary lattices and related
codes’, IEEE Trans., 1988, IT-34, (5), pp. 1152-1187
3 MUOER, D.I.: ’Minimal trellises for block codes’, IEEE Trans. 1988,
IT-34, (3,pp. 1049-1053
4 HONARY, B., MARKAMAN, G., and FARRELL, P.: ‘Generalised array
codes and their trellis structure’, Electron. Lett., 1993, 29, (6), pp.
541-542
5 LIN, s., and COSTELM, D.: ‘Error control coding: fundamentals and
applications’ (Prentice-Hall, Englewood Cliffs, NJ, 1983)
6 PHELPS, K. T.: ‘A combinatorial construction 01 perfect codes’,
SIAM J . Algebraic Discrete Mathematics, 1983, (4), pp. 398-403

EFFECT OF GATE STRUCTURES O N NOISE

CHARACTERISTICS OF 0.15pm
AI GaAs/l nGa As PS EU DO M0R PH IC H E M T s
K. C. Hwang, P. Ho and F. R. Bardsley

Indexing terms Transistors, Semiconductor devices and

materials
b 1897111
I

The Letter investigates the effect 01 two different gate layouts Fig. 1 Photograph of 0.15 x 1 5 0 w 2 A1GaAsllnGaAslGaAs pseudo-
on the noise characteristics of 150pm2 AIGaAsl
InCaAs pseudomorphic HEMTs; the T-structure with a gate morphic
width of 2 x 75flm, multigate finger n-structure with a gate a Multigate finger II-structure (6 x 25pm gate width)
width of 6 x 25um and with airbridees. Whereas the DC b T-structure(2 x 75 pm gate width)
characteristics 01 both types are verysimilar, the RF noise
Characteristics of the multifinger gate n-structure devices is reduced to 0.15pm, the lithography for longer gate width is
exhibit a minimum noise figure 01 0.67dB with an associated more diffcult and reduces device yield, especially for mush-
gain of 12.2dB at IXGHz, which is superior to that of T- room gate HEMTs. Another advantage of this n-structure is
structure devices. that the airbridge is not covering the gate channel area.
The pseudomorphic HEMT structure was grown by molec-
ular beam epitaxy (MBE) on a semi-insulating GaAs sub-
High electron mobility transistors (HEMTs) have recently strate. It consists of a 350A GaAs cap layer doped at
stimulated a great deal of interest for high performance micro- approximately 4 x 1018/cm3, a 400A AI,,,,Ga,,,,As gate
wave and millimetre-wave low noise applications due to the layer doped at 2 x 10’7/cm3, a silicon planar doped layer at
extremely low noise performance and high associated gain [l, 5 x 101z/cm2,a 45A undoped Alo.z3Gao.77Asspacer layer, a
21. Since HEMTs were first demonstrated, considerable l 0 0 A undoped In,,,2Ga,,8As active layer, and a 1 pm thick
progress has been made in GaAs-based HEMTs by intro- undoped GaAs buffer layer. In the device fabrication, the
ducing an InGaAs strained channel which improves electron mesa area was defined using a wet chemical etch for isolation.
confinement and mobility, and the AlGaAs layer DX-centre The AuGe/Ni/Au ohmic contact was alloyed with rapid
effect [ 3 , 4 ] . thermal annealing. The mushroom Schottky gate was defined
We report a significant improvement of the noise per- by direct write electron-beam lithography.
formance of the 0.15 x 150pm2 pseudomorphic HEMT by
using a multifinger gate structure. In a popular T-shaped gate
The DC parameters of both the n-and T-structure
devices were obtained using an H P 4145B parameter analyser.
structure as shown in Fig. la, the total gate width is divided The DC characteristics of both types of HEMT are very
into two gate strips (2 x 75 pm) connected in parallel. On the similar. Fig. 2 shows the typical g, and drain-source current
other hand, a multifinger n-shaped gate structure, as shown (I,) against gate-source voltage (VJ at a drain bias (cr)
of
in Fig. lb, has the same total gate width, but it is divided into 2.5 V for both IL and T-structure devices. Very good pinchoff
six equivalent parallel-connected strips, where each gate strip characteristics are observed on both structures. Mesa isolation
is 25pm in length with two source pads interconnected by in the device fabrication results in sidewalls where the low
airbridges. The gate resistance is a key factor affecting the R F bandgap InGaS channel is exposed and comes into contact
noise performance of the low-noise HEMT according to the with the gate metallisation running up the mesa. The low
following expression [SI: F,,(dB) = 10 log (1 + kfC [ ( R , Schottky barrier height on InGaAs results in a sidewall
+ R s ) / g , ] ” z ) where F,,, is the minimum noise figure, k 3; the leakage path from the gate to the channel [ S I , Although the
Fukui factor, f is the frequency, C,, is the input gate capac- n-structure has more crossover contact points than the T-
itance, R , is the gate resistance, R , is the source resistance, and structure, the gate leakage currents and the breakdown volt-
g, is the transconductance. When the R , , C,,, and g, are ages on both devices are almost identical.
identical for both types of device, it is expected that the low R , For R F measurements, devices having similar DC charac-
of the shorter n - g a t e width gives better noise characteristics teristics were selected from both types of gate structure. The S
than those of the longer T-gate width. In addition, the short parameters were measured from 1 to 30GHz using an
1116 E L E C T R O N I C S LETTERS 10th June 1993 Vol. 29 N o . 12