IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. IT-28, NO.

5, SEPTEMBER 1982 753

Square Root Kalman Filtering for

High-Speed Data Received
over Fading Dispersive
HF Channels
FRANK M. HSU, MEMBER,
IEEE

[5] to be unstable (worse than conventional Kalman algo-

Abs,stmct-This paper shows how to apply square root Kalman filtering to
high speed data transmission on fading dispersive high-frequency (HF) rithm) in time-varying channel environments. In other ap-
channels. High data-rate HF transmission can be achieved by use of a
plications, researchers have shown that square root formu-
single carrier bandwidth efficiency modulation that is demodulated with
decision-feedback equalization (DFE). The DFE coefficients are updatedlations of the Kalman algorithm have inherently better
stability and numerical accuracy than does the conven-
adaptively by square root Kalman algorithms. Theoretical formulation and
tional Kalman algorithm [3], [6]-[13]. The improved
mechanization procedures for the square root Kalman algorithms are given.
Computer simulation indicates that good error-rate performance can be numerical behavior of the square root algorithms is due in
achieved with these algorithms for rapid fading HF radio channels.
large part to a reduction of the numerical ranges of the
I. INTRODUCTION variables. Loosely speaking, one can say that computations
which involve numbers ranging between 10-N and 10 N are

T HIS PAPER is concerned with square root Kalman reduced to ranges between lOeN/* and 10N/‘. Thus the
algorithms for adaptive equalization of fading disper- square root algorithm achieve accuracies that are compara-
sive high-frequency (HF) radio channels. High-speed adap- ble with a conventional Kalman algorithm that uses twice
tive communication systems are constrained by the non- the numerical precision [3].
ideal characteristics of HF channels such as bandwidth The application of conventional Kalman filtering theory
restrictions, signal fading, and multipath dispersion. These to the adaptive equalization techniques has been consid-
channel limitations present the major obstacles to achiev- ered by Godard [14]. In this paper, new algorithms based
ing increased digital transmission rates with the required on the Kalman/Godard algorithm and the Carlson/Bier-
accuracies. man [3], [6], [12] (or Gentleman [13]) U-D covariance fac-
High data-rate HF transmission (up to 9600 bits/s in 3 torization filter are proposed for updating the tap gains of
KHz bandwidths) can be achieved by use of a single carrier a decision-feedback equalizer for fading dispersive HF
bandwidth-efficient modulation that is demodulated with a channels. These new algorithms are applicable to complex
decision feedback equalizer (DFE) operated adaptively [I]. (i.e., angle-modulated) input signals and time-varying
The DFE coefficients can be updated adaptively by gradi- channels. Theoretical formulation and mechanization pro-
ent, Kalman algorithms, etc. It is found [2] that only the cedures for the square root Kalman algorithms are given
Kalman algorithm provides a tracking rate sufficient to here. Computer simulation indicates that good error-rate
follow the time-varying HF channel. Unfortunately, the performance can be achieved by these algorithms for rapidly
Kalman algorithm has been found to be sensitive to com- fading HF radio channels.
puter roundoff errors; numerical accuracy due to roundoff The computational requirements of the square root algo-
sometimes degrades performance to the point where the rithms for an equalizer with N taps are 1.5 N2 complex
results become meaningless [3]. Simulations that we have multiplications per cycle, which are roughly equivalent to
conducted in fixed-point arithmetic have shown the perfor- the requirements of the conventional Kalman algorithm.
mance of a DFE with conventional Kalman updating does Typically, the keying rate of an HF modem with 3 KHz
indeed degrade rapidly as the computer-word size is transmission bandwidth is set to 2400 bits/s. A decision
decreased. Recently, Falconer and Ljung [4] have shown a feedback equalizer with 15 feedforward taps and 14 feed-
“fast Kalman algorithm” which requires only n rather than back taps can handle multipath time spread up to 6.25 ms.
n2 operations per cycle. In fact, the fast Kalman algorithm The multipath spread is less than 5 ms in typical HF
was found, independently, by us and by Lim and Mueller channels [ 151. The least-squares lattice structure introduced
by Morf et al. and adapted by Satorius and Park [16] for
Manuscript received March 25, 1980; revised July 6, 198 1. equalizer adjustment algorithm is a linear (all zero) equali-
The author is with the Sylvania Systems Group, Communications
Systems Division, GTE Products Corporation, 77 A Street, Needham zation technique. It is well known [17] that linear filters are
Heights, M A 02 194. unable to cope with severe fading-dispersive channels. The

OOIS-9448,‘82/0900-0753$00.75 01982 IEEE

754 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. IT-28, NO. 5, SEPTEMBER 1982

least squares lattice decision-feedback equalizer formulated input to the feedback section is the output symbol decision
by Shensa [ 181 requires about 97 N multiplications per sequence from the nonlinear symbol detector. The equalizer
cycle, where N is the number of stage (in a N-zero, N-pole output can be expressed as
channel). The least-squares lattice DFE requires more com- M N-M- I
putation than the square root algorithms shown in this ik = 2 ajYkij + 2 b/ik-j> (2.4
paper for multipath time spreads up to 6.25 ms. Conse- j=O j=l

quently, DFE’s with square root Kalman algorithms are where ik is an estimate of the k th information symbol and
more efficient than DFE’s with least-squares lattice algo- {fk-,; . .,‘k-N+M+I) are P reviously detected symbols. The
rithm for HF modem implementation. The square root decision Ik is formed by quantizing the estimate fk to the
Kalman algorithms described are shown to be remarkably nearest information symbol. The least mean-square error
stable with rapid convergence and tracking abilities for (LMSE) criterion provides a practical means for selecting
short computer-word sizes. The implementation of these the equalizer tap coefficients {a,} and {b,}. Based on the
algorithms for high-speed HF modems has been accom- assumption that previously detected symbols in the feed-
plished recently [ 191. back section are correct, the minimization of the mean
square error
II. COMMUNICATIONS
SYSTEMMODEL
G2 = EIIk - fk12 (2.5)
The HF channel can be modeled as a tapped delay-line leads to the following set of linear equations for the tap
with time-varying coefficients followed by addition of coefficients of the feedforward filter
zero-mean white-Gaussian noise, v(t) with variance, cr,”
[13]. The tapped delay-line has taps spaced at no more
than the symbol interval T and nonzero tap coefficients
corresponding to discrete multipath propagation. The dis-
m = O,l;..,M, (2.6)
crete tapped delay-line model can combine the transmis-
sion filter, the channel, the matched filter, the symbol rate f; = Ai- (2.7)
samples and a noise whitening filter [l]. The tap coeffi- :Hk)
cients are assumed to be independent complex-Gaussian
random variables { fi, i = 0,. . . , L} with zero means and nk=j$, (2.8)
fixed variances, E{ ] f; I’}. The tap coefficients are generated
by filtering white-Gaussian noise through a filter whose
bandwidth is on the order of the fade rate. where es2and ui’, are the variances of the signal Ik and noise
It can be shown that passage of the random signals nk, respectively. The tap coefficients of the feedback filter
through the discrete tapped delay-line results in a received are given in terms of coefficients for the feedforward filters
signal sequence {Rk} which can be expressed as by the following expression

R, = i fmIk--m + vk, k = 0,1,2;.., (2.1) 6, = - 5 ajf’,+j, m = 1,2;**, N - A4 - 1. (2.9)

m=O
where {Ik} is the random signal sequence. In order to
reduce the signal fluctuations due to the fading channel,
the received signal is normally regulated by an automatic
gain control (AGC) filter that keeps the magnitude of the
output signal yk nearly constant:
Hk = XR,R*, + (1 - x)H,-,,
where h is the AGC constant and asterisk (*) denotes
complex conjugate. The signal y, can be demodulated with
a decision-feedback equalizer (DFE) operated adaptively.
The DFE consists of two sections, a feedforward section
and a feedback section. The feedforward section is a linear
transversal filter with tap coefficients {uj, j = 0,. . . ,M}
spaced at no more than the symbol interval T. The input to
the feedforward section is the sequence { yk} of noisy
received signal samples. The feedback section is also a
linear transversal filter with tap coefficients {bj, j =
1; . *, N - M - 1) spaced at the symbol interval. The
j=O
One can easily show that these values of the feedback
coefficients result in complete elimination of intersymbol
interference from previously detected symbols, provided
that previous decisions are correct. The optimum number
N - M - 1 of feedback taps may be set equal to the
number L of interfering symbols given in the discrete
channel model, if the tap a, of the feedforward filter is
(2.2)
selected as the main tap for synchronization purposes.
The equalizer tap coefficients {u,} and {bj} may be
(2.3) computed exactly from (2.6) and (2.9) provided the values
of u,‘, q”,, and {f;} are known. Normally, however, tap
coefficients are computed adaptively because the channel is
not known a priori at the receiver. The calculation of the
theoretical tap coefficients from (2.6) and (2.9) is primarily
useful in providing a baseline performance level for com-
parison with the various adaptive schemes.
III. KALMAN/GODARD FILTER
The solution of (2.6) and (2.9) for the optimum tap
coefficients involves the inversion of a matrix that depends
on the signal and noise statistics as well as the characteris-
HSU: SQUARE ROOT KALMAN FILTERING 755

tics of the channel. Generally, this is not known to the The state transition matrix @(k, k - 1) is defined in a
receiver so that iterative procedures must be used to de- system modeled as
termine the tap coefficients. The Kalman/Godard filter is C(k) = @(k, k - l)C(k - 1) + W(k), (3.8)
designed to update the equalizer tap coefficients, adap-
tively, from the received signal samples. During the train- where W(k) is N-dimensional vector of zero-mean white-
ing period, the equalizer continuously seeks to minimize noise process. It is assumed that the noise processes W and
the deviation of its sampled output signal from an ideal c are statistically independent. The parameter 5’ denotes
reference signal internally generated by the receiver. When the approximate final mean square error, E[] c 1’1. The
the residual distortion is small enough, the equalizer is matrix Q(k) represents the covariance matrix of W . Gener-
switched into the decision-directed mode, using as a refer- ally, the state transition matrix is not known to the receiver
ence a reconstructed signal obtained by quantizing the in the equalization process. Nevertheless, one can assume
output signal of the equalizer to the nearest symbol. Then that the optimum tap-gain values are randomly varying
the equalizer has the ability to self-adapt to changes in about a mean value [14], i.e.,
channel characteristics during the transmission process. = C( k - l)opt + AC(k),
c(%t (3.9)
The Kalman/Godard algorithm will be described here for
updating the DFE coefficients. This algorithm is by far the where AC(k) is considered to be a noise process. The
most powerful, i.e., it is fastest, known adaptive algorithm matrix Q(k) is then reduced to
for adjusting the equalizer coefficients.
Q(k) = +W))@C(k))*‘] (3.10)
-At time k, the @puts { y,, JJ~+,, . . . ,yK+ M} and decisions
(Ik--l> Ik--2,. . .Jk-,v+M+J are in the feedforward and and, at each step from (3.7), the predicted-error covariance
feedback sections of decision feedback equalizer, respec- matrix is given by
tively. The signals can be represented by a vector
P’(k) = P(k - 1) + Q(k). (3.11)
X(k)’ = (Yk, Y~+I,’ ’ ‘,Yk+M,
Exact calculation of the correlation matrix Q(k) is impos-
‘ik-,,~k-2,‘.‘,~k~N+M+,), (34 sible because the optimum tap coefficients are unknown to
where t denotes the transpose. the receiver. It has been shown [14] that C(k) converges to
The equalizer tap coefficients at time k are represented C(k)*,t within less than 2N steps. If the matrix Q(k) is
by the vector assumed to be negligible for a slowly time-varying channel,
then one has [ 141
C(k - I)‘= (ao, al,-“,aM, b,, bz,-,b,+,v-,I,

and its output is C(k - l)‘X(k), which may differ from

(3.2)

and
X*(m)X(m)’
1-’ (3.12)

the ideal output I, by an error r(k)

E[X*(k)X(k)‘] = r$‘k-‘P(k - l)-‘, (3.13)
c(k) = Ik - C(k - l)‘X(k). (3.3)
as k goes to infinity. It can be verified [ 141 that the P(k)
The goal of the adaptation algorithm is to adjust C(k)
(or P’(k)) matrix elements converge to zero. Consequently,
iteratively to converge toward an optimum vector C(k),,,,
the parameter (Y’= X( k)‘P’( k)X*( k) + E’ converges to E’.
which minimizes the mean-squared value of E(k). The new
In the steady-state limit, the Kalman gains in (3.5) are
estimate of C(k) is given by
reduced to
C(k) = C(k - 1) + G(k)c(k), (3.4)
G(k) = P(k - l)X*(k)/[‘. (3.14)
where
Substituting (3.4), (3.13), and (3.14) into (3.10) yields
G(k) =[X(k)‘P’(k)X*(k) + $P’(k)X*(k), (3.5)
Q(k) = k-‘P(k - 1). (3.15)
P(k) = P’(k) - G(k)X(k)?‘(k), (3.6)
The matrix Q(k) is identically zero in the case of sta-
P’(k) = @(k, k - l)P(k - l)@*‘(k, k - 1) + Q(k).
tionary channels. For slowly time-varying channels,
(3.7) Q(k) may be assumed proportional to P( k - l), Q(k) =
@(k - l), where q (K 1) is a small constant. Substituting
All constants, vectors and matrices (except E’) defined in
(3.15) into (3.1 l), (3.5), and (3.6), we obtain
(3.1), (3.2), (3.3), (3.4), (3.5), (3.6), and (3.7) are complex
quantities. The quantities ak, b,, I,, and c(k) are scalars;
X(k), C(k), and G(k) are vectors; and P(k), P’(k),
G(k) =[X(k)‘P(k - 1)X*(k) + $‘P(k - 1)X*(k),
Ca(k, k - l), and Q(k) are square matrices. The asterisk (3.16)
and superscript t denote complex conjugate and transpose
respectively. The matrix P(k) represents the covariance of P(k) = (1 + q)[P(k - 1) - G(k)X(k)‘P(k
the error in the estimate of the actual equalizer coefficients.
156 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. IT-28, NO. 5, SEPTEMBER 1982

where The algorithm presented in this section gets around both

of the above limitations. It involves an upper complex
+5’ (3.18) triangular factorization of the filter error covariance ma-
1 +q’
trix, i.e., P = U*DU’. Efficient and stable measurement-
The parameter .$ and q should be adjusted to obtain updating recursions are achieved for the upper triangular
optimum performance. It is interesting to note that (3.16) matrix U and the diagonal matrix D. This method is based
and (3.17) have the form that would result from deriving on (3.17) in which the correlation matrix Q is computed
the Kalman updating algorithm to minimize the exponen- implicitly in each iteration. The details of the algorithm
tially weighted square error at time k: will be given in the following discussion.
The price to be paid in direct implementation of (3.17) is
F2(k) = ; wk-’ 1I, - C(k)‘X(n) 12, (3.19) quite high; however, the square root factorization algo-
n=l
rithm allows the multiplications by (1 + q) to be incorpo-
Specifically, (3.16) and (3.17) are the solution to this mini- rated into a simple (real-valued) diagonal matrix. This is
mization problem if we let (1 + q)-’ = w and t; = w. The accomplished as follows.
exponentially weighted error criterion has intuitive appeal Suppose that the error covariance matrix P is expressed
in that the distant past is “forgotten”, which is a desirable in factored form as
property for following time-variations of the channel and
P(k) = U*(k)D(k)U’(k), (4.1)
helping avoid problems associated with digital-roundoff
errors. This method has been investigated by Satorious [ 161 where U is an upper triangular matrix with unit diagonal
and Shensa [ 181, etc. Unfortunately, the optimum values of elements and nonequal off-diagonal elements pij, i =
q and 5 are quite different from that specified in the 1,2;. . ,N - 1;j = i + 1, i + 2;. .,N, and D is a diagonal
exponentially weighted method. matrix with real-valued diagonal elements (d,, d,, + * . , d,).
The starting conditions for C(k) and P(k) are given by Also let the vectors F and V be defined by
c(0) = 0, (3.20) F(k - 1) = U’(k - 1)X*(k) (4.2)
P(0) = I, (3.21) and
where 0 is the zero vector and I is the identity matrix. V(k - 1) = D(k - l)F(k - 1). (4.3)
Substituting (3.16), (4.1), (4.2), and (4.3) into (3.17) yields
IV. SQUARER• OTKALMANFILTER
U*(k)D(k)U’(k) = (1 + q)U*(k - l){D(k - 1)
The classic Kalman filter presented in Section III pro-
-a-‘V(k - l)V*‘(k - l)} U’(k - l), (4.4)
vides a tracking rate sufficient for fading HF radio chan-
nels, thereby achieving good error-rate digital communica- where a = X(k)‘P(k - 1)X*(k) + 5. If we define Uand 5
tion performance. Unfortunately the covariance update factors of the terms inside the braces of (4.4), we obtain
formula (3.6) is numerically unstable. The main reason for U(k - l)*D(k - l)ut(k - 1)
this instability is that P is computed as a difference of two
positive semidefinite matrices. Numerical accuracy is re- = D(k - 1) - (Y-‘V(k - l)l/*‘(k - l), (4.5)
duced in every iteration. Moreover, numerical deterioration and
is also associated with high dimensionality of G and P on
X and with the accumulated effects of roundoff error. The U*(k)D(k)U’(k) = (1 + q)(U(k - l)c(k - l))*
accuracy degeneration of the classic Kalman filter may
.6(k - l)(U(k - l)U(k - 1))‘. (4.6)
result in a P matrix which is indefinite (having both
positive and negative eigenvalues). From (4.6) we can identify
The Carlson/Bierman U-D covariance factorization filter U(k) = U(k - l)i?(k - 1) (4.7)
[3], [6], [ 121 involves an upper triangular factorization of a
real-valued filter error covariance matrix, i.e., P = UDU’. and
This algorithm can be linked to square root Kalman filter- D(k) = (1 + q)B(k (4.8) - 1).
ing, although the computation of square roots is not re-
Thus the updated U-D factors are determined in terms of
quired. The U-D factorization procedure guarantees non-
the U-D factors of D(k - 1) - a-‘V(k - l)V*‘(k - 1)
negativity of the computed covariance matrix. However,
and U(k - 1).
the Carlson/Bierman algorithm cannot be directly applied
to HF communication systems. First, the baseband signals Theorem 1: Let the factors X, F, and V be denoted as
in the equalizer are generally represented by complex vari-
X’(k) = (x,, x2,-.., xi,& (4.9)
ables. Second, the correlation matrix Q shown in (3.7), is
not included in the Carlson/Bierman square root Kalman F’(k - 1) = tf,, f2,-,.&), (4.10)
algorithm. Godard suggested the need for this modification and
in order to apply the Kalman algorithm to time-varying
channels [ 141. V’(k - 1) = (o,, tag,...,+), (4.11)
HSU: SQUARE ROOT KALMAN FILTERING 151

then the U-D factors of D(k - 1) - a-‘V(k - l)V*‘( k - utilized for updating the tap gains. On the other hand, the
1) can be obtained by evaluating the following ordered convergence rate of the a1gorith.m is primarily controlled
equations recursively for j = 2,. . . , N: by the parameter 6. A large value of [ will produce a slow
rate of stable convergence, whereas a small value of 5 will
d, = d,(k - l)$ (4.12) result in a fast rate of convergence but at the sacrifice of
stability. From (4.9) to (4.22), it is found that d,(l) is
d, = dj(k - l)?, (4.13) significantly different from dj(l), j = 2,3, * * *, N for small
J values of [(e.g., E = 0.001; (Y, = 1; d,(l) = 0.001; dj(l) = 1,
and for i = : 1,2; f *,j - 1, j = 2,3,-s. , N; q e 1). The magnitude disparity of d, and
dj in the start-up period can significantly disturb the
vi*vj equalization process. The disparity can also be ,expected to
Fij = - - (4.14)
oljd, ’ degrade the performance of the algorithm if a periodic
resetting scheme as described in Section VII is adopted.
where
The algorithm can be modified in such a way that it is
fi = x1*, (4.15) stable for various values of 5 without sacrificing the rate of
1-l convergence. A revised square-root formulation described
4 = 2 pi,j(k - 1)x,* + xj*, j = 2,3;*.,N, below will serve this purpose. This algorithm is more suited
i=l for implementing the receiver with a processor of limited
(4.16) word size.
vj = d,(k - l)& j= 1,2;..,N, (4.17) Theorem 3: If (3.6) is modified according to the follow-
ing equation
a, = E + qfi*, (4.18)
aJ = aj-, + vj4.*, j= 2,3;**,N. (4.19) P(k) = P(k - 1) + Q(k) - G(k)X(k)‘P(k - l),
(54
Theorem 2: The intermediate Kalman gain G and the
updated error covariance factors U and D can be obtained then the intermediate Kalman gain G and the updated
from the following algorithm error covariance factors U and D can be obtained from the
following algorithm
gj = vj> j = 1,2;.e,N, (4.20)
gj = v~o/, j = 1,2;.+,N, (5 4
d,(k) = (1 + q)& (4.21)
d,(k) = (1 + q)d,, j= 2,3;..,N, (4.22) d,(k) = d,( k - I)h (E + hr)
(5.3)
’ ((~1 + hr) ’
‘J = -h/ffj- 12 j= 2,3;.+,N, (4.23)
(0(/-l + ht)
Pii = PiJ(k - 1) + gi*hj, i = 1,2;..,j - 1, d,(k) = dj(k - l)h, (aj+ h,) , J = 2,3,.-.,N,
(4.24)
(5.4)
gj = g, + vjpTj(k - l), i = 1,2;..,j - 1,
(4.25) j= 2,3;..,N, (5.5)
A, = - (aJ-,‘+ h,) ’
where g, on the right-hand side of (4.25) is gi obtained from
(j - 1)th iteration, and Pij(k) = PrJ(k- 1) ’ gi*‘j, i = 1,2;..,j - 1,
G’= tg,,g2>->g,v)> (4.26) (5.6)
and the final Kalman gain is given by g, = gi + vjpTj(k - l), i = 1,2,.-e ,j - 1, (5.7)
G(k) = G/a,. (4.27)
where gi on the right-hand side of (5.7) is gi obtained from
The square root Kalman algorithm is given by Theorems (j - 1)th iteration, and
1 and 2.
f, = x1*,
V. A REVISEDSQUAREROOTFORMULATION j-l

h= 2 pij(k- 1)x,*+x,*, j=2,3;..,N,

The square root algorithm described in the previous i=l
section is a stable algorithm as long as the parameters 5
(5.9)
and q are well chosen. 5 and q are sensitive to the rate of
convergence of equalization and fading rate of the channel, vj = dj(k - l)& j = 1,2;**,N, (5.10)
respectively. Intuitively, the value of q after equalization
should be proportional to the magnitude of the fading rate.
a, = E + v,f,*, (5.11)
The optimum value of q may also depend on the algorithm aJ = ffj-, + vj5.J;*, j = 2,3;**,N, (5.12)
758 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. IT-28, NO. 5, SEPTEMBER 1982

h, = aNq, (5.13) 0.001 gave identical results for the same sequence {Ik} and
the same sequence of noise samples nk after equalization.
h, = 1 + q, (5.14)

G’= (81, &,***,gN), (5.15) VI. COMPUTATIONALPROCEDURES

and the final Kalman gain is given by Significant computer storage can be saved for numerical
computation of the square root Kalman algorithms. The
G(k) = G/a,. (5.16)
computational procedures described here are such that the
Proof: Let F and V be given in (4.2) and (4.3), the lower triangular portion and the diagonal elements of U
U-D factors of P given in (5.1) become are not used (the diagonal elements of U are unity). The
vector V is replaced by vector G for storage reduction. The
U*(k)D(k)U’(k) = U*(k - l){D’(k - 1)
computational procedures of the square root (U-D factori-
-a-‘V(k - l)V*/*‘(k - l)}U’(k - l), (5.17) zation) formulation defined in Section IV are given by
where D’ is a diagonal matrix with diagonal elements N
(1 + q)d,). With extensive c(k) = & - 2 xjcj, (6.1)
((1 + q)d,, (1 + qM,,---,
j=l
mathematical manipulations (see Appendix), it can be
shown that the U-D factors of terms inside the braces of f, = x1*, (6.2)
(5.17) j-l

P= D’(k - 1) - a- ‘V(k - l)V*‘*‘(k - 1) (5.18) fi= 2 pi,jxi* + xj*, j= 2,3;**,N, (6.3)

i=l
are given by j= 1,2;..,N,
gj = dj(k - l>fi> (6.4)
d, = d,(k - 1)h (’ + hf) (5.19) al = 5 + g,f,*, (6.5)
4 6% + 4) ’
aj=aj-, +gjfj*, j=2,3;**,N, (6.6)
(aj-1 + ht)
dj = dj(k - l)h, (aj + h,) > J = 2,3,- . .,N, h, = 1 + q, (6.7)

(5.20) ,=1 (6.8)

a1
hpi*vj
pij= - (aj+h,)d,, i= 1,2,..-,j-lT (5-2l) d,(k) 7 d,(k - l)h&y, (6.9)
P = aj-19 (6.10)
where the variablesh, vj, aj, h,, and h, are given in (5.8) to
hj= -fjYY (6.11)
(5.14). Then the U-D factors of P follow immediately from
Theorem 2, i.e., &, (6.12)
d,(k) = 21, (5.22) aj
dj(k) = Jj, j= 2,3;..,N, (5.23) dj(k) = dj(k - l)h&, (6.13)

PI = Pij? (6.14)
j= 2,3;.*,N, (5.24)
Aj = - (ajel’+ h,) ’ Pij = P* + gi*xjY (6.15)

Pij(k) = Pij(k - 1) + gi*Aj, i = 1,2;**,j- 1, gi = gi + gjPl*’ (6.16)

(5.25) All constants, except 5, aj, j = 1,2,. . . , N, y, h,, q, dj,

j = l,;.., N, and /? defined in (6.1) to (6.16), are complex
gi = gi + vjpTj(k - l), i = 1,2;.,,j - 1,
quantities. The quantities defined in (6.1) to (6.9) are
(5.26) computed first. Equations (6.10) to (6.16) are evaluated
where recursively for j = 2,3,. * a,N, and (6.14) to (6.16) are
computed in order for i = 1,2,. . . J - 1. The equalizer tap
G’ = (g,, g,,. . -,gN)? (5.27)
weights C’ = (C,, C,; . .,C,) are given by
and the final Kalman gain is given by (6.17)
e = c(k)u,
G(k) = G/a,. (5.28) q.=q.+gje, j = 1,2;..,N, (6.18)
Q.E.D.
where the computation of the final Kalman gain, gj = gjy,
The revised square root formulation is stable in terms of j = 1,2;**, N, is avoided by adding an intermediate step
various values of E. A small value of 5 could be used for a defined in (6.17). Equations (6.1) to (6.18) are evaluated for
brief portion of the start-up phase to achieve a fast reduc- every new input signal X defined in (3.1).
tion of mean-square distortion. However, for a fading For the revised square root formulation, the computa-
channel, four computer runs with 5 = 1 .O, 0.1, 0.01, and tional procedures are given in (6.1) to (6.18) except ,(6.8),
HSU: SQUARE ROOT KALMAN FILTERING 159

TABLE I
COMPUTATIONALREQUIRRMENTSFORSQUAREROOTKALMANALGORITHMS

Equivalent Equivalent
Real Real Real Computer
Algorithms Multiplications Additions Reciprocals Storage
Square
Root 6N2 f IlN 6N2 + 6N N N2+8N+ 14
Algorithm
Revised
Square
Root 6N2+ llN+ 1 6N2+7N+ 1 N+ 1 N2t8Nf 16
Algorithm

(6.9) (6.10), and (6.12) are modified as
1 equalizer tap weight with a two-path fading channel model
Y=- a, + h,’
d,(k) = d,tk - l)h,(E + h,)y,
P = aj-, + h,,
1 Fig. 3 shows the convergence properties of the equalizer
y=-----aj + h, ’
where h, = aNq.
The computational requirements of the square root algo-
rithms for an equalizer with N taps are listed in Table I.
The requirements may be slightly different for various
signal processors.
VII. SIMULATION RESULTS
In this section we present some results of computer
simulation of the properties and performance of the algo-
rithms presented in the previous sections. W e have used the
fading dispersive HF radio channel model described previ-
ously. A complex Gaussian noise is also generated and
added to the received signal.
This section contains the properties of convergence, the-
oretical and simulated performance results for the DFE
with 8-ary (8 - $) phase-shift keying (PSK) modulation.
Extensive simulations were conducted in the investigation
of the performance of the receiver using the DFE with
square root Kalman algorithms for mitigating the inter-
symbol interference and signal fading. The channel fade
rate is 1 Hz and the channel model consists of two fading
paths (except in Fig. 3). The received signal is regulated by
an AGC filter in the front end with an AGC constant,
X = 0.02. The DFE consists of two sections, a feedforward
section and a feedback section. The numbers of feedfor-
ward and feedback taps are set to be greater and equal to
the number of interfering symbols given in the discrete
channel model, respectively. The initial values of U and D
in the square root algorithms are
d, = 1.0, j= 1,2;..,N,
pij = (0.0, o.o), i = 1,2;*.,N;
-j = i + 1, i + 2;..,N,
where x, y, and (x, y) denote the real part, imaginary part
and complex quantities.
Figs. 1 and 2 depict the convergence properties of the
(6.19) in the first 40 iterations. The DFE with square root and
revised square root Kalman algorithms converges very
(6.20) rapidly following the initial start-up. The speed of conver-
(6.21) gence and stability of the algorithms will depend on the
parameters h, 6, and q chosen for various design objectives.
(6.22)
tap-weight with a two-path fixed channel using the revised
square root Kalman algorithm. It is clear that the conver-
gence of a DFE with the revised square root algorithm is
extremely fast and stable on fixed channels. These results
indicate that the square root Kalman algorithms can also
be used to obtain high stability and fast start-up on tele-
phone channels.
Figs. 4 and 5 depict the performance of the DFE for 8 +
PSK with a two-path fading channel model using the
square root and revised square root Kalman algorithms.
The quantized symbol ik is assumed to be equal to the
actual transmitted symbol I,. For purposes of comparison,
the error rate based on exact knowledge of the equalizer
coefficients (LMS sense) is also illustrated. Both the perfor-
mance based on theoretical tap-weights and square root
Kalman filters are depicted. In Fig. 4, the performance of
the DFE with the square root and revised square root
Kalman algorithms exhibit 1.25 and 1 dB degradation
respectively related to the ideal performance at P, = 10e3.
The performance of the DFE with the revised square root
algorithm is slightly better than that with the square root
algorithm. This result is due to the improvement of stabil-
ity of the revised square root Kalman algorithm. In Fig. 5,
the revised square root Kalman algorithm has been mod-
ified to incorporate the periodic resetting of the U-D
factors according to (7.1) and (7.2). The resetting of matrices
may slightly degrade the performance in floating-point
arithmetic (as indicated in Fig. 5) but will considerably
improve the performance in fixed-point arithmetic. The
resetting scheme is used to prevent the propagation of
roundoff errors associated with the adaptation history.
(74
APPENDIX
(7.2)
The U-D factors of (5.18)
P= D’(k - 1) - cC’V’(k - l)V*‘(k - 1) (A.11
760 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. IT-28, NO. 5, SEPTEMBER 1982

0.9 - THEORETICAL TAP GAIN

0.8 -

0.7 -

0.6 -
z
d

TAP GAlN OBTAINED B Y SOUARE ROOT

KALMAN ALGORITHM
(A = 0.02: t = 0.001: q = 0.21

I I I
0 10 20 30 40
N U M B E R OF ITERATION

Fig. 1. Properties of convergence of the equalizer tap weight (real part of the main tap shown; DFE; S+J; two-path fading
channel; 1 hz fade rate; N = 5; SNR = 20 dB).

1.0 ,TAP GAIN OBTAINED B Y REVISED SQUARE ROOT KALMAN ALGORITHM

(h = 0.02: $ = 0.001: q = 0.08)
0.9 t , \

0.8 - I vv-
AA
THEORETICAL TAP GAIN

0.7 - I

zd /vi.,
,/'.i.,.,*,.,.~'~.-.-. /'
0.6

020.5 '1.-.+,./*\“‘\.,./ .-./'\.,.

I i i
TAP GAIN OBTAINED B Y REVISED SQUARE ROOT KALMAN ALGORITHM
(h = 0.02; [ = 1.0: q = 0.08)
0.3

I I
0 10 20 30 40
N U M B E R OF ITERATION

Fig. 2. Properties of convergence of the equalizer tap weight (real part of the main tap shown; DFE; 8+; two-path fading
channel; 1 Hz fade rate; N = 5; SNR = 20 dB).

may be found by forming the quadratic form X’pX*, Let us suppose wN = --a-‘, then (A.2) becomes
- -- X’PX* = X’D’X* + w,X’YV*‘X*
X+X* = X’U*D U’X*
= F*~DF

N-l

where

F= u’x*, (A.31
f, = xl*, (A.41

i I
j-1 N-l
f, = x j&,x; + x*
J'
j = 2,3, . . . . N. (A4 +WNUN*XN* 2 vjx, .
i=l j=l
HSU: SQUARE ROOT KALMAN FILTERING
761

1.0

0.9

0.8
TAP GAIN OBTAINED BY REVISED SDUARE ROOT KALMAN ALGORlTHM
(t = 0.001: 9 = 0.081
0.7

0.6

z
3 0.5

3
0.4

0.3

0.2

0.1

0.0 I t I I
10 20 30 40
NUMBER OF ITERATION

Fig. 3. Properties of Convergence of the Equalizer tap weight (main tap shown; DFE; S+; two-path fixed channel; N = 5;
SNR = 30 dB).

THEORETICAL

SNR Eb/N, (dB)

Fig. 4. Error rate performance (DFE; two-path fading channel; 1 Hz fade rate; X = 0.02).
162 IEEE TRANSACTIONS ON INFORMATIONTHEORY,VOL.IT-28,~o. 5, SEPTEMBER 1982

10-l

i
E 10-2 TAP WEIGHT

d
ii
5
b
c FOR EVERY 100
3 ITERATIONS)

2
m 10.:
0
E

lo”

10-l I I I \ I 1
0 5 10 15 20 25 30

SNR E,,/N, IdB)

Fig. 5. Error rate performance (DFE; two-path fading channel; 1 Hz fade rate; X = 0.02).

Denote the quantities dN, W,- , , and ijN by

d,=d~+w,Iu,12, (A.71

WrJ:,
wp,=- (A.8)
dN ’
wNvN
ll,N= -u * j=1,2;..,N-1, (A.91
dN J’

Equation (A.6) may be written as

N-l d~+wNluN12 IV-’
x’Px*= 2 d;IXj12+dNIXN12+WN d 1 jz, ‘jxJ I ’ + WN”NXN
j=l N

N-I N-l N-l

= 2 d;lxjl’+ WN-11 ,zl oj’jXj12 f ,“, FJ~xT+xg '2~

j=l

N-l
= jz, d;lXj12fWN-I (A. 10)

Similarly, the inductive reduction follows recursively, and after N - 1 steps the quadratic form has been reduced to

xFx*= 5 d,lj12 (A.1 1)

j=l
HSII: SQUARE ROOT KALMAN FILTERING 163

provided that ACKNOWLEDGMENT

2, = d; + w, 1u, I*, j = 1,2;..,N, (A.12)
The author expresses his appreciation to P. Anderson for
wj d; (A 13) help and reviews of this work.
w,-, =-r-) j= 1,2;..,N,
dj
and REFERENCES

i= 1,2;..,j- l;j=2,3;..,N. [‘I H. de Pedro, F. M. Hsu, A. A. Giordano, and J. G. Proakis, “Signal

design for high speed serial transmission on fading dispersive chan-
nels,” in Nat. Telecommun. Conf., Birmingham, AL, Dec. 1978.
(A.141 [21 F. M. Hsu, A. A. Giordano, H. E. de Pedro, and J. G. Proakis,
Multiplying (A.12) by l/dJw,, one obtains “High speed modem techniques for fading dispersive channels,” in
NUSC Workshop Comm. Fading Dispersive Medium, New London,
j= 1,2;..,N. CT, June 5-6, 1979.
(A.15) [3] G. J. Bierman, Factorization Methods for Discrete Sequential Estima-
tion. New York: Academic, 1977.
Substituting (A. 15) by the following equation ]41 D. Falconer and L. Ljung, “Application of fast Kalman estimation
to adaptive equalization,” IEEE Trans. Comm., Oct. 1978.
d; = (1 + q)d, = h,d,, j= 1,2;..,N, G4.w [5] T. L. Lim and M. S. Mueller, “Rapid equalizer start-up using
least-squares algorithms,” in Int. Conf. Comm., 1980.
we have [61 N. A. Carlson, “Fast triangular formulation of the square root
filter,” AIAA .I., vol. 11, no. 9, Sept., 1973.
[71 N. A. Carlson and A. F. Culmone, “Efficient algorithms for on-
board array processing,” in Int. Conf. Commun., Boston, MA, June
10-14, 1979.
[sl A. Andrews, “A square root formulation of the Kalman covariance
equations,” AIAA J., vol. 6, no. 6, June 1968.
. _ J. F. Bellantoni and K. W . Dodge, “A square root formulation of
191
=-- 4a + a,-1 j = 1,2;. . ,N, (A.17)
the Kalman-Schmidt filter,” AIAA J., vol. 5, no. 7, July 1967.
[loI R. H. Battin, Astronautical Guidance. New York: McGraw-Hill,
hq ’ 1964, pp. 338-339.
where [Ill S. F. Schmidt, “Computational techniques in Kalman filtering,” in
Theoty and Applications of Kalman Filtering, C. T. Leondes, Ed.,
aj= i diIf,]*+<> j= 1,2;..,N, (A.18) Advisory Group for Aerospace Res. Develop., AGARDOgraph 139,
i=l
1970.
N
1121 G. T. Bierman, “Measurement updating using the U-D factoriza-
tion,” in Proc. IEEE Conf. Decision Contr., Houston, TX, 1975.
a=-wi’=aN= zdiIf,l*+[, (A.19) ~131 W . M. Gentleman, “Least squares computations by Givens trans-
i=l formations without square roots,” J. Inst. Math. Appl. vol. 12, 1973.
a0 = E. (A.20) [14l D. Godard, “Channel equalization using a Kalman filter for fast
data transmission,” IBM J. Res. Develop., May 1974.
Substituting (A.16) and (A.17) into (A.13) and (A.14), we obtain [15] F. H. Hsu, A. A. Giordano, H. de Pedro and J. G. Proakis,
“Adaptive equalization techniques for high-speed transmission on
ajp, + h, fading dispersive HF channels,” in NTC ‘80, Houston, TX, 1980.
jj=dJ(k- l)hq a-+h > j= 1,2;..,N, 64.21) [16] E. H. Satorius and J. D. Pack, “Application of least square lattice
J I
algorithms to adaptive equalization,” IEEE Commun., vol. COM-29,
h ‘Iv,?v, Feb., 1981.
- =-
pi, i= 1,2;.. J- 1; j=2,3;..,N, [171 J. G. Proakis, “Advances in equalization for intersymbol inter-
(a, + h,)d, ’ ference,” in Advances in Communication Systems, Theory and Appli-
cation., New York: Academic, 1975.
(A.22) [IQ M. J. Shensa, “A least-squares decision feedback equalizer,” in Int.
Conf. Commun., 1980.
where V. Allen and P. Anderson, “Wideband digital modem for HF
h, = aq. (A.23) [I91 radio,” in 26th Annual Southeastern ISA Conf., Orlando, FL, 1980.