Professional Documents
Culture Documents
Optimal Normalization Scheme
Optimal Normalization Scheme
lntroduction
tion wq is defined by
wq= (a + d) (w + 0) =
X w + dw - w.
Manuscript received May 16, 1974. Copyright 1975 by IEEE Trans. Once q = ai+bj + ck + d is obtained, the attitude matrix is
Aerospace and Electronic Systems, vol. AES-11, no. 4, July 1975. immediately given by the Euler four parameter matrix
IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-11, NO. 4 JULY 1975 443
Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on April 04,2020 at 05:16:26 UTC from IEEE Xplore. Restrictions apply.
d2 _a2 -b2 _ C2 2(ab + cd) 2(ac -bd)
The outputs of the gyros, being physical quantities, are where {X1(D), X2(D), X3(D)} are the eigenvalues of D and
never exact and the numerical methods used for solving (4) the maximum is taken over all orthonormal bases of vectors.
also introduce errors so an approximation to q, q= + bj +
ck + d, is actually calculated. When these quantities are Lemma 2: Let A be an arbitrary real 3 X 3 matrix. An
substituted into (5), we obtain B, which is usually not or- orthogonal matrix C will minimize hIA Cl if and only if -
of A, and if v
(vl,
1, , v,) represents a three-dimensional mal bases, then so too is {U'Q1* U'i42, U'i,&3} if U is orthog-
vector, onal. Also,
/ 3 1/2 3
lvllI Z; ( i))
2
tr (AGC) ;c
(AC ja>
We now prove that the unique orthogonal matrix C which where (<, () denotes the usual dot product of ct and ß and
minimizes hIB - Cli, B nonsingular, is given by (6). To this {°c} represents the standard orthonormal basis; i.e., c, has
end, we recall a well-known result in matrixes [3j. a one for its ith component and zeros for all other compo-
nents. Then, using the Cauchy Schwarz inequality, we have
Theorem 1: If A is an arbitrary nonsingular real matrix,
~~~~~3
then there exists a unique positive-definite matrix P and a tr (AC')<
11<
max tr (AC') -max -1 <AC'ai, cv.)
unique orthogonal matrix U such that A UP. Further-
more, max ZE JIACa°i11ll°Ei1
3
P -\IA-A and U-A(A 'A) -1/2. =-max 2; IIAC'ilI
The following result, which is proved in the Appendix, Since C is orthogonal, { CG' C'c2, C'ca3}
3 , { 1, QP2, 3},
is crucial. another set of orthonormal vectors, and
Lemma 1: Let D be a real 3 X 3 diagonal matrix and let 3 3
max ih1AIAC o 11- {aX il IA 1k
i
{01, 0,21,3 } represent an orthonormal bases of vectors.
Then
3 3 where the maximum is now taken over all orthonormal sets
max 11D 11 IXAID)l of vectors {~ i}. Then, using A = UP, it follows that
Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on April 04,2020 at 05:16:26 UTC from IEEE Xplore. Restrictions apply.
(
trC1 ~5 3 3 where Bo is determined from the alignment and Q2 is the
}i-l Pi1
{-pm}x=11
t i}l {l1 a }i-l
i 1
il skew symmetric angular velocity matrix whose elements are
But, P is a positive-definite real symmetric matrix, so it can computed directly from the outputs of the gyros. Equation
be diagonalized by an orthogonal similarity transformation. (7) is solved by a numerical scheme which results in ß. The
Hence, matrix B is orthogonalized (normalized) by one of many
suggested procedures [4] .
P-= UDU-1 = U1DU' The method given in Section II is superior to methods in-
volving direction cosines in two important respects. First,
and (4) contains only four unknowns, the components of q, as
compared to (7), which contains nine unknowns, the ele-
3 ments of B. Second, none of the orthogonalization proce-
tr (AC) . max 2; 11 U,DU4,~~i dures suggested for obtaining C from the solution of (7) are
{Ü i} i= 1
known to be optimal, whereas the orthogonalization proce-
3 3 dure given by (6) is optimal and requires at most thirteen
- max Z IIDUj 'i 11 = max i IID0 11 multiplications, three additions, and one division.
Theoretically, Theorem 3 could be used to find an opti-
where, again, {tk} = {U'1 4} is another set of orthonormal mal C even for the B obtained from (7). In practice, how-
vectors. We note that the eigenvalues are invarient under a ever, it is quite difficult to compute (B'B)-1/2 for most
similarity transformation, so \.(D) = X,(P) and, since P is matrixes B. lt is a characteristic of the Euler four parameter
positive-definite, the eigenvalues of D must be positive. matrix that
Theorem 2 is now a direct consequence of Lemma 1.
The following result which is discussed in a different (R'B)-112 (1/lq 12)I
form in [6] is the theoretical base for the normalization
scheme. Another well-known method involves the Euler angle
Theorem 3: If A is an arbitrary real invertible 3 X 3 representation. Although this representation generates only
matrix, then there exists an orthogonal matrix C which three unknowns, the differential equations that must be
minimizes jIA - CIi. Furthermore, C = A(A .A)-1/2. solved are nonlinear, as compared to (4) and (7), which are
linear, and, in addition, there exists a mathematical singular-
Proof: From Lemma 2, it is sufficient to maximize ity (gimbal lock). Methods based on the Euler angle reprc-
tr (AC'). Using Theorem 2, we do so by showing that sentation are known to be less efficient in general than
3 attitude determination by directional cosines; hence, they
tr(AC')= X.(P) for C=A(A'A)-'/2.
/
One of the most commonly used methods for obtaining We now prove Theorem 2. To do so we need the follow-
C employs directional cosines. Here the attitude matrix B ing result [5].
is given as the solution of the linear matrix differential equa-
tion Lemma 3: For any 3 X 3 real matrix A,
Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on April 04,2020 at 05:16:26 UTC from IEEE Xplore. Restrictions apply.
and the maximum occurs when ß is the normalized eigen- max IIDyl 1= 1Ä2 (D)
vector associated with the largest eigenvalue of v'4 . ll 11i= 1
In particular, if A = D, a diagonal matrix, Lemma 3 when y = (0, 1, 0)', which in turn implies that
reduces to 73 =(0, O, 1)'and that
max IIDY3 11 = 1X3 (D)
max IlDßll = max XD) = max L.(D) I . {y3}
[lll=31 i i I
Now, without loss of generality, let D = diag (X1, X2, X3) Thus,
with X1 . X2 - X3. If {01, 02, 03 } is any orthonormal set 3
of three-dimensional vectors, then max -E IIDO.Il . 1X1(D)l + max IIDyll + max IIDy3 11
3
max . IIDi ll < max IIDßlI + max (IlDß2 11 + IDß3 11) < IX1(D)I + 1X2(D)I + 1X3(D)I
4il 1 13l= 1 {1ß2 '3 1
where ß is any unit vector and {ß, ß2 , ß3 1 is an orthonormal as was to be proved.
set of vectors. From Lemma 3,
Charles R. Giardina was born in the Bronx, New York December 29, 1942. He is presently associated with The Singer
Company, Kearfott Division and with Fairleigh Dickinson University as Professor of Mathematics, Electrical Engineering
and Computer Sceince. His current research interests are Stochastic Functional Analysis and Stochastic Integral Equations.
He resides in Mahwah, N.J. with his wife Betty and their three children; Barbara Ann, Steven and Mikey.
Richard D. Bronson was born in New York City on August 5, 1941. He received the B.S., M.S., and PhD. degrees in
mathematics from Stevens Institute of Technology in 1963, 1965, and 1968, respectively.
He joined the faculty at Fairleigh Dickinson University, Teaneck, N.J. where he is currently Chairman, Department of
Mathematics and Computer Science. He is the author of MATRIX METHODS (Academic Press) and MODERN INTRO-
DUCTORY DIFFERENTIAL EQUATIONS (Schaum's Outline Series, McGraw-Hill).
Dr. Bronson is a member of Sigma Xi, the American Mathematical Society, the Mathematical Association of America,
and the Association for Computing Machinery.
Lawrence J. Wallen was born in 1933 in Philadelphia, PA. He received his Ph.D. from M.I.T. in 1967. Has been on the
faculties of Stevens Institute of Technology, University of Michigan, and is at present Professor of Mathematics at the
University of Hawaii. His main research is in Operator Theory and Semi-groups of Operators.
446 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS JULY 1975
Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on April 04,2020 at 05:16:26 UTC from IEEE Xplore. Restrictions apply.