Research Article Riemannian Gradient Algorithm...
Transcript of Research Article Riemannian Gradient Algorithm...
Research ArticleRiemannian Gradient Algorithm for the NumericalSolution of Linear Matrix Equations
Xiaomin Duan12 Huafei Sun1 and Xinyu Zhao34
1 School of Mathematics Beijing Institute of Technology Beijing 100081 China2 School of Science Dalian Jiaotong University Dalian 116028 China3 School of Mechanical Engineering Beijing Institute of Technology Beijing 100081 China4 School of Materials Science and Engineering Dalian Jiaotong University Dalian 116028 China
Correspondence should be addressed to Huafei Sun huafeisunbiteducn
Received 6 August 2013 Revised 10 December 2013 Accepted 10 December 2013 Published 6 January 2014
Academic Editor Zhi-Hong Guan
Copyright copy 2014 Xiaomin Duan et alThis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
ARiemannian gradient algorithmbased on geometric structures of amanifold consisting of all positive definitematrices is proposedto calculate the numerical solution of the linear matrix equation 119876 = 119883 + sum
119898
119894=1119860119879
119894119883119860119894 In this algorithm the geodesic distance on
the curved Riemannian manifold is taken as an objective function and the geodesic curve is treated as the convergence path Alsothe optimal variable step sizes corresponding to the minimum value of the objective function are provided in order to improvethe convergence speed Furthermore the convergence speed of the Riemannian gradient algorithm is compared with that of thetraditional conjugate gradientmethod in two simulation examples It is found that the convergence speed of the provided algorithmis faster than that of the conjugate gradient method
1 Introduction
The linear matrix equation
119876 = 119883 +
119898
sum
119894=1
119860119879
119894119883119860119894 (1)
where 1198601 1198602 119860
119898are arbitrary 119899 times 119899 real matrices
119898 is a nonnegative integer and 119860119879
119894denotes the transpose
of the matrix 119860119894 arises from many fields such as the
control theory the dynamic programming and the stochasticfiltering [1ndash4] In the past decades there has been increasinginterest in the solution problems of this equation In thecase of 119898 = 1 some numerical methods such as Bartels-Stewart method Hessenberg-Schur method and Schur andQR decompositions method were proposed in [5 6] Basedon the Kronecker product and the fixed point theorem inpartially ordered sets some sufficient conditions for theexistence of a unique symmetric positive definite solutionare given in [7 8] Ran and Reurings ([7 Theorem 33] and
[9 Theorem 31]) pointed out that if 119876 minus sum119898
119894=1119860119879
119894119876119860119894is a
positive definite matrix then there exists a unique solutionand it is symmetric positive definite Recently under thecondition that (1) is consistent Su and Chen presented anefficient numerical iterative method based on the conjugategradient method (CGM) [10]
In addition based on geometric structures on a Rie-mannian manifold Duan et al proposed a natural gradientdescent algorithm to solve algebraic Lyapunov equations [1112] Following them we investigate the solution problemof (1) in the view of Riemannian manifolds Note that thissolution of (1) is a symmetric positive definite matrix andthe set of all the symmetric positive definite matrices can beconsidered as a manifold Thus it is more convenient toinvestigate the solution problem with the help of these geo-metric structures on this manifold To address such a needa new framework is presented in this paper to calculate thenumerical solution which is based on the geometric struc-tures on the Riemannian manifold of positive definitesymmetric matrices We will first describe briefly some
Hindawi Publishing CorporationJournal of Applied MathematicsVolume 2014 Article ID 507175 7 pageshttpdxdoiorg1011552014507175
2 Journal of Applied Mathematics
fundamental knowledge on manifold Then a Riemanniangradient algorithm is proposed where the geodesic distanceon the curved Riemannian manifold is taken as an objectivefunction and the geodesic curve is treated as the convergencepath Also the optimal variable step sizes corresponding tothe minimum value of the objective function are providedin order to improve the convergence speed Finally thebehavior of the provided algorithm and the traditional CGMis compared and demonstrated in two simulation examples
2 A Survey of Some Geometrical Concepts
In the present section we briefly survey the geometry ofsmoothmanifolds by recalling concepts as tangent spaces theRiemannian gradient geodesic curves and geodesic distanceMore details can be found in [13 14] Specially these conceptsof manifold consisting of all symmetric positive definitematrices have also been introduced in [15ndash20]
21 Riemannian Manifolds Let us denote a Riemannianmanifold byM Its tangent space at point119884 isin M is denoted by119879119884M Given any pair of points 119881 119882 isin 119879
119884M an inner product
⟨119881 119882⟩119884
isin R is definedThe specification of the inner productfor a Riemannian manifold turns it into a metric space Infact the length of a curve 120601
119884119881 [0 1] rarr M such that
120601119884119881
(0) = 119884 isin M and 120601119884119881
(0) = 119881 isin 119879119884M is given by
ℓ (120601119884119881
) = int
1
0
radic⟨ 120601119884119881
(119905) 120601119884119881
(119905)⟩120601119884119881(119905)
d119905 (2)
Given arbitrary 119884 isin M and 119881 isin 119879119884M the curve 120574
119884119881
[0 1] rarr M of shortest length is called geodesic curve Suchminimal length ℓ(120574
119884119881) is called geodesic distance between
two points namely 120574119884119881
(0) = 119884 and 120574119884119881
(1) The Riemanniandistance between two points is denoted by
d (120574119884119881
(0) 120574119884119881
(1)) = ℓ (120574119884119881
) (3)
It is obvious that if any pair of points on manifold M can beconnected by a geodesic curve then it is possible to measurethe distance between any given pair of points on it Given aregular function 119891 M rarr R its Riemannian gradient nabla
119884119891
in the direction of the vector 119881 isin 119879119884M measures the rate of
change of the function119891 in the direction119881 Namely given anysmooth curve 120601
119884119881 [0 1] rarr M such that 120601
119884119881(0) = 119884 isin M
and 120601119884119881
(0) = 119881 the Riemannian gradient nabla119884119891 is the unique
vector in 119879119884M such that
⟨119881 nabla119884119891⟩119884
=dd119905
119891 (120601119884119881
(119905))1003816100381610038161003816119905=0
(4)
Instead of the Euclidean space by using the Riemanniangradient the optimization method may be readily extendedto smooth manifolds [21 Chapter 7] For this purpose let usconsider the differential equation on manifoldM
119910 = minusnabla119910(119905)
119891 119910 (0) = 119910 isin M (5)
If the function 119891 M rarr R is bounded and the smoothmanifoldM is compact then the solution of such differential
equation tends to a local minimum of the function 119891 in Mdepending on the initial point 119910 In fact it is achieved bydefinition of the Riemannian gradient
dd119905
119891 (119910 (119905)) = ⟨ 119910 (119905) nabla119910(119905)
119891⟩119910(119905)
= minus⟨nabla119910(119905)
119891 nabla119910(119905)
119891⟩119910(119905)
le 0
(6)
22 Manifold of Symmetric Positive Definite Matrices Let usdenote the general linear group by GL(119899) which consists ofall nonsingular real matrices And 119883 gt 0 means that 119883 is asymmetric positive definite matrix For a different notation119883 minus 119884 gt 0 we will use 119883 gt 119884 And the manifold consisting ofall symmetric positive definite matrices is defined by S+(119899) =
119883 isin R119899times119899 | 119883119879
= 119883 119883 gt 0 The tangent space at a point119883 isin S+(119899) is given by 119879
119883S+(119899) = 119881 isin R119899times119899 | 119881
119879
= 119881On manifold S+(119899) the Riemannian metric at the point 119883 isdefined by
⟨1198811 1198812⟩119883
= tr [119883minus1
1198811119883minus1
1198812] (7)
where 1198811 1198812
isin 119879119883S+(119899) and tr denotes the trace of the
matrix Then manifold S+(119899) with the Riemannian metric(7) becomes a Riemannian manifold Moreover it is a curvedmanifold since thismetric is invariant under the group actionof GL(119899) The geodesic curve at the point 119883 isin S+(119899) in thedirection 119881 isin 119879
119883S+(119899) can be expressed as
120574119883119881
(119905) = 11988312 exp (119905119883
minus12
119881119883minus12
) 11988312
(8)
Obviously this geodesic curve is entirely contained in man-ifold The Hopf-Rinow theorem implies that manifold S+(119899)
with the Riemannianmetric (7) is geodesically complete ([22Theorem 28]) This means for any given pair 119883 119884 isin S+(119899)we can find a geodesic curve 120574
119883119881(119905) such that 120574
119883119881(0) = 119883
and 120574119883119881
(1) = 119884 by taking the initial velocity 120574119883119881
(0) =
11988312 log(119883
minus12
119884119883minus12
)11988312
isin 119879119883S+(119899) According to (2) the
geodesic distance d(119883 119884) can be computed explicitly by
d2 (119883 119884) = tr [log2 (119883minus12
119884119883minus12
)] (9)
3 Riemannian Gradient Algorithm
Some researches show that the solution equation (1) is uniquepositive definite if 119860
119894and 119876 satisfy ([7 Theorem 33] and [9
Theorem 31])
119876 gt
119898
sum
119894=1
119860119879
119894119876119860119894 (10)
For convenience we set L(119883) = 119883 + sum119898
119894=1119860119879
119894119883119860119894 And it
is obvious that L(119883) is positive definite for any 119883 isin S+(119899)Our purpose is to seek a matrix 119883 isin S+(119899) so that L(119883) isas close as possible to the given matrix 119876 We figure out somekey points in designing such an algorithm as follows
(1) For some 119883 the difference between the given positivedefinite matrix 119876 and the matrix L(119883) should bedetermined and calculated firstly
Journal of Applied Mathematics 3
(2) Then the Riemannian gradient descent algorithm canbe used to adjust 119883 so that the difference will be assmall as possible
Note that the Riemannian manifold S+(119899) is geodesicallycomplete hence we take the geodesic distance between119876 andL(119883) as the objective function 119869(119883) namely
119869 (119883) = d2 (119876L (119883)) (11)
Then the exact solution of (1) may be defined as
119883lowast
= arg min119883isinS+(119899)
119869 (119883) (12)
In order to find aminimizer for the criterion d2(119876L(119883)) wemay make use of the differential equation (5) The minimumis apparently achieved for a value 119883
lowastsuch that
nabla119883
119869 (119883)1003816100381610038161003816119883=119883lowast
= 0 (13)
It is necessary to solve the optimization problem (12) by anumerical optimization algorithm that takes the geometry ofthe Riemannian manifold S+(119899) into account In particularstarting from an initial guess 119883
0 it is possible to provide an
iterative learning algorithm that generates a pattern 119883119905
isin
S+(119899) at any learning step 119905 isin N Following [21 23] suchsequence may be generated by moving from each point 119883
119905isin
S+(119899) to the next point 119883119905+1
along a short geodesic curvein the opposite direction of the Riemannian gradient of theobjective function 119869(119883
119905) Namely if we set
119881119905
= nabla119883119905
119869 (119883119905) (14)
then the Riemannian gradient algorithmmay be expressed as
119883119905+1
= 120574119883119905minus119881119905
(120578) (15)
where 120578 denotes any suitable step-size schedule that drives theiterative algorithm (15) to be convergent Moreover with theaim to employ the learning algorithm (14) and (15) we firstneed to compute the Riemannian gradient of the objectivefunction 119869(119883
119905) to be optimized In the definition of the
Riemannian gradient (4) the generic smooth curve 120601119884119881
maybe replaced with a geodesic curve We obtain the followingtheorem
Theorem 1 TheRiemannian gradient of the objective function119869(119883) at point 119883 is given by
nabla119883
119869 (119883) = 2119883 ( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119883
(16)
Proof First we compute the derivative of the real-valuedfunction
119869 (119904 (119905)) = tr [log2 (119876minus12
L (119904 (119905)) 119876minus12
)] (17)
with respect to 119905 where 119904(119905) = 11988312 exp(119905119883
minus12
119881119883minus12
)11988312
is the geodesic curve emanating from 119883 in the direction 119881 =
119904(0) Using Proposition 21 in [16] and some properties of thetrace we have
dd119905
119869 (119904 (119905))|119905=0
= 2 tr[
[
( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119881]
]
= ⟨119881 2119883 ( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119883⟩
119883
(18)
From (4) it is shown that formula (16) is valid
Remark 2 In fact note that log(119860minus1
119861119860) = 119860minus1
(log119861)119860 forall 119860 isin GL(119899) and for all 119861 isin S+(119899) then nabla
119883119869(119883) can be
rewritten as
nabla119883
119869 (119883)
= 2119883 (L(119883)minus12 log (L(119883)
12
119876minus1
L(119883)12
)L(119883)minus12
+
119898
sum
119895=1
119860119895L(119883)
minus12
times log (L(119883)12
119876minus1
L(119883)12
)L(119883)minus12
119860119879
119895) 119883
(19)
Therefore it can be shown that nabla119883
119869(119883) belongs to 119879119883S+(119899)
Corollary 3 TheRiemannian gradientnabla119883
119869(119883) has the uniquezero point 119883
lowast Furthermore for the suitable step-size 120578 the
sequence 119883119905generated by (15) converges to the solution 119883
lowastof
(1)
Proof Obviously 119883lowastis a zero point of the Riemannian gra-
dient nabla119883
119869(119883) If there exists 1198831015840
lowast= 119883lowastsuch that the iterative
process (15) stops then it means nabla1198831015840
lowast
119869(1198831015840
lowast) = 0 which is
equivalent to
log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
119860119879
119895= 0
(20)
4 Journal of Applied Mathematics
On (20) the matrices are multiplied by 1198831015840
lowastand the traces of
the matrices are taken then we have
tr [log (119876minus1
L (1198831015840
lowast))] = 0 (21)
Also if we replace 1198831015840
lowastwith 119883
lowastand follow the similar
procedure as the above equation then the following equalitiesare valid
tr[log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
(119883lowast
+
119898
sum
119894=1
119860119879
119894119883lowast119860119894)]
= tr [log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
119876]
= 0
(22)
Combining (21) and (22) it can be obtained that
tr [log (119876minus1
L (1198831015840
lowast))] minus tr [log (119876
minus1
L (1198831015840
lowast)) 119876]
= tr [log (119876minus1
L (1198831015840
lowast)) (119868 minus L(119883
1015840
lowast)minus1
119876)]
= tr [log (119876minus12
L (1198831015840
lowast) 119876minus12
)
times (119868 minus 11987612
L(1198831015840
lowast)minus1
11987612
)]
= 0
(23)
Note that the matrix 11987612L(119883
1015840
lowast)minus1
11987612 is positive defi-
nite hence its orthogonal similarity diagonalization can beexpressed as 119880Λ119880
119879 where 119880 is an orthogonal matrix Λ =
diag (1205821 120582
119899) and 120582
119894gt 0 (1 le 119894 le 119899) Thus we have
tr [(Λ minus 119868) logΛ] =
119899
sum
119894=1
(120582119894minus 1) log 120582
119894
= 0 (24)
Note that the second equality of (24) is equivalent to
119899
prod
119894=1
120582119894
120582119894minus1 = 1 (25)
Thus we have 1205821
= 1205822
= sdot sdot sdot = 120582119899
= 1 That means that 1198831015840
lowastis
also the solution of (1) which is contrary to the uniquenessof the solution Therefore the Riemannian gradient nabla
119883119869(119883)
has the unique zero point 119883lowast This completes the proof of
Corollary 3
Now the Riemannian gradient algorithm with a constantstep-size (RGACS) becomes
119883119905+1
= 11988312
119905exp (minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905) 11988312
119905 (26)
where the initial value 1198830
isin S+(119899) The step-size 120578 corres-ponding to the best convergence speed can be obtained exper-imentally and the details will be described in our followingsimulations
31 Riemannian Gradient Algorithm with Variable Step SizesIn order to improve the convergence speed of the RGACSan optimal variable step-size schedule 120578
119905will be considered
and defined as follows The step-size 120578119905should be evaluated
in such a way that the objective function 119869(119883119905+1
) at step 119905 + 1
is reduced as much as possible with respect to 119869(119883119905) We may
regard 119869(119883119905+1
) as a function of the step-size 120578119905that
119869 (119883119905+1
) = d2 (119876L (120574119883119905minus119881119905
(120578119905))) (27)
Thus we can optimize the value of 120578119905to ensure that the
difference 119869(119883119905) minus 119869(119883
119905+1) could be as large as possible Since
it is difficult to solve this nonlinear problem exactly we use asuboptimal approximation in practice Under the hypothesisthat the step-size value 120578
119905is small enough we may invoke the
expansion of the function 119869(119883119905+1
) around the point 120578119905
= 0 asfollows
119869 (119883119905+1
) asymp 119869 (119883119905) + 1198621119905
120578119905+
1
21198622119905
1205782
119905 (28)
Then the optimal step-size 120578lowast
119905 corresponding to the maxi-
mum of the difference 119869(119883119905) minus 119869(119883
119905+1) can be expressed as
120578lowast
119905asymp minus
1198621119905
1198622119905
(29)
The coefficients 1198621119905 and 119862
2119905can be calculated using the
first-order and the second-order derivatives of the function119869(119883119905+1
) with respect to the parameter 120578119905at the point 120578
119905= 0
On the basis of the above discussion we may obtain theoptimal step-size
120578lowast
119905asymp
tr [(119883L (log (119876minus1L (119883
119905))L(119883
119905)minus1
))2
]
tr [(L(119883119905)minus1
L (nabla119883119905
119869 (119883119905)))2
]
(30)
Then the Riemannian gradient algorithm with variablestep-sizes (RGAVS) can be written explicitly as
119883119905+1
= 11988312
119905exp (minus120578
lowast
119905119883minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905) 11988312
119905 (31)
where 1198830
isin S+(119899) is the initial value
32 Volume Feature of Riemannian Gradient Algorithm TheRGACS shows an interesting volume feature that the ratiodet(119883
119905+1) det(119883
119905) always keeps a certain relationship In
order to prove the volume feature let us recall two propertiesof determinant operator det namely for each 119860 119861 isin GL(119899)there hold
det (119860119861) = det (119860) det (119861)
det (exp (119860)) = exp (tr (119860)) (32)
Journal of Applied Mathematics 5
By applying the determinant operator to both sides of (31)we obtain
det (119883119905+1
)
= det (119883119905) exp (tr [minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905])
= det (119883119905) exp(tr[
[
120578 ( log (L(119883119905)minus1
119876)L(119883119905)minus1
119883119905
+ log (L(119883119905)minus1
119876)L(119883119905)minus1
times
119898
sum
119895=1
119860119879
119895119883119905119860119895)]
]
)
= det (119883119905) exp (tr [120578 log (L(119883
119905)minus1
119876)])
= det (119883119905) det ((L(119883
119905)minus1
119876)120578
)
(33)
Therefore the following property is valid
det (119883119905+1
)
det (119883119905)
= (det (119876)
det(L (119883119905))
120578
(34)
This equation shows that the determinant of (1) isclosely linked with that of the sequence 119883
119905 Moreover
the ratio det(119883119905+1
) det(119883119905) is directly proportional to
det(119876) det(L(119883119905)) for the fixed step-size 120578 In the case of
the RGAVS the similar conclusion can also be obtained
4 Simulations
Two simulation examples are given to compare the con-vergence speed of the Riemannian gradient algorithm (theRGACS and the RGAVS) with those of the traditional CGMThe error criterion used in these simulations is120588(L(119883)minus119876) lt
10minus15 where 120588(sdot) denotes the spectral radius of the matrix
Example 4 First we consider the matrices equation
119876 = 119883 + 119860119879
11198831198601
+ 119860119879
21198831198602
+ 119860119879
31198831198603
(35)
in the case of 119899 = 5 And 119876 1198601 1198602 1198603are as follows
119876 = (
46350 00698 03104 01990 02918
00698 40376 00742 00694 00690
03104 00742 42242 01554 01826
01990 00694 01554 41472 01572
02918 00690 01826 01572 42134
)
1198601
= (
00834 00007 00293 00142 00139
00141 02975 00154 00289 00339
00244 00152 00317 00270 00117
00251 00097 00140 00135 00240
00158 01499 00275 00077 00157
)
02 03 04 06 07 08050
20
40
60
80
100
120
140
160
Step size
Itera
tions
Figure 1 Step sizes versus the iterations in the RGACS
1198602
= (
02624 00166 00454 00172 00160
00893 00608 00242 00029 00061
00054 00190 00589 00036 00160
00278 00050 00255 00689 00048
00319 00064 00103 00219 00018
)
1198603
= (
81622 06020 04505 08258 01067
27943 70630 00838 05383 09619
03112 06541 62290 09961 00046
05285 06892 09133 10782 07749
01656 07482 01524 04427 90073
)
(36)
which satisfy the condition (10) Both the RGACS andRGAVS are used and compared with the traditional CGMand 119876 is taken as the initial value 119883
0in the following
simulation In the RGACS case the step sizes versus theiterations are shown in Figure 1 It can be found that theiterations will reduce gradually as the step-size changes fromzero to a critical value then the iterations will increasegradually above this critical valueTherefore the critical valueabout 064 can be selected experimentally as an optimal step-size and used in the RGACS
Figure 2 shows the convergence comparison between theRGACS and the RGAVS and the traditional CGM It can befound that the RGAVS possesses themost steady convergenceand the fastest convergence speed and it needs 23 iterationsto obtain the numerical solution of (35) as follows
(
00856 minus00265 00151 minus01065 00148
minus00265 00970 00166 minus01581 minus00023
00151 00166 01591 minus03608 00283
minus01065 minus01581 minus03608 23761 minus01723
00148 minus00023 00283 minus01723 00654
)
(37)
The case of the CGM realizes the given error through 33iterations and the slowest one among them is the RGACSwith 31 steps By comparison it is shown that the convergence
6 Journal of Applied Mathematics
5 10 15 20 25 30 35Iteration
Erro
r
RGACSRGAVS
CGM
105
100
10minus5
10minus10
Figure 2 Comparison of convergence speeds (log scale)
speed of the RGAVS is faster than both those of the RGACSand the traditional CGM in solving (35)
Example 5 A 20-order linear matrix equation with 119898 =
2 is also considered here following the similar procedureas Example 4 where 119876 isin S+(20) and 119860
1 1198602
isin R20times20
satisfy the condition (10) The simulation indicates that thenumerical solution needs 26 45 and 37 iterations in theRGAVS the RGACS (120578 = 074) and the traditional CGMrespectively Therefore the convergence speed of the RGAVSis also better than those of the RGACS and the CGM
5 Conclusion
Using geometric structures of manifold consisting of all sym-metric positive definite matrices the Riemannian gradientalgorithm is developed to calculate the numerical solutionfor the linear matrix equation 119876 = 119883 + sum
119898
119894=1119860119879
119894119883119860119894 In this
algorithm the geodesic distance on the curved Riemannianmanifold is taken as an objective function and the geodesiccurve is treated as the convergence path Also the variablestep sizes algorithm corresponding to the minimum valueof the objective function is provided in order to improvethe convergence speed in the constant step-size case Finallythe convergence speed of the Riemannian gradient algorithmhas been compared with the traditional conjugate gradientmethod by two simulation examples in 5-order and 20-orderlinear matrix equations respectively It is shown that theconvergence speed of the Riemannian gradient algorithmwith variable step sizes is faster than that of the conjugate gra-dient method Although the Riemannian gradient algorithmis used here for the linear matrix equation in the real numberfield it may be generalized to the complex case as well
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to express their sincere thanks tothe referees for their valuable suggestions which greatlyimproved the presentation of this paper This subject issupported by the National Natural Science Foundations ofChina (nos 61179031 10932002 and 51105033)
References
[1] M Berzig ldquoSolving a class of matrix equations via the Bhaskar-Lakshmikantham coupled fixed point theoremrdquo Applied Math-ematics Letters vol 25 no 11 pp 1638ndash1643 2012
[2] J C Engwerda ldquoOn the existence of a positive definite solutionof the matrix equation 119883 + 119860
119879
119883minus1
119860 = 119868rdquo Linear Algebra and ItsApplications vol 194 pp 91ndash108 1993
[3] W L Green and E W Kamen ldquoStabilizability of linear sys-tems over a commutative normed algebra with applications tospatially-distributed and parameter-dependent systemsrdquo SIAMJournal onControl andOptimization vol 23 no 1 pp 1ndash18 1985
[4] W Pusz and S L Woronowicz ldquoFunctional calculus forsesquilinear forms and the purification maprdquo Reports on Math-ematical Physics vol 8 no 2 pp 159ndash170 1975
[5] C-Y Chiang E K-W Chu and W-W Lin ldquoOn the 995333-Sylvester equation 119860119883 plusmn 119883
995333
119861995333
= 119862rdquo Applied Mathematics andComputation vol 218 no 17 pp 8393ndash8407 2012
[6] Z Gajic and M T J Qureshi Lyapunov Matrix Equation inSystem Stability and Control vol 195 Academic Press LondonUK 1995
[7] A C M Ran and M C B Reurings ldquoThe symmetric linearmatrix equationrdquo Electronic Journal of Linear Algebra vol 9 pp93ndash107 2002
[8] M C B Reurings Symmetric matrix equations [PhD thesis]Vrije Universiteit Amsterdam The Netherlands 2003
[9] A C M Ran and M C B Reurings ldquoA fixed point theoremin partially ordered sets and some applications to matrixequationsrdquo Proceedings of the American Mathematical Societyvol 132 no 5 pp 1435ndash1443 2004
[10] Y Su and G Chen ldquoIterative methods for solving linear matrixequation and linear matrix systemrdquo International Journal ofComputer Mathematics vol 87 no 4 pp 763ndash774 2010
[11] X Duan H Sun L Peng and X Zhao ldquoA natural gradi-ent descent algorithm for the solution of discrete algebraicLyapunov equations based on the geodesic distancerdquo AppliedMathematics and Computation vol 219 no 19 pp 9899ndash99052013
[12] X Duan H Sun and Z Zhang ldquoA natural gradient descentalgorithm for the solution of Lyapunov equations based on thegeodesic distancerdquo Journal of Computational Mathematics vol219 no 19 pp 9899ndash9905 2013
[13] M Spivak A Comprehensive Introduction to Differential Geom-etry vol 1 Publish or Perish Berkeley Calif USA 3rd edition1999
[14] S Lang Fundamentals of Differential Geometry vol 191Springer 1999
Journal of Applied Mathematics 7
[15] F Barbaresco ldquoInteractions between symmetric cones andinformation geometrics Bruhat-tits and Siegel spaces modelsfor high resolution autoregressive Doppler imageryrdquo in Emerg-ing Trends in Visual Computing vol 5416 of Lecture Notes inComputer Science pp 124ndash163 2009
[16] M Moakher ldquoA differential geometric approach to the geo-metric mean of symmetric positive-definite matricesrdquo SIAMJournal on Matrix Analysis and Applications vol 26 no 3 pp735ndash747 2005
[17] M Moakher ldquoOn the averaging of symmetric positive-definitetensorsrdquo Journal of Elasticity vol 82 no 3 pp 273ndash296 2006
[18] A Schwartzman Random ellipsoids and false discovery ratesstatistics for diffusion tensor imaging data [PhD thesis] StanfordUniversity 2006
[19] J D Lawson and Y Lim ldquoThe geometric mean matricesmetrics and morerdquo The American Mathematical Monthly vol108 no 9 pp 797ndash812 2001
[20] R Bhatia Positive definite Matrices Princeton Series in AppliedMathematics Princeton University Press Princeton NJ USA2007
[21] C Udriste Convex Functions and Optimization Methods onRiemannian Manifolds vol 297 ofMathematics and its Applica-tions Kluwer Academic Publishers Dordrecht Germany 1994
[22] M P D Carmo Riemannian Geometry Springer 1992[23] D G Luenberger ldquoThe gradient projection method along
geodesicsrdquoManagement Science vol 18 pp 620ndash631 1972
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
2 Journal of Applied Mathematics
fundamental knowledge on manifold Then a Riemanniangradient algorithm is proposed where the geodesic distanceon the curved Riemannian manifold is taken as an objectivefunction and the geodesic curve is treated as the convergencepath Also the optimal variable step sizes corresponding tothe minimum value of the objective function are providedin order to improve the convergence speed Finally thebehavior of the provided algorithm and the traditional CGMis compared and demonstrated in two simulation examples
2 A Survey of Some Geometrical Concepts
In the present section we briefly survey the geometry ofsmoothmanifolds by recalling concepts as tangent spaces theRiemannian gradient geodesic curves and geodesic distanceMore details can be found in [13 14] Specially these conceptsof manifold consisting of all symmetric positive definitematrices have also been introduced in [15ndash20]
21 Riemannian Manifolds Let us denote a Riemannianmanifold byM Its tangent space at point119884 isin M is denoted by119879119884M Given any pair of points 119881 119882 isin 119879
119884M an inner product
⟨119881 119882⟩119884
isin R is definedThe specification of the inner productfor a Riemannian manifold turns it into a metric space Infact the length of a curve 120601
119884119881 [0 1] rarr M such that
120601119884119881
(0) = 119884 isin M and 120601119884119881
(0) = 119881 isin 119879119884M is given by
ℓ (120601119884119881
) = int
1
0
radic⟨ 120601119884119881
(119905) 120601119884119881
(119905)⟩120601119884119881(119905)
d119905 (2)
Given arbitrary 119884 isin M and 119881 isin 119879119884M the curve 120574
119884119881
[0 1] rarr M of shortest length is called geodesic curve Suchminimal length ℓ(120574
119884119881) is called geodesic distance between
two points namely 120574119884119881
(0) = 119884 and 120574119884119881
(1) The Riemanniandistance between two points is denoted by
d (120574119884119881
(0) 120574119884119881
(1)) = ℓ (120574119884119881
) (3)
It is obvious that if any pair of points on manifold M can beconnected by a geodesic curve then it is possible to measurethe distance between any given pair of points on it Given aregular function 119891 M rarr R its Riemannian gradient nabla
119884119891
in the direction of the vector 119881 isin 119879119884M measures the rate of
change of the function119891 in the direction119881 Namely given anysmooth curve 120601
119884119881 [0 1] rarr M such that 120601
119884119881(0) = 119884 isin M
and 120601119884119881
(0) = 119881 the Riemannian gradient nabla119884119891 is the unique
vector in 119879119884M such that
⟨119881 nabla119884119891⟩119884
=dd119905
119891 (120601119884119881
(119905))1003816100381610038161003816119905=0
(4)
Instead of the Euclidean space by using the Riemanniangradient the optimization method may be readily extendedto smooth manifolds [21 Chapter 7] For this purpose let usconsider the differential equation on manifoldM
119910 = minusnabla119910(119905)
119891 119910 (0) = 119910 isin M (5)
If the function 119891 M rarr R is bounded and the smoothmanifoldM is compact then the solution of such differential
equation tends to a local minimum of the function 119891 in Mdepending on the initial point 119910 In fact it is achieved bydefinition of the Riemannian gradient
dd119905
119891 (119910 (119905)) = ⟨ 119910 (119905) nabla119910(119905)
119891⟩119910(119905)
= minus⟨nabla119910(119905)
119891 nabla119910(119905)
119891⟩119910(119905)
le 0
(6)
22 Manifold of Symmetric Positive Definite Matrices Let usdenote the general linear group by GL(119899) which consists ofall nonsingular real matrices And 119883 gt 0 means that 119883 is asymmetric positive definite matrix For a different notation119883 minus 119884 gt 0 we will use 119883 gt 119884 And the manifold consisting ofall symmetric positive definite matrices is defined by S+(119899) =
119883 isin R119899times119899 | 119883119879
= 119883 119883 gt 0 The tangent space at a point119883 isin S+(119899) is given by 119879
119883S+(119899) = 119881 isin R119899times119899 | 119881
119879
= 119881On manifold S+(119899) the Riemannian metric at the point 119883 isdefined by
⟨1198811 1198812⟩119883
= tr [119883minus1
1198811119883minus1
1198812] (7)
where 1198811 1198812
isin 119879119883S+(119899) and tr denotes the trace of the
matrix Then manifold S+(119899) with the Riemannian metric(7) becomes a Riemannian manifold Moreover it is a curvedmanifold since thismetric is invariant under the group actionof GL(119899) The geodesic curve at the point 119883 isin S+(119899) in thedirection 119881 isin 119879
119883S+(119899) can be expressed as
120574119883119881
(119905) = 11988312 exp (119905119883
minus12
119881119883minus12
) 11988312
(8)
Obviously this geodesic curve is entirely contained in man-ifold The Hopf-Rinow theorem implies that manifold S+(119899)
with the Riemannianmetric (7) is geodesically complete ([22Theorem 28]) This means for any given pair 119883 119884 isin S+(119899)we can find a geodesic curve 120574
119883119881(119905) such that 120574
119883119881(0) = 119883
and 120574119883119881
(1) = 119884 by taking the initial velocity 120574119883119881
(0) =
11988312 log(119883
minus12
119884119883minus12
)11988312
isin 119879119883S+(119899) According to (2) the
geodesic distance d(119883 119884) can be computed explicitly by
d2 (119883 119884) = tr [log2 (119883minus12
119884119883minus12
)] (9)
3 Riemannian Gradient Algorithm
Some researches show that the solution equation (1) is uniquepositive definite if 119860
119894and 119876 satisfy ([7 Theorem 33] and [9
Theorem 31])
119876 gt
119898
sum
119894=1
119860119879
119894119876119860119894 (10)
For convenience we set L(119883) = 119883 + sum119898
119894=1119860119879
119894119883119860119894 And it
is obvious that L(119883) is positive definite for any 119883 isin S+(119899)Our purpose is to seek a matrix 119883 isin S+(119899) so that L(119883) isas close as possible to the given matrix 119876 We figure out somekey points in designing such an algorithm as follows
(1) For some 119883 the difference between the given positivedefinite matrix 119876 and the matrix L(119883) should bedetermined and calculated firstly
Journal of Applied Mathematics 3
(2) Then the Riemannian gradient descent algorithm canbe used to adjust 119883 so that the difference will be assmall as possible
Note that the Riemannian manifold S+(119899) is geodesicallycomplete hence we take the geodesic distance between119876 andL(119883) as the objective function 119869(119883) namely
119869 (119883) = d2 (119876L (119883)) (11)
Then the exact solution of (1) may be defined as
119883lowast
= arg min119883isinS+(119899)
119869 (119883) (12)
In order to find aminimizer for the criterion d2(119876L(119883)) wemay make use of the differential equation (5) The minimumis apparently achieved for a value 119883
lowastsuch that
nabla119883
119869 (119883)1003816100381610038161003816119883=119883lowast
= 0 (13)
It is necessary to solve the optimization problem (12) by anumerical optimization algorithm that takes the geometry ofthe Riemannian manifold S+(119899) into account In particularstarting from an initial guess 119883
0 it is possible to provide an
iterative learning algorithm that generates a pattern 119883119905
isin
S+(119899) at any learning step 119905 isin N Following [21 23] suchsequence may be generated by moving from each point 119883
119905isin
S+(119899) to the next point 119883119905+1
along a short geodesic curvein the opposite direction of the Riemannian gradient of theobjective function 119869(119883
119905) Namely if we set
119881119905
= nabla119883119905
119869 (119883119905) (14)
then the Riemannian gradient algorithmmay be expressed as
119883119905+1
= 120574119883119905minus119881119905
(120578) (15)
where 120578 denotes any suitable step-size schedule that drives theiterative algorithm (15) to be convergent Moreover with theaim to employ the learning algorithm (14) and (15) we firstneed to compute the Riemannian gradient of the objectivefunction 119869(119883
119905) to be optimized In the definition of the
Riemannian gradient (4) the generic smooth curve 120601119884119881
maybe replaced with a geodesic curve We obtain the followingtheorem
Theorem 1 TheRiemannian gradient of the objective function119869(119883) at point 119883 is given by
nabla119883
119869 (119883) = 2119883 ( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119883
(16)
Proof First we compute the derivative of the real-valuedfunction
119869 (119904 (119905)) = tr [log2 (119876minus12
L (119904 (119905)) 119876minus12
)] (17)
with respect to 119905 where 119904(119905) = 11988312 exp(119905119883
minus12
119881119883minus12
)11988312
is the geodesic curve emanating from 119883 in the direction 119881 =
119904(0) Using Proposition 21 in [16] and some properties of thetrace we have
dd119905
119869 (119904 (119905))|119905=0
= 2 tr[
[
( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119881]
]
= ⟨119881 2119883 ( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119883⟩
119883
(18)
From (4) it is shown that formula (16) is valid
Remark 2 In fact note that log(119860minus1
119861119860) = 119860minus1
(log119861)119860 forall 119860 isin GL(119899) and for all 119861 isin S+(119899) then nabla
119883119869(119883) can be
rewritten as
nabla119883
119869 (119883)
= 2119883 (L(119883)minus12 log (L(119883)
12
119876minus1
L(119883)12
)L(119883)minus12
+
119898
sum
119895=1
119860119895L(119883)
minus12
times log (L(119883)12
119876minus1
L(119883)12
)L(119883)minus12
119860119879
119895) 119883
(19)
Therefore it can be shown that nabla119883
119869(119883) belongs to 119879119883S+(119899)
Corollary 3 TheRiemannian gradientnabla119883
119869(119883) has the uniquezero point 119883
lowast Furthermore for the suitable step-size 120578 the
sequence 119883119905generated by (15) converges to the solution 119883
lowastof
(1)
Proof Obviously 119883lowastis a zero point of the Riemannian gra-
dient nabla119883
119869(119883) If there exists 1198831015840
lowast= 119883lowastsuch that the iterative
process (15) stops then it means nabla1198831015840
lowast
119869(1198831015840
lowast) = 0 which is
equivalent to
log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
119860119879
119895= 0
(20)
4 Journal of Applied Mathematics
On (20) the matrices are multiplied by 1198831015840
lowastand the traces of
the matrices are taken then we have
tr [log (119876minus1
L (1198831015840
lowast))] = 0 (21)
Also if we replace 1198831015840
lowastwith 119883
lowastand follow the similar
procedure as the above equation then the following equalitiesare valid
tr[log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
(119883lowast
+
119898
sum
119894=1
119860119879
119894119883lowast119860119894)]
= tr [log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
119876]
= 0
(22)
Combining (21) and (22) it can be obtained that
tr [log (119876minus1
L (1198831015840
lowast))] minus tr [log (119876
minus1
L (1198831015840
lowast)) 119876]
= tr [log (119876minus1
L (1198831015840
lowast)) (119868 minus L(119883
1015840
lowast)minus1
119876)]
= tr [log (119876minus12
L (1198831015840
lowast) 119876minus12
)
times (119868 minus 11987612
L(1198831015840
lowast)minus1
11987612
)]
= 0
(23)
Note that the matrix 11987612L(119883
1015840
lowast)minus1
11987612 is positive defi-
nite hence its orthogonal similarity diagonalization can beexpressed as 119880Λ119880
119879 where 119880 is an orthogonal matrix Λ =
diag (1205821 120582
119899) and 120582
119894gt 0 (1 le 119894 le 119899) Thus we have
tr [(Λ minus 119868) logΛ] =
119899
sum
119894=1
(120582119894minus 1) log 120582
119894
= 0 (24)
Note that the second equality of (24) is equivalent to
119899
prod
119894=1
120582119894
120582119894minus1 = 1 (25)
Thus we have 1205821
= 1205822
= sdot sdot sdot = 120582119899
= 1 That means that 1198831015840
lowastis
also the solution of (1) which is contrary to the uniquenessof the solution Therefore the Riemannian gradient nabla
119883119869(119883)
has the unique zero point 119883lowast This completes the proof of
Corollary 3
Now the Riemannian gradient algorithm with a constantstep-size (RGACS) becomes
119883119905+1
= 11988312
119905exp (minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905) 11988312
119905 (26)
where the initial value 1198830
isin S+(119899) The step-size 120578 corres-ponding to the best convergence speed can be obtained exper-imentally and the details will be described in our followingsimulations
31 Riemannian Gradient Algorithm with Variable Step SizesIn order to improve the convergence speed of the RGACSan optimal variable step-size schedule 120578
119905will be considered
and defined as follows The step-size 120578119905should be evaluated
in such a way that the objective function 119869(119883119905+1
) at step 119905 + 1
is reduced as much as possible with respect to 119869(119883119905) We may
regard 119869(119883119905+1
) as a function of the step-size 120578119905that
119869 (119883119905+1
) = d2 (119876L (120574119883119905minus119881119905
(120578119905))) (27)
Thus we can optimize the value of 120578119905to ensure that the
difference 119869(119883119905) minus 119869(119883
119905+1) could be as large as possible Since
it is difficult to solve this nonlinear problem exactly we use asuboptimal approximation in practice Under the hypothesisthat the step-size value 120578
119905is small enough we may invoke the
expansion of the function 119869(119883119905+1
) around the point 120578119905
= 0 asfollows
119869 (119883119905+1
) asymp 119869 (119883119905) + 1198621119905
120578119905+
1
21198622119905
1205782
119905 (28)
Then the optimal step-size 120578lowast
119905 corresponding to the maxi-
mum of the difference 119869(119883119905) minus 119869(119883
119905+1) can be expressed as
120578lowast
119905asymp minus
1198621119905
1198622119905
(29)
The coefficients 1198621119905 and 119862
2119905can be calculated using the
first-order and the second-order derivatives of the function119869(119883119905+1
) with respect to the parameter 120578119905at the point 120578
119905= 0
On the basis of the above discussion we may obtain theoptimal step-size
120578lowast
119905asymp
tr [(119883L (log (119876minus1L (119883
119905))L(119883
119905)minus1
))2
]
tr [(L(119883119905)minus1
L (nabla119883119905
119869 (119883119905)))2
]
(30)
Then the Riemannian gradient algorithm with variablestep-sizes (RGAVS) can be written explicitly as
119883119905+1
= 11988312
119905exp (minus120578
lowast
119905119883minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905) 11988312
119905 (31)
where 1198830
isin S+(119899) is the initial value
32 Volume Feature of Riemannian Gradient Algorithm TheRGACS shows an interesting volume feature that the ratiodet(119883
119905+1) det(119883
119905) always keeps a certain relationship In
order to prove the volume feature let us recall two propertiesof determinant operator det namely for each 119860 119861 isin GL(119899)there hold
det (119860119861) = det (119860) det (119861)
det (exp (119860)) = exp (tr (119860)) (32)
Journal of Applied Mathematics 5
By applying the determinant operator to both sides of (31)we obtain
det (119883119905+1
)
= det (119883119905) exp (tr [minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905])
= det (119883119905) exp(tr[
[
120578 ( log (L(119883119905)minus1
119876)L(119883119905)minus1
119883119905
+ log (L(119883119905)minus1
119876)L(119883119905)minus1
times
119898
sum
119895=1
119860119879
119895119883119905119860119895)]
]
)
= det (119883119905) exp (tr [120578 log (L(119883
119905)minus1
119876)])
= det (119883119905) det ((L(119883
119905)minus1
119876)120578
)
(33)
Therefore the following property is valid
det (119883119905+1
)
det (119883119905)
= (det (119876)
det(L (119883119905))
120578
(34)
This equation shows that the determinant of (1) isclosely linked with that of the sequence 119883
119905 Moreover
the ratio det(119883119905+1
) det(119883119905) is directly proportional to
det(119876) det(L(119883119905)) for the fixed step-size 120578 In the case of
the RGAVS the similar conclusion can also be obtained
4 Simulations
Two simulation examples are given to compare the con-vergence speed of the Riemannian gradient algorithm (theRGACS and the RGAVS) with those of the traditional CGMThe error criterion used in these simulations is120588(L(119883)minus119876) lt
10minus15 where 120588(sdot) denotes the spectral radius of the matrix
Example 4 First we consider the matrices equation
119876 = 119883 + 119860119879
11198831198601
+ 119860119879
21198831198602
+ 119860119879
31198831198603
(35)
in the case of 119899 = 5 And 119876 1198601 1198602 1198603are as follows
119876 = (
46350 00698 03104 01990 02918
00698 40376 00742 00694 00690
03104 00742 42242 01554 01826
01990 00694 01554 41472 01572
02918 00690 01826 01572 42134
)
1198601
= (
00834 00007 00293 00142 00139
00141 02975 00154 00289 00339
00244 00152 00317 00270 00117
00251 00097 00140 00135 00240
00158 01499 00275 00077 00157
)
02 03 04 06 07 08050
20
40
60
80
100
120
140
160
Step size
Itera
tions
Figure 1 Step sizes versus the iterations in the RGACS
1198602
= (
02624 00166 00454 00172 00160
00893 00608 00242 00029 00061
00054 00190 00589 00036 00160
00278 00050 00255 00689 00048
00319 00064 00103 00219 00018
)
1198603
= (
81622 06020 04505 08258 01067
27943 70630 00838 05383 09619
03112 06541 62290 09961 00046
05285 06892 09133 10782 07749
01656 07482 01524 04427 90073
)
(36)
which satisfy the condition (10) Both the RGACS andRGAVS are used and compared with the traditional CGMand 119876 is taken as the initial value 119883
0in the following
simulation In the RGACS case the step sizes versus theiterations are shown in Figure 1 It can be found that theiterations will reduce gradually as the step-size changes fromzero to a critical value then the iterations will increasegradually above this critical valueTherefore the critical valueabout 064 can be selected experimentally as an optimal step-size and used in the RGACS
Figure 2 shows the convergence comparison between theRGACS and the RGAVS and the traditional CGM It can befound that the RGAVS possesses themost steady convergenceand the fastest convergence speed and it needs 23 iterationsto obtain the numerical solution of (35) as follows
(
00856 minus00265 00151 minus01065 00148
minus00265 00970 00166 minus01581 minus00023
00151 00166 01591 minus03608 00283
minus01065 minus01581 minus03608 23761 minus01723
00148 minus00023 00283 minus01723 00654
)
(37)
The case of the CGM realizes the given error through 33iterations and the slowest one among them is the RGACSwith 31 steps By comparison it is shown that the convergence
6 Journal of Applied Mathematics
5 10 15 20 25 30 35Iteration
Erro
r
RGACSRGAVS
CGM
105
100
10minus5
10minus10
Figure 2 Comparison of convergence speeds (log scale)
speed of the RGAVS is faster than both those of the RGACSand the traditional CGM in solving (35)
Example 5 A 20-order linear matrix equation with 119898 =
2 is also considered here following the similar procedureas Example 4 where 119876 isin S+(20) and 119860
1 1198602
isin R20times20
satisfy the condition (10) The simulation indicates that thenumerical solution needs 26 45 and 37 iterations in theRGAVS the RGACS (120578 = 074) and the traditional CGMrespectively Therefore the convergence speed of the RGAVSis also better than those of the RGACS and the CGM
5 Conclusion
Using geometric structures of manifold consisting of all sym-metric positive definite matrices the Riemannian gradientalgorithm is developed to calculate the numerical solutionfor the linear matrix equation 119876 = 119883 + sum
119898
119894=1119860119879
119894119883119860119894 In this
algorithm the geodesic distance on the curved Riemannianmanifold is taken as an objective function and the geodesiccurve is treated as the convergence path Also the variablestep sizes algorithm corresponding to the minimum valueof the objective function is provided in order to improvethe convergence speed in the constant step-size case Finallythe convergence speed of the Riemannian gradient algorithmhas been compared with the traditional conjugate gradientmethod by two simulation examples in 5-order and 20-orderlinear matrix equations respectively It is shown that theconvergence speed of the Riemannian gradient algorithmwith variable step sizes is faster than that of the conjugate gra-dient method Although the Riemannian gradient algorithmis used here for the linear matrix equation in the real numberfield it may be generalized to the complex case as well
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to express their sincere thanks tothe referees for their valuable suggestions which greatlyimproved the presentation of this paper This subject issupported by the National Natural Science Foundations ofChina (nos 61179031 10932002 and 51105033)
References
[1] M Berzig ldquoSolving a class of matrix equations via the Bhaskar-Lakshmikantham coupled fixed point theoremrdquo Applied Math-ematics Letters vol 25 no 11 pp 1638ndash1643 2012
[2] J C Engwerda ldquoOn the existence of a positive definite solutionof the matrix equation 119883 + 119860
119879
119883minus1
119860 = 119868rdquo Linear Algebra and ItsApplications vol 194 pp 91ndash108 1993
[3] W L Green and E W Kamen ldquoStabilizability of linear sys-tems over a commutative normed algebra with applications tospatially-distributed and parameter-dependent systemsrdquo SIAMJournal onControl andOptimization vol 23 no 1 pp 1ndash18 1985
[4] W Pusz and S L Woronowicz ldquoFunctional calculus forsesquilinear forms and the purification maprdquo Reports on Math-ematical Physics vol 8 no 2 pp 159ndash170 1975
[5] C-Y Chiang E K-W Chu and W-W Lin ldquoOn the 995333-Sylvester equation 119860119883 plusmn 119883
995333
119861995333
= 119862rdquo Applied Mathematics andComputation vol 218 no 17 pp 8393ndash8407 2012
[6] Z Gajic and M T J Qureshi Lyapunov Matrix Equation inSystem Stability and Control vol 195 Academic Press LondonUK 1995
[7] A C M Ran and M C B Reurings ldquoThe symmetric linearmatrix equationrdquo Electronic Journal of Linear Algebra vol 9 pp93ndash107 2002
[8] M C B Reurings Symmetric matrix equations [PhD thesis]Vrije Universiteit Amsterdam The Netherlands 2003
[9] A C M Ran and M C B Reurings ldquoA fixed point theoremin partially ordered sets and some applications to matrixequationsrdquo Proceedings of the American Mathematical Societyvol 132 no 5 pp 1435ndash1443 2004
[10] Y Su and G Chen ldquoIterative methods for solving linear matrixequation and linear matrix systemrdquo International Journal ofComputer Mathematics vol 87 no 4 pp 763ndash774 2010
[11] X Duan H Sun L Peng and X Zhao ldquoA natural gradi-ent descent algorithm for the solution of discrete algebraicLyapunov equations based on the geodesic distancerdquo AppliedMathematics and Computation vol 219 no 19 pp 9899ndash99052013
[12] X Duan H Sun and Z Zhang ldquoA natural gradient descentalgorithm for the solution of Lyapunov equations based on thegeodesic distancerdquo Journal of Computational Mathematics vol219 no 19 pp 9899ndash9905 2013
[13] M Spivak A Comprehensive Introduction to Differential Geom-etry vol 1 Publish or Perish Berkeley Calif USA 3rd edition1999
[14] S Lang Fundamentals of Differential Geometry vol 191Springer 1999
Journal of Applied Mathematics 7
[15] F Barbaresco ldquoInteractions between symmetric cones andinformation geometrics Bruhat-tits and Siegel spaces modelsfor high resolution autoregressive Doppler imageryrdquo in Emerg-ing Trends in Visual Computing vol 5416 of Lecture Notes inComputer Science pp 124ndash163 2009
[16] M Moakher ldquoA differential geometric approach to the geo-metric mean of symmetric positive-definite matricesrdquo SIAMJournal on Matrix Analysis and Applications vol 26 no 3 pp735ndash747 2005
[17] M Moakher ldquoOn the averaging of symmetric positive-definitetensorsrdquo Journal of Elasticity vol 82 no 3 pp 273ndash296 2006
[18] A Schwartzman Random ellipsoids and false discovery ratesstatistics for diffusion tensor imaging data [PhD thesis] StanfordUniversity 2006
[19] J D Lawson and Y Lim ldquoThe geometric mean matricesmetrics and morerdquo The American Mathematical Monthly vol108 no 9 pp 797ndash812 2001
[20] R Bhatia Positive definite Matrices Princeton Series in AppliedMathematics Princeton University Press Princeton NJ USA2007
[21] C Udriste Convex Functions and Optimization Methods onRiemannian Manifolds vol 297 ofMathematics and its Applica-tions Kluwer Academic Publishers Dordrecht Germany 1994
[22] M P D Carmo Riemannian Geometry Springer 1992[23] D G Luenberger ldquoThe gradient projection method along
geodesicsrdquoManagement Science vol 18 pp 620ndash631 1972
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 3
(2) Then the Riemannian gradient descent algorithm canbe used to adjust 119883 so that the difference will be assmall as possible
Note that the Riemannian manifold S+(119899) is geodesicallycomplete hence we take the geodesic distance between119876 andL(119883) as the objective function 119869(119883) namely
119869 (119883) = d2 (119876L (119883)) (11)
Then the exact solution of (1) may be defined as
119883lowast
= arg min119883isinS+(119899)
119869 (119883) (12)
In order to find aminimizer for the criterion d2(119876L(119883)) wemay make use of the differential equation (5) The minimumis apparently achieved for a value 119883
lowastsuch that
nabla119883
119869 (119883)1003816100381610038161003816119883=119883lowast
= 0 (13)
It is necessary to solve the optimization problem (12) by anumerical optimization algorithm that takes the geometry ofthe Riemannian manifold S+(119899) into account In particularstarting from an initial guess 119883
0 it is possible to provide an
iterative learning algorithm that generates a pattern 119883119905
isin
S+(119899) at any learning step 119905 isin N Following [21 23] suchsequence may be generated by moving from each point 119883
119905isin
S+(119899) to the next point 119883119905+1
along a short geodesic curvein the opposite direction of the Riemannian gradient of theobjective function 119869(119883
119905) Namely if we set
119881119905
= nabla119883119905
119869 (119883119905) (14)
then the Riemannian gradient algorithmmay be expressed as
119883119905+1
= 120574119883119905minus119881119905
(120578) (15)
where 120578 denotes any suitable step-size schedule that drives theiterative algorithm (15) to be convergent Moreover with theaim to employ the learning algorithm (14) and (15) we firstneed to compute the Riemannian gradient of the objectivefunction 119869(119883
119905) to be optimized In the definition of the
Riemannian gradient (4) the generic smooth curve 120601119884119881
maybe replaced with a geodesic curve We obtain the followingtheorem
Theorem 1 TheRiemannian gradient of the objective function119869(119883) at point 119883 is given by
nabla119883
119869 (119883) = 2119883 ( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119883
(16)
Proof First we compute the derivative of the real-valuedfunction
119869 (119904 (119905)) = tr [log2 (119876minus12
L (119904 (119905)) 119876minus12
)] (17)
with respect to 119905 where 119904(119905) = 11988312 exp(119905119883
minus12
119881119883minus12
)11988312
is the geodesic curve emanating from 119883 in the direction 119881 =
119904(0) Using Proposition 21 in [16] and some properties of thetrace we have
dd119905
119869 (119904 (119905))|119905=0
= 2 tr[
[
( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119881]
]
= ⟨119881 2119883 ( log (119876minus1
L (119883))L(119883)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (119883))L(119883)minus1
119860119879
119895) 119883⟩
119883
(18)
From (4) it is shown that formula (16) is valid
Remark 2 In fact note that log(119860minus1
119861119860) = 119860minus1
(log119861)119860 forall 119860 isin GL(119899) and for all 119861 isin S+(119899) then nabla
119883119869(119883) can be
rewritten as
nabla119883
119869 (119883)
= 2119883 (L(119883)minus12 log (L(119883)
12
119876minus1
L(119883)12
)L(119883)minus12
+
119898
sum
119895=1
119860119895L(119883)
minus12
times log (L(119883)12
119876minus1
L(119883)12
)L(119883)minus12
119860119879
119895) 119883
(19)
Therefore it can be shown that nabla119883
119869(119883) belongs to 119879119883S+(119899)
Corollary 3 TheRiemannian gradientnabla119883
119869(119883) has the uniquezero point 119883
lowast Furthermore for the suitable step-size 120578 the
sequence 119883119905generated by (15) converges to the solution 119883
lowastof
(1)
Proof Obviously 119883lowastis a zero point of the Riemannian gra-
dient nabla119883
119869(119883) If there exists 1198831015840
lowast= 119883lowastsuch that the iterative
process (15) stops then it means nabla1198831015840
lowast
119869(1198831015840
lowast) = 0 which is
equivalent to
log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
+
119898
sum
119895=1
119860119895log (119876
minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
119860119879
119895= 0
(20)
4 Journal of Applied Mathematics
On (20) the matrices are multiplied by 1198831015840
lowastand the traces of
the matrices are taken then we have
tr [log (119876minus1
L (1198831015840
lowast))] = 0 (21)
Also if we replace 1198831015840
lowastwith 119883
lowastand follow the similar
procedure as the above equation then the following equalitiesare valid
tr[log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
(119883lowast
+
119898
sum
119894=1
119860119879
119894119883lowast119860119894)]
= tr [log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
119876]
= 0
(22)
Combining (21) and (22) it can be obtained that
tr [log (119876minus1
L (1198831015840
lowast))] minus tr [log (119876
minus1
L (1198831015840
lowast)) 119876]
= tr [log (119876minus1
L (1198831015840
lowast)) (119868 minus L(119883
1015840
lowast)minus1
119876)]
= tr [log (119876minus12
L (1198831015840
lowast) 119876minus12
)
times (119868 minus 11987612
L(1198831015840
lowast)minus1
11987612
)]
= 0
(23)
Note that the matrix 11987612L(119883
1015840
lowast)minus1
11987612 is positive defi-
nite hence its orthogonal similarity diagonalization can beexpressed as 119880Λ119880
119879 where 119880 is an orthogonal matrix Λ =
diag (1205821 120582
119899) and 120582
119894gt 0 (1 le 119894 le 119899) Thus we have
tr [(Λ minus 119868) logΛ] =
119899
sum
119894=1
(120582119894minus 1) log 120582
119894
= 0 (24)
Note that the second equality of (24) is equivalent to
119899
prod
119894=1
120582119894
120582119894minus1 = 1 (25)
Thus we have 1205821
= 1205822
= sdot sdot sdot = 120582119899
= 1 That means that 1198831015840
lowastis
also the solution of (1) which is contrary to the uniquenessof the solution Therefore the Riemannian gradient nabla
119883119869(119883)
has the unique zero point 119883lowast This completes the proof of
Corollary 3
Now the Riemannian gradient algorithm with a constantstep-size (RGACS) becomes
119883119905+1
= 11988312
119905exp (minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905) 11988312
119905 (26)
where the initial value 1198830
isin S+(119899) The step-size 120578 corres-ponding to the best convergence speed can be obtained exper-imentally and the details will be described in our followingsimulations
31 Riemannian Gradient Algorithm with Variable Step SizesIn order to improve the convergence speed of the RGACSan optimal variable step-size schedule 120578
119905will be considered
and defined as follows The step-size 120578119905should be evaluated
in such a way that the objective function 119869(119883119905+1
) at step 119905 + 1
is reduced as much as possible with respect to 119869(119883119905) We may
regard 119869(119883119905+1
) as a function of the step-size 120578119905that
119869 (119883119905+1
) = d2 (119876L (120574119883119905minus119881119905
(120578119905))) (27)
Thus we can optimize the value of 120578119905to ensure that the
difference 119869(119883119905) minus 119869(119883
119905+1) could be as large as possible Since
it is difficult to solve this nonlinear problem exactly we use asuboptimal approximation in practice Under the hypothesisthat the step-size value 120578
119905is small enough we may invoke the
expansion of the function 119869(119883119905+1
) around the point 120578119905
= 0 asfollows
119869 (119883119905+1
) asymp 119869 (119883119905) + 1198621119905
120578119905+
1
21198622119905
1205782
119905 (28)
Then the optimal step-size 120578lowast
119905 corresponding to the maxi-
mum of the difference 119869(119883119905) minus 119869(119883
119905+1) can be expressed as
120578lowast
119905asymp minus
1198621119905
1198622119905
(29)
The coefficients 1198621119905 and 119862
2119905can be calculated using the
first-order and the second-order derivatives of the function119869(119883119905+1
) with respect to the parameter 120578119905at the point 120578
119905= 0
On the basis of the above discussion we may obtain theoptimal step-size
120578lowast
119905asymp
tr [(119883L (log (119876minus1L (119883
119905))L(119883
119905)minus1
))2
]
tr [(L(119883119905)minus1
L (nabla119883119905
119869 (119883119905)))2
]
(30)
Then the Riemannian gradient algorithm with variablestep-sizes (RGAVS) can be written explicitly as
119883119905+1
= 11988312
119905exp (minus120578
lowast
119905119883minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905) 11988312
119905 (31)
where 1198830
isin S+(119899) is the initial value
32 Volume Feature of Riemannian Gradient Algorithm TheRGACS shows an interesting volume feature that the ratiodet(119883
119905+1) det(119883
119905) always keeps a certain relationship In
order to prove the volume feature let us recall two propertiesof determinant operator det namely for each 119860 119861 isin GL(119899)there hold
det (119860119861) = det (119860) det (119861)
det (exp (119860)) = exp (tr (119860)) (32)
Journal of Applied Mathematics 5
By applying the determinant operator to both sides of (31)we obtain
det (119883119905+1
)
= det (119883119905) exp (tr [minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905])
= det (119883119905) exp(tr[
[
120578 ( log (L(119883119905)minus1
119876)L(119883119905)minus1
119883119905
+ log (L(119883119905)minus1
119876)L(119883119905)minus1
times
119898
sum
119895=1
119860119879
119895119883119905119860119895)]
]
)
= det (119883119905) exp (tr [120578 log (L(119883
119905)minus1
119876)])
= det (119883119905) det ((L(119883
119905)minus1
119876)120578
)
(33)
Therefore the following property is valid
det (119883119905+1
)
det (119883119905)
= (det (119876)
det(L (119883119905))
120578
(34)
This equation shows that the determinant of (1) isclosely linked with that of the sequence 119883
119905 Moreover
the ratio det(119883119905+1
) det(119883119905) is directly proportional to
det(119876) det(L(119883119905)) for the fixed step-size 120578 In the case of
the RGAVS the similar conclusion can also be obtained
4 Simulations
Two simulation examples are given to compare the con-vergence speed of the Riemannian gradient algorithm (theRGACS and the RGAVS) with those of the traditional CGMThe error criterion used in these simulations is120588(L(119883)minus119876) lt
10minus15 where 120588(sdot) denotes the spectral radius of the matrix
Example 4 First we consider the matrices equation
119876 = 119883 + 119860119879
11198831198601
+ 119860119879
21198831198602
+ 119860119879
31198831198603
(35)
in the case of 119899 = 5 And 119876 1198601 1198602 1198603are as follows
119876 = (
46350 00698 03104 01990 02918
00698 40376 00742 00694 00690
03104 00742 42242 01554 01826
01990 00694 01554 41472 01572
02918 00690 01826 01572 42134
)
1198601
= (
00834 00007 00293 00142 00139
00141 02975 00154 00289 00339
00244 00152 00317 00270 00117
00251 00097 00140 00135 00240
00158 01499 00275 00077 00157
)
02 03 04 06 07 08050
20
40
60
80
100
120
140
160
Step size
Itera
tions
Figure 1 Step sizes versus the iterations in the RGACS
1198602
= (
02624 00166 00454 00172 00160
00893 00608 00242 00029 00061
00054 00190 00589 00036 00160
00278 00050 00255 00689 00048
00319 00064 00103 00219 00018
)
1198603
= (
81622 06020 04505 08258 01067
27943 70630 00838 05383 09619
03112 06541 62290 09961 00046
05285 06892 09133 10782 07749
01656 07482 01524 04427 90073
)
(36)
which satisfy the condition (10) Both the RGACS andRGAVS are used and compared with the traditional CGMand 119876 is taken as the initial value 119883
0in the following
simulation In the RGACS case the step sizes versus theiterations are shown in Figure 1 It can be found that theiterations will reduce gradually as the step-size changes fromzero to a critical value then the iterations will increasegradually above this critical valueTherefore the critical valueabout 064 can be selected experimentally as an optimal step-size and used in the RGACS
Figure 2 shows the convergence comparison between theRGACS and the RGAVS and the traditional CGM It can befound that the RGAVS possesses themost steady convergenceand the fastest convergence speed and it needs 23 iterationsto obtain the numerical solution of (35) as follows
(
00856 minus00265 00151 minus01065 00148
minus00265 00970 00166 minus01581 minus00023
00151 00166 01591 minus03608 00283
minus01065 minus01581 minus03608 23761 minus01723
00148 minus00023 00283 minus01723 00654
)
(37)
The case of the CGM realizes the given error through 33iterations and the slowest one among them is the RGACSwith 31 steps By comparison it is shown that the convergence
6 Journal of Applied Mathematics
5 10 15 20 25 30 35Iteration
Erro
r
RGACSRGAVS
CGM
105
100
10minus5
10minus10
Figure 2 Comparison of convergence speeds (log scale)
speed of the RGAVS is faster than both those of the RGACSand the traditional CGM in solving (35)
Example 5 A 20-order linear matrix equation with 119898 =
2 is also considered here following the similar procedureas Example 4 where 119876 isin S+(20) and 119860
1 1198602
isin R20times20
satisfy the condition (10) The simulation indicates that thenumerical solution needs 26 45 and 37 iterations in theRGAVS the RGACS (120578 = 074) and the traditional CGMrespectively Therefore the convergence speed of the RGAVSis also better than those of the RGACS and the CGM
5 Conclusion
Using geometric structures of manifold consisting of all sym-metric positive definite matrices the Riemannian gradientalgorithm is developed to calculate the numerical solutionfor the linear matrix equation 119876 = 119883 + sum
119898
119894=1119860119879
119894119883119860119894 In this
algorithm the geodesic distance on the curved Riemannianmanifold is taken as an objective function and the geodesiccurve is treated as the convergence path Also the variablestep sizes algorithm corresponding to the minimum valueof the objective function is provided in order to improvethe convergence speed in the constant step-size case Finallythe convergence speed of the Riemannian gradient algorithmhas been compared with the traditional conjugate gradientmethod by two simulation examples in 5-order and 20-orderlinear matrix equations respectively It is shown that theconvergence speed of the Riemannian gradient algorithmwith variable step sizes is faster than that of the conjugate gra-dient method Although the Riemannian gradient algorithmis used here for the linear matrix equation in the real numberfield it may be generalized to the complex case as well
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to express their sincere thanks tothe referees for their valuable suggestions which greatlyimproved the presentation of this paper This subject issupported by the National Natural Science Foundations ofChina (nos 61179031 10932002 and 51105033)
References
[1] M Berzig ldquoSolving a class of matrix equations via the Bhaskar-Lakshmikantham coupled fixed point theoremrdquo Applied Math-ematics Letters vol 25 no 11 pp 1638ndash1643 2012
[2] J C Engwerda ldquoOn the existence of a positive definite solutionof the matrix equation 119883 + 119860
119879
119883minus1
119860 = 119868rdquo Linear Algebra and ItsApplications vol 194 pp 91ndash108 1993
[3] W L Green and E W Kamen ldquoStabilizability of linear sys-tems over a commutative normed algebra with applications tospatially-distributed and parameter-dependent systemsrdquo SIAMJournal onControl andOptimization vol 23 no 1 pp 1ndash18 1985
[4] W Pusz and S L Woronowicz ldquoFunctional calculus forsesquilinear forms and the purification maprdquo Reports on Math-ematical Physics vol 8 no 2 pp 159ndash170 1975
[5] C-Y Chiang E K-W Chu and W-W Lin ldquoOn the 995333-Sylvester equation 119860119883 plusmn 119883
995333
119861995333
= 119862rdquo Applied Mathematics andComputation vol 218 no 17 pp 8393ndash8407 2012
[6] Z Gajic and M T J Qureshi Lyapunov Matrix Equation inSystem Stability and Control vol 195 Academic Press LondonUK 1995
[7] A C M Ran and M C B Reurings ldquoThe symmetric linearmatrix equationrdquo Electronic Journal of Linear Algebra vol 9 pp93ndash107 2002
[8] M C B Reurings Symmetric matrix equations [PhD thesis]Vrije Universiteit Amsterdam The Netherlands 2003
[9] A C M Ran and M C B Reurings ldquoA fixed point theoremin partially ordered sets and some applications to matrixequationsrdquo Proceedings of the American Mathematical Societyvol 132 no 5 pp 1435ndash1443 2004
[10] Y Su and G Chen ldquoIterative methods for solving linear matrixequation and linear matrix systemrdquo International Journal ofComputer Mathematics vol 87 no 4 pp 763ndash774 2010
[11] X Duan H Sun L Peng and X Zhao ldquoA natural gradi-ent descent algorithm for the solution of discrete algebraicLyapunov equations based on the geodesic distancerdquo AppliedMathematics and Computation vol 219 no 19 pp 9899ndash99052013
[12] X Duan H Sun and Z Zhang ldquoA natural gradient descentalgorithm for the solution of Lyapunov equations based on thegeodesic distancerdquo Journal of Computational Mathematics vol219 no 19 pp 9899ndash9905 2013
[13] M Spivak A Comprehensive Introduction to Differential Geom-etry vol 1 Publish or Perish Berkeley Calif USA 3rd edition1999
[14] S Lang Fundamentals of Differential Geometry vol 191Springer 1999
Journal of Applied Mathematics 7
[15] F Barbaresco ldquoInteractions between symmetric cones andinformation geometrics Bruhat-tits and Siegel spaces modelsfor high resolution autoregressive Doppler imageryrdquo in Emerg-ing Trends in Visual Computing vol 5416 of Lecture Notes inComputer Science pp 124ndash163 2009
[16] M Moakher ldquoA differential geometric approach to the geo-metric mean of symmetric positive-definite matricesrdquo SIAMJournal on Matrix Analysis and Applications vol 26 no 3 pp735ndash747 2005
[17] M Moakher ldquoOn the averaging of symmetric positive-definitetensorsrdquo Journal of Elasticity vol 82 no 3 pp 273ndash296 2006
[18] A Schwartzman Random ellipsoids and false discovery ratesstatistics for diffusion tensor imaging data [PhD thesis] StanfordUniversity 2006
[19] J D Lawson and Y Lim ldquoThe geometric mean matricesmetrics and morerdquo The American Mathematical Monthly vol108 no 9 pp 797ndash812 2001
[20] R Bhatia Positive definite Matrices Princeton Series in AppliedMathematics Princeton University Press Princeton NJ USA2007
[21] C Udriste Convex Functions and Optimization Methods onRiemannian Manifolds vol 297 ofMathematics and its Applica-tions Kluwer Academic Publishers Dordrecht Germany 1994
[22] M P D Carmo Riemannian Geometry Springer 1992[23] D G Luenberger ldquoThe gradient projection method along
geodesicsrdquoManagement Science vol 18 pp 620ndash631 1972
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
4 Journal of Applied Mathematics
On (20) the matrices are multiplied by 1198831015840
lowastand the traces of
the matrices are taken then we have
tr [log (119876minus1
L (1198831015840
lowast))] = 0 (21)
Also if we replace 1198831015840
lowastwith 119883
lowastand follow the similar
procedure as the above equation then the following equalitiesare valid
tr[log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
(119883lowast
+
119898
sum
119894=1
119860119879
119894119883lowast119860119894)]
= tr [log (119876minus1
L (1198831015840
lowast))L(119883
1015840
lowast)minus1
119876]
= 0
(22)
Combining (21) and (22) it can be obtained that
tr [log (119876minus1
L (1198831015840
lowast))] minus tr [log (119876
minus1
L (1198831015840
lowast)) 119876]
= tr [log (119876minus1
L (1198831015840
lowast)) (119868 minus L(119883
1015840
lowast)minus1
119876)]
= tr [log (119876minus12
L (1198831015840
lowast) 119876minus12
)
times (119868 minus 11987612
L(1198831015840
lowast)minus1
11987612
)]
= 0
(23)
Note that the matrix 11987612L(119883
1015840
lowast)minus1
11987612 is positive defi-
nite hence its orthogonal similarity diagonalization can beexpressed as 119880Λ119880
119879 where 119880 is an orthogonal matrix Λ =
diag (1205821 120582
119899) and 120582
119894gt 0 (1 le 119894 le 119899) Thus we have
tr [(Λ minus 119868) logΛ] =
119899
sum
119894=1
(120582119894minus 1) log 120582
119894
= 0 (24)
Note that the second equality of (24) is equivalent to
119899
prod
119894=1
120582119894
120582119894minus1 = 1 (25)
Thus we have 1205821
= 1205822
= sdot sdot sdot = 120582119899
= 1 That means that 1198831015840
lowastis
also the solution of (1) which is contrary to the uniquenessof the solution Therefore the Riemannian gradient nabla
119883119869(119883)
has the unique zero point 119883lowast This completes the proof of
Corollary 3
Now the Riemannian gradient algorithm with a constantstep-size (RGACS) becomes
119883119905+1
= 11988312
119905exp (minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905) 11988312
119905 (26)
where the initial value 1198830
isin S+(119899) The step-size 120578 corres-ponding to the best convergence speed can be obtained exper-imentally and the details will be described in our followingsimulations
31 Riemannian Gradient Algorithm with Variable Step SizesIn order to improve the convergence speed of the RGACSan optimal variable step-size schedule 120578
119905will be considered
and defined as follows The step-size 120578119905should be evaluated
in such a way that the objective function 119869(119883119905+1
) at step 119905 + 1
is reduced as much as possible with respect to 119869(119883119905) We may
regard 119869(119883119905+1
) as a function of the step-size 120578119905that
119869 (119883119905+1
) = d2 (119876L (120574119883119905minus119881119905
(120578119905))) (27)
Thus we can optimize the value of 120578119905to ensure that the
difference 119869(119883119905) minus 119869(119883
119905+1) could be as large as possible Since
it is difficult to solve this nonlinear problem exactly we use asuboptimal approximation in practice Under the hypothesisthat the step-size value 120578
119905is small enough we may invoke the
expansion of the function 119869(119883119905+1
) around the point 120578119905
= 0 asfollows
119869 (119883119905+1
) asymp 119869 (119883119905) + 1198621119905
120578119905+
1
21198622119905
1205782
119905 (28)
Then the optimal step-size 120578lowast
119905 corresponding to the maxi-
mum of the difference 119869(119883119905) minus 119869(119883
119905+1) can be expressed as
120578lowast
119905asymp minus
1198621119905
1198622119905
(29)
The coefficients 1198621119905 and 119862
2119905can be calculated using the
first-order and the second-order derivatives of the function119869(119883119905+1
) with respect to the parameter 120578119905at the point 120578
119905= 0
On the basis of the above discussion we may obtain theoptimal step-size
120578lowast
119905asymp
tr [(119883L (log (119876minus1L (119883
119905))L(119883
119905)minus1
))2
]
tr [(L(119883119905)minus1
L (nabla119883119905
119869 (119883119905)))2
]
(30)
Then the Riemannian gradient algorithm with variablestep-sizes (RGAVS) can be written explicitly as
119883119905+1
= 11988312
119905exp (minus120578
lowast
119905119883minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905) 11988312
119905 (31)
where 1198830
isin S+(119899) is the initial value
32 Volume Feature of Riemannian Gradient Algorithm TheRGACS shows an interesting volume feature that the ratiodet(119883
119905+1) det(119883
119905) always keeps a certain relationship In
order to prove the volume feature let us recall two propertiesof determinant operator det namely for each 119860 119861 isin GL(119899)there hold
det (119860119861) = det (119860) det (119861)
det (exp (119860)) = exp (tr (119860)) (32)
Journal of Applied Mathematics 5
By applying the determinant operator to both sides of (31)we obtain
det (119883119905+1
)
= det (119883119905) exp (tr [minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905])
= det (119883119905) exp(tr[
[
120578 ( log (L(119883119905)minus1
119876)L(119883119905)minus1
119883119905
+ log (L(119883119905)minus1
119876)L(119883119905)minus1
times
119898
sum
119895=1
119860119879
119895119883119905119860119895)]
]
)
= det (119883119905) exp (tr [120578 log (L(119883
119905)minus1
119876)])
= det (119883119905) det ((L(119883
119905)minus1
119876)120578
)
(33)
Therefore the following property is valid
det (119883119905+1
)
det (119883119905)
= (det (119876)
det(L (119883119905))
120578
(34)
This equation shows that the determinant of (1) isclosely linked with that of the sequence 119883
119905 Moreover
the ratio det(119883119905+1
) det(119883119905) is directly proportional to
det(119876) det(L(119883119905)) for the fixed step-size 120578 In the case of
the RGAVS the similar conclusion can also be obtained
4 Simulations
Two simulation examples are given to compare the con-vergence speed of the Riemannian gradient algorithm (theRGACS and the RGAVS) with those of the traditional CGMThe error criterion used in these simulations is120588(L(119883)minus119876) lt
10minus15 where 120588(sdot) denotes the spectral radius of the matrix
Example 4 First we consider the matrices equation
119876 = 119883 + 119860119879
11198831198601
+ 119860119879
21198831198602
+ 119860119879
31198831198603
(35)
in the case of 119899 = 5 And 119876 1198601 1198602 1198603are as follows
119876 = (
46350 00698 03104 01990 02918
00698 40376 00742 00694 00690
03104 00742 42242 01554 01826
01990 00694 01554 41472 01572
02918 00690 01826 01572 42134
)
1198601
= (
00834 00007 00293 00142 00139
00141 02975 00154 00289 00339
00244 00152 00317 00270 00117
00251 00097 00140 00135 00240
00158 01499 00275 00077 00157
)
02 03 04 06 07 08050
20
40
60
80
100
120
140
160
Step size
Itera
tions
Figure 1 Step sizes versus the iterations in the RGACS
1198602
= (
02624 00166 00454 00172 00160
00893 00608 00242 00029 00061
00054 00190 00589 00036 00160
00278 00050 00255 00689 00048
00319 00064 00103 00219 00018
)
1198603
= (
81622 06020 04505 08258 01067
27943 70630 00838 05383 09619
03112 06541 62290 09961 00046
05285 06892 09133 10782 07749
01656 07482 01524 04427 90073
)
(36)
which satisfy the condition (10) Both the RGACS andRGAVS are used and compared with the traditional CGMand 119876 is taken as the initial value 119883
0in the following
simulation In the RGACS case the step sizes versus theiterations are shown in Figure 1 It can be found that theiterations will reduce gradually as the step-size changes fromzero to a critical value then the iterations will increasegradually above this critical valueTherefore the critical valueabout 064 can be selected experimentally as an optimal step-size and used in the RGACS
Figure 2 shows the convergence comparison between theRGACS and the RGAVS and the traditional CGM It can befound that the RGAVS possesses themost steady convergenceand the fastest convergence speed and it needs 23 iterationsto obtain the numerical solution of (35) as follows
(
00856 minus00265 00151 minus01065 00148
minus00265 00970 00166 minus01581 minus00023
00151 00166 01591 minus03608 00283
minus01065 minus01581 minus03608 23761 minus01723
00148 minus00023 00283 minus01723 00654
)
(37)
The case of the CGM realizes the given error through 33iterations and the slowest one among them is the RGACSwith 31 steps By comparison it is shown that the convergence
6 Journal of Applied Mathematics
5 10 15 20 25 30 35Iteration
Erro
r
RGACSRGAVS
CGM
105
100
10minus5
10minus10
Figure 2 Comparison of convergence speeds (log scale)
speed of the RGAVS is faster than both those of the RGACSand the traditional CGM in solving (35)
Example 5 A 20-order linear matrix equation with 119898 =
2 is also considered here following the similar procedureas Example 4 where 119876 isin S+(20) and 119860
1 1198602
isin R20times20
satisfy the condition (10) The simulation indicates that thenumerical solution needs 26 45 and 37 iterations in theRGAVS the RGACS (120578 = 074) and the traditional CGMrespectively Therefore the convergence speed of the RGAVSis also better than those of the RGACS and the CGM
5 Conclusion
Using geometric structures of manifold consisting of all sym-metric positive definite matrices the Riemannian gradientalgorithm is developed to calculate the numerical solutionfor the linear matrix equation 119876 = 119883 + sum
119898
119894=1119860119879
119894119883119860119894 In this
algorithm the geodesic distance on the curved Riemannianmanifold is taken as an objective function and the geodesiccurve is treated as the convergence path Also the variablestep sizes algorithm corresponding to the minimum valueof the objective function is provided in order to improvethe convergence speed in the constant step-size case Finallythe convergence speed of the Riemannian gradient algorithmhas been compared with the traditional conjugate gradientmethod by two simulation examples in 5-order and 20-orderlinear matrix equations respectively It is shown that theconvergence speed of the Riemannian gradient algorithmwith variable step sizes is faster than that of the conjugate gra-dient method Although the Riemannian gradient algorithmis used here for the linear matrix equation in the real numberfield it may be generalized to the complex case as well
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to express their sincere thanks tothe referees for their valuable suggestions which greatlyimproved the presentation of this paper This subject issupported by the National Natural Science Foundations ofChina (nos 61179031 10932002 and 51105033)
References
[1] M Berzig ldquoSolving a class of matrix equations via the Bhaskar-Lakshmikantham coupled fixed point theoremrdquo Applied Math-ematics Letters vol 25 no 11 pp 1638ndash1643 2012
[2] J C Engwerda ldquoOn the existence of a positive definite solutionof the matrix equation 119883 + 119860
119879
119883minus1
119860 = 119868rdquo Linear Algebra and ItsApplications vol 194 pp 91ndash108 1993
[3] W L Green and E W Kamen ldquoStabilizability of linear sys-tems over a commutative normed algebra with applications tospatially-distributed and parameter-dependent systemsrdquo SIAMJournal onControl andOptimization vol 23 no 1 pp 1ndash18 1985
[4] W Pusz and S L Woronowicz ldquoFunctional calculus forsesquilinear forms and the purification maprdquo Reports on Math-ematical Physics vol 8 no 2 pp 159ndash170 1975
[5] C-Y Chiang E K-W Chu and W-W Lin ldquoOn the 995333-Sylvester equation 119860119883 plusmn 119883
995333
119861995333
= 119862rdquo Applied Mathematics andComputation vol 218 no 17 pp 8393ndash8407 2012
[6] Z Gajic and M T J Qureshi Lyapunov Matrix Equation inSystem Stability and Control vol 195 Academic Press LondonUK 1995
[7] A C M Ran and M C B Reurings ldquoThe symmetric linearmatrix equationrdquo Electronic Journal of Linear Algebra vol 9 pp93ndash107 2002
[8] M C B Reurings Symmetric matrix equations [PhD thesis]Vrije Universiteit Amsterdam The Netherlands 2003
[9] A C M Ran and M C B Reurings ldquoA fixed point theoremin partially ordered sets and some applications to matrixequationsrdquo Proceedings of the American Mathematical Societyvol 132 no 5 pp 1435ndash1443 2004
[10] Y Su and G Chen ldquoIterative methods for solving linear matrixequation and linear matrix systemrdquo International Journal ofComputer Mathematics vol 87 no 4 pp 763ndash774 2010
[11] X Duan H Sun L Peng and X Zhao ldquoA natural gradi-ent descent algorithm for the solution of discrete algebraicLyapunov equations based on the geodesic distancerdquo AppliedMathematics and Computation vol 219 no 19 pp 9899ndash99052013
[12] X Duan H Sun and Z Zhang ldquoA natural gradient descentalgorithm for the solution of Lyapunov equations based on thegeodesic distancerdquo Journal of Computational Mathematics vol219 no 19 pp 9899ndash9905 2013
[13] M Spivak A Comprehensive Introduction to Differential Geom-etry vol 1 Publish or Perish Berkeley Calif USA 3rd edition1999
[14] S Lang Fundamentals of Differential Geometry vol 191Springer 1999
Journal of Applied Mathematics 7
[15] F Barbaresco ldquoInteractions between symmetric cones andinformation geometrics Bruhat-tits and Siegel spaces modelsfor high resolution autoregressive Doppler imageryrdquo in Emerg-ing Trends in Visual Computing vol 5416 of Lecture Notes inComputer Science pp 124ndash163 2009
[16] M Moakher ldquoA differential geometric approach to the geo-metric mean of symmetric positive-definite matricesrdquo SIAMJournal on Matrix Analysis and Applications vol 26 no 3 pp735ndash747 2005
[17] M Moakher ldquoOn the averaging of symmetric positive-definitetensorsrdquo Journal of Elasticity vol 82 no 3 pp 273ndash296 2006
[18] A Schwartzman Random ellipsoids and false discovery ratesstatistics for diffusion tensor imaging data [PhD thesis] StanfordUniversity 2006
[19] J D Lawson and Y Lim ldquoThe geometric mean matricesmetrics and morerdquo The American Mathematical Monthly vol108 no 9 pp 797ndash812 2001
[20] R Bhatia Positive definite Matrices Princeton Series in AppliedMathematics Princeton University Press Princeton NJ USA2007
[21] C Udriste Convex Functions and Optimization Methods onRiemannian Manifolds vol 297 ofMathematics and its Applica-tions Kluwer Academic Publishers Dordrecht Germany 1994
[22] M P D Carmo Riemannian Geometry Springer 1992[23] D G Luenberger ldquoThe gradient projection method along
geodesicsrdquoManagement Science vol 18 pp 620ndash631 1972
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 5
By applying the determinant operator to both sides of (31)we obtain
det (119883119905+1
)
= det (119883119905) exp (tr [minus120578119883
minus12
119905nabla119883119905
119869 (119883119905) 119883minus12
119905])
= det (119883119905) exp(tr[
[
120578 ( log (L(119883119905)minus1
119876)L(119883119905)minus1
119883119905
+ log (L(119883119905)minus1
119876)L(119883119905)minus1
times
119898
sum
119895=1
119860119879
119895119883119905119860119895)]
]
)
= det (119883119905) exp (tr [120578 log (L(119883
119905)minus1
119876)])
= det (119883119905) det ((L(119883
119905)minus1
119876)120578
)
(33)
Therefore the following property is valid
det (119883119905+1
)
det (119883119905)
= (det (119876)
det(L (119883119905))
120578
(34)
This equation shows that the determinant of (1) isclosely linked with that of the sequence 119883
119905 Moreover
the ratio det(119883119905+1
) det(119883119905) is directly proportional to
det(119876) det(L(119883119905)) for the fixed step-size 120578 In the case of
the RGAVS the similar conclusion can also be obtained
4 Simulations
Two simulation examples are given to compare the con-vergence speed of the Riemannian gradient algorithm (theRGACS and the RGAVS) with those of the traditional CGMThe error criterion used in these simulations is120588(L(119883)minus119876) lt
10minus15 where 120588(sdot) denotes the spectral radius of the matrix
Example 4 First we consider the matrices equation
119876 = 119883 + 119860119879
11198831198601
+ 119860119879
21198831198602
+ 119860119879
31198831198603
(35)
in the case of 119899 = 5 And 119876 1198601 1198602 1198603are as follows
119876 = (
46350 00698 03104 01990 02918
00698 40376 00742 00694 00690
03104 00742 42242 01554 01826
01990 00694 01554 41472 01572
02918 00690 01826 01572 42134
)
1198601
= (
00834 00007 00293 00142 00139
00141 02975 00154 00289 00339
00244 00152 00317 00270 00117
00251 00097 00140 00135 00240
00158 01499 00275 00077 00157
)
02 03 04 06 07 08050
20
40
60
80
100
120
140
160
Step size
Itera
tions
Figure 1 Step sizes versus the iterations in the RGACS
1198602
= (
02624 00166 00454 00172 00160
00893 00608 00242 00029 00061
00054 00190 00589 00036 00160
00278 00050 00255 00689 00048
00319 00064 00103 00219 00018
)
1198603
= (
81622 06020 04505 08258 01067
27943 70630 00838 05383 09619
03112 06541 62290 09961 00046
05285 06892 09133 10782 07749
01656 07482 01524 04427 90073
)
(36)
which satisfy the condition (10) Both the RGACS andRGAVS are used and compared with the traditional CGMand 119876 is taken as the initial value 119883
0in the following
simulation In the RGACS case the step sizes versus theiterations are shown in Figure 1 It can be found that theiterations will reduce gradually as the step-size changes fromzero to a critical value then the iterations will increasegradually above this critical valueTherefore the critical valueabout 064 can be selected experimentally as an optimal step-size and used in the RGACS
Figure 2 shows the convergence comparison between theRGACS and the RGAVS and the traditional CGM It can befound that the RGAVS possesses themost steady convergenceand the fastest convergence speed and it needs 23 iterationsto obtain the numerical solution of (35) as follows
(
00856 minus00265 00151 minus01065 00148
minus00265 00970 00166 minus01581 minus00023
00151 00166 01591 minus03608 00283
minus01065 minus01581 minus03608 23761 minus01723
00148 minus00023 00283 minus01723 00654
)
(37)
The case of the CGM realizes the given error through 33iterations and the slowest one among them is the RGACSwith 31 steps By comparison it is shown that the convergence
6 Journal of Applied Mathematics
5 10 15 20 25 30 35Iteration
Erro
r
RGACSRGAVS
CGM
105
100
10minus5
10minus10
Figure 2 Comparison of convergence speeds (log scale)
speed of the RGAVS is faster than both those of the RGACSand the traditional CGM in solving (35)
Example 5 A 20-order linear matrix equation with 119898 =
2 is also considered here following the similar procedureas Example 4 where 119876 isin S+(20) and 119860
1 1198602
isin R20times20
satisfy the condition (10) The simulation indicates that thenumerical solution needs 26 45 and 37 iterations in theRGAVS the RGACS (120578 = 074) and the traditional CGMrespectively Therefore the convergence speed of the RGAVSis also better than those of the RGACS and the CGM
5 Conclusion
Using geometric structures of manifold consisting of all sym-metric positive definite matrices the Riemannian gradientalgorithm is developed to calculate the numerical solutionfor the linear matrix equation 119876 = 119883 + sum
119898
119894=1119860119879
119894119883119860119894 In this
algorithm the geodesic distance on the curved Riemannianmanifold is taken as an objective function and the geodesiccurve is treated as the convergence path Also the variablestep sizes algorithm corresponding to the minimum valueof the objective function is provided in order to improvethe convergence speed in the constant step-size case Finallythe convergence speed of the Riemannian gradient algorithmhas been compared with the traditional conjugate gradientmethod by two simulation examples in 5-order and 20-orderlinear matrix equations respectively It is shown that theconvergence speed of the Riemannian gradient algorithmwith variable step sizes is faster than that of the conjugate gra-dient method Although the Riemannian gradient algorithmis used here for the linear matrix equation in the real numberfield it may be generalized to the complex case as well
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to express their sincere thanks tothe referees for their valuable suggestions which greatlyimproved the presentation of this paper This subject issupported by the National Natural Science Foundations ofChina (nos 61179031 10932002 and 51105033)
References
[1] M Berzig ldquoSolving a class of matrix equations via the Bhaskar-Lakshmikantham coupled fixed point theoremrdquo Applied Math-ematics Letters vol 25 no 11 pp 1638ndash1643 2012
[2] J C Engwerda ldquoOn the existence of a positive definite solutionof the matrix equation 119883 + 119860
119879
119883minus1
119860 = 119868rdquo Linear Algebra and ItsApplications vol 194 pp 91ndash108 1993
[3] W L Green and E W Kamen ldquoStabilizability of linear sys-tems over a commutative normed algebra with applications tospatially-distributed and parameter-dependent systemsrdquo SIAMJournal onControl andOptimization vol 23 no 1 pp 1ndash18 1985
[4] W Pusz and S L Woronowicz ldquoFunctional calculus forsesquilinear forms and the purification maprdquo Reports on Math-ematical Physics vol 8 no 2 pp 159ndash170 1975
[5] C-Y Chiang E K-W Chu and W-W Lin ldquoOn the 995333-Sylvester equation 119860119883 plusmn 119883
995333
119861995333
= 119862rdquo Applied Mathematics andComputation vol 218 no 17 pp 8393ndash8407 2012
[6] Z Gajic and M T J Qureshi Lyapunov Matrix Equation inSystem Stability and Control vol 195 Academic Press LondonUK 1995
[7] A C M Ran and M C B Reurings ldquoThe symmetric linearmatrix equationrdquo Electronic Journal of Linear Algebra vol 9 pp93ndash107 2002
[8] M C B Reurings Symmetric matrix equations [PhD thesis]Vrije Universiteit Amsterdam The Netherlands 2003
[9] A C M Ran and M C B Reurings ldquoA fixed point theoremin partially ordered sets and some applications to matrixequationsrdquo Proceedings of the American Mathematical Societyvol 132 no 5 pp 1435ndash1443 2004
[10] Y Su and G Chen ldquoIterative methods for solving linear matrixequation and linear matrix systemrdquo International Journal ofComputer Mathematics vol 87 no 4 pp 763ndash774 2010
[11] X Duan H Sun L Peng and X Zhao ldquoA natural gradi-ent descent algorithm for the solution of discrete algebraicLyapunov equations based on the geodesic distancerdquo AppliedMathematics and Computation vol 219 no 19 pp 9899ndash99052013
[12] X Duan H Sun and Z Zhang ldquoA natural gradient descentalgorithm for the solution of Lyapunov equations based on thegeodesic distancerdquo Journal of Computational Mathematics vol219 no 19 pp 9899ndash9905 2013
[13] M Spivak A Comprehensive Introduction to Differential Geom-etry vol 1 Publish or Perish Berkeley Calif USA 3rd edition1999
[14] S Lang Fundamentals of Differential Geometry vol 191Springer 1999
Journal of Applied Mathematics 7
[15] F Barbaresco ldquoInteractions between symmetric cones andinformation geometrics Bruhat-tits and Siegel spaces modelsfor high resolution autoregressive Doppler imageryrdquo in Emerg-ing Trends in Visual Computing vol 5416 of Lecture Notes inComputer Science pp 124ndash163 2009
[16] M Moakher ldquoA differential geometric approach to the geo-metric mean of symmetric positive-definite matricesrdquo SIAMJournal on Matrix Analysis and Applications vol 26 no 3 pp735ndash747 2005
[17] M Moakher ldquoOn the averaging of symmetric positive-definitetensorsrdquo Journal of Elasticity vol 82 no 3 pp 273ndash296 2006
[18] A Schwartzman Random ellipsoids and false discovery ratesstatistics for diffusion tensor imaging data [PhD thesis] StanfordUniversity 2006
[19] J D Lawson and Y Lim ldquoThe geometric mean matricesmetrics and morerdquo The American Mathematical Monthly vol108 no 9 pp 797ndash812 2001
[20] R Bhatia Positive definite Matrices Princeton Series in AppliedMathematics Princeton University Press Princeton NJ USA2007
[21] C Udriste Convex Functions and Optimization Methods onRiemannian Manifolds vol 297 ofMathematics and its Applica-tions Kluwer Academic Publishers Dordrecht Germany 1994
[22] M P D Carmo Riemannian Geometry Springer 1992[23] D G Luenberger ldquoThe gradient projection method along
geodesicsrdquoManagement Science vol 18 pp 620ndash631 1972
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
6 Journal of Applied Mathematics
5 10 15 20 25 30 35Iteration
Erro
r
RGACSRGAVS
CGM
105
100
10minus5
10minus10
Figure 2 Comparison of convergence speeds (log scale)
speed of the RGAVS is faster than both those of the RGACSand the traditional CGM in solving (35)
Example 5 A 20-order linear matrix equation with 119898 =
2 is also considered here following the similar procedureas Example 4 where 119876 isin S+(20) and 119860
1 1198602
isin R20times20
satisfy the condition (10) The simulation indicates that thenumerical solution needs 26 45 and 37 iterations in theRGAVS the RGACS (120578 = 074) and the traditional CGMrespectively Therefore the convergence speed of the RGAVSis also better than those of the RGACS and the CGM
5 Conclusion
Using geometric structures of manifold consisting of all sym-metric positive definite matrices the Riemannian gradientalgorithm is developed to calculate the numerical solutionfor the linear matrix equation 119876 = 119883 + sum
119898
119894=1119860119879
119894119883119860119894 In this
algorithm the geodesic distance on the curved Riemannianmanifold is taken as an objective function and the geodesiccurve is treated as the convergence path Also the variablestep sizes algorithm corresponding to the minimum valueof the objective function is provided in order to improvethe convergence speed in the constant step-size case Finallythe convergence speed of the Riemannian gradient algorithmhas been compared with the traditional conjugate gradientmethod by two simulation examples in 5-order and 20-orderlinear matrix equations respectively It is shown that theconvergence speed of the Riemannian gradient algorithmwith variable step sizes is faster than that of the conjugate gra-dient method Although the Riemannian gradient algorithmis used here for the linear matrix equation in the real numberfield it may be generalized to the complex case as well
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to express their sincere thanks tothe referees for their valuable suggestions which greatlyimproved the presentation of this paper This subject issupported by the National Natural Science Foundations ofChina (nos 61179031 10932002 and 51105033)
References
[1] M Berzig ldquoSolving a class of matrix equations via the Bhaskar-Lakshmikantham coupled fixed point theoremrdquo Applied Math-ematics Letters vol 25 no 11 pp 1638ndash1643 2012
[2] J C Engwerda ldquoOn the existence of a positive definite solutionof the matrix equation 119883 + 119860
119879
119883minus1
119860 = 119868rdquo Linear Algebra and ItsApplications vol 194 pp 91ndash108 1993
[3] W L Green and E W Kamen ldquoStabilizability of linear sys-tems over a commutative normed algebra with applications tospatially-distributed and parameter-dependent systemsrdquo SIAMJournal onControl andOptimization vol 23 no 1 pp 1ndash18 1985
[4] W Pusz and S L Woronowicz ldquoFunctional calculus forsesquilinear forms and the purification maprdquo Reports on Math-ematical Physics vol 8 no 2 pp 159ndash170 1975
[5] C-Y Chiang E K-W Chu and W-W Lin ldquoOn the 995333-Sylvester equation 119860119883 plusmn 119883
995333
119861995333
= 119862rdquo Applied Mathematics andComputation vol 218 no 17 pp 8393ndash8407 2012
[6] Z Gajic and M T J Qureshi Lyapunov Matrix Equation inSystem Stability and Control vol 195 Academic Press LondonUK 1995
[7] A C M Ran and M C B Reurings ldquoThe symmetric linearmatrix equationrdquo Electronic Journal of Linear Algebra vol 9 pp93ndash107 2002
[8] M C B Reurings Symmetric matrix equations [PhD thesis]Vrije Universiteit Amsterdam The Netherlands 2003
[9] A C M Ran and M C B Reurings ldquoA fixed point theoremin partially ordered sets and some applications to matrixequationsrdquo Proceedings of the American Mathematical Societyvol 132 no 5 pp 1435ndash1443 2004
[10] Y Su and G Chen ldquoIterative methods for solving linear matrixequation and linear matrix systemrdquo International Journal ofComputer Mathematics vol 87 no 4 pp 763ndash774 2010
[11] X Duan H Sun L Peng and X Zhao ldquoA natural gradi-ent descent algorithm for the solution of discrete algebraicLyapunov equations based on the geodesic distancerdquo AppliedMathematics and Computation vol 219 no 19 pp 9899ndash99052013
[12] X Duan H Sun and Z Zhang ldquoA natural gradient descentalgorithm for the solution of Lyapunov equations based on thegeodesic distancerdquo Journal of Computational Mathematics vol219 no 19 pp 9899ndash9905 2013
[13] M Spivak A Comprehensive Introduction to Differential Geom-etry vol 1 Publish or Perish Berkeley Calif USA 3rd edition1999
[14] S Lang Fundamentals of Differential Geometry vol 191Springer 1999
Journal of Applied Mathematics 7
[15] F Barbaresco ldquoInteractions between symmetric cones andinformation geometrics Bruhat-tits and Siegel spaces modelsfor high resolution autoregressive Doppler imageryrdquo in Emerg-ing Trends in Visual Computing vol 5416 of Lecture Notes inComputer Science pp 124ndash163 2009
[16] M Moakher ldquoA differential geometric approach to the geo-metric mean of symmetric positive-definite matricesrdquo SIAMJournal on Matrix Analysis and Applications vol 26 no 3 pp735ndash747 2005
[17] M Moakher ldquoOn the averaging of symmetric positive-definitetensorsrdquo Journal of Elasticity vol 82 no 3 pp 273ndash296 2006
[18] A Schwartzman Random ellipsoids and false discovery ratesstatistics for diffusion tensor imaging data [PhD thesis] StanfordUniversity 2006
[19] J D Lawson and Y Lim ldquoThe geometric mean matricesmetrics and morerdquo The American Mathematical Monthly vol108 no 9 pp 797ndash812 2001
[20] R Bhatia Positive definite Matrices Princeton Series in AppliedMathematics Princeton University Press Princeton NJ USA2007
[21] C Udriste Convex Functions and Optimization Methods onRiemannian Manifolds vol 297 ofMathematics and its Applica-tions Kluwer Academic Publishers Dordrecht Germany 1994
[22] M P D Carmo Riemannian Geometry Springer 1992[23] D G Luenberger ldquoThe gradient projection method along
geodesicsrdquoManagement Science vol 18 pp 620ndash631 1972
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 7
[15] F Barbaresco ldquoInteractions between symmetric cones andinformation geometrics Bruhat-tits and Siegel spaces modelsfor high resolution autoregressive Doppler imageryrdquo in Emerg-ing Trends in Visual Computing vol 5416 of Lecture Notes inComputer Science pp 124ndash163 2009
[16] M Moakher ldquoA differential geometric approach to the geo-metric mean of symmetric positive-definite matricesrdquo SIAMJournal on Matrix Analysis and Applications vol 26 no 3 pp735ndash747 2005
[17] M Moakher ldquoOn the averaging of symmetric positive-definitetensorsrdquo Journal of Elasticity vol 82 no 3 pp 273ndash296 2006
[18] A Schwartzman Random ellipsoids and false discovery ratesstatistics for diffusion tensor imaging data [PhD thesis] StanfordUniversity 2006
[19] J D Lawson and Y Lim ldquoThe geometric mean matricesmetrics and morerdquo The American Mathematical Monthly vol108 no 9 pp 797ndash812 2001
[20] R Bhatia Positive definite Matrices Princeton Series in AppliedMathematics Princeton University Press Princeton NJ USA2007
[21] C Udriste Convex Functions and Optimization Methods onRiemannian Manifolds vol 297 ofMathematics and its Applica-tions Kluwer Academic Publishers Dordrecht Germany 1994
[22] M P D Carmo Riemannian Geometry Springer 1992[23] D G Luenberger ldquoThe gradient projection method along
geodesicsrdquoManagement Science vol 18 pp 620ndash631 1972
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of