hamilton-jacobi approach to contrast functions and...
TRANSCRIPT
Hamilton-Jacobi approach to contrast functions and geodetical motion in
information geometry
Juan Manuel Pérez-Pardo
J.M. Pérez-Pardo
o j f n z g p s c g t b l j z t s l
F l o r i o M C i a g l i a c u a m
r d j e i x a d s p b i o t n u b q
a n g i m u r x t e n s a u g n w r
n i w x b l c u w m l P t S z a h m
c q j G a i o y n n l a a t v o w x
o y o i g s L l y z s t b e m l g t
V g w u b l a p F c m r t f q f x g
e j r s j h u k a l b i t a z u c i
n r w e s t d n b v f z h n t m v b
t e f p b F a b i o D i C o s m o m
r z z p y n t m o k m a j M d r p w
i p a e p w o i M t x V z a c c t d
g t v M b x l f e x g i o n l z y o
l q k a e w p s l n c t h c v k k m
i b n r r t h t e c p a p i d b s h
a k s m d r q m j n b l y n b z c i
d q D o m e n i c o F e l i c e m a
J.M. Pérez-Pardo
o j f n z g p s c g t b l j z t s l
F l o r i o M C i a g l i a c u a m
r d j e i x a d s p b i o t n u b q
a n g i m u r x t e n s a u g n w r
n i w x b l c u w m l P t S z a h m
c q j G a i o y n n l a a t v o w x
o y o i g s L l y z s t b e m l g t
V g w u b l a p F c m r t f q f x g
e j r s j h u k a l b i t a z u c i
n r w e s t d n b v f z h n t m v b
t e f p b F a b i o D i C o s m o m
r z z p y n t m o k m a j M d r p w
i p a e p w o i M t x V z a c c t d
g t v M b x l f e x g i o n l z y o
l q k a e w p s l n c t h c v k k m
i b n r r t h t e c p a p i d b s h
a k s m d r q m j n b l y n b z c i
d q D o m e n i c o F e l i c e m a
J.M. Pérez-Pardo
F l o r i o M C i a g l i a
r a
a r
n c P S
c G o a t
o i L t e
V u a F r f
e s u a i a
n e d b z n
t p F a b i o D i C o s m o
r p t o a M
i e o M V a
g M e i n
l a l t c
i r e a i
a m l n
D o m e n i c o F e l i c e
Joint work with
J.M. Pérez-Pardo
Outline
Introduction and motivation
Generating Potentials for arbitrary tensors
Hamilton-Jacobi approach to contrast functions
J.M. Pérez-Pardo
Geometrical Quantum Mechanics
Pure states are a complex projective space
Orthonormal basis ,
h̃ =hd ⌦ d ih | i � h |d i
h | i ⌦ hd | ih | i
⇡ : H ! PH⇡(| i) = | ih |
h | i{|eji} hej |eki = �jk
zj = hej | i z̄j = h |eji
h̃ =dz̄j ⌦ dzj
kzk2 � (z̄jdzj⌦)⌦ (dz̄kzk)
kzk4
J.M. Pérez-Pardo
Geometrical Quantum Mechanics
non-linear coordinateszj =
ppje
i'j
⇣pj � 0 ,
X
j
pj = 1⌘
h̃ =1
4
X
j
1
pjdpj ⌦ dpj +
⇥dpj ⌦ d'k + d'j ⌦ d'k
⇤
h̃ =dz̄j ⌦ dzj
kzk2 � (z̄jdzj⌦)⌦ (dz̄kzk)
kzk4
J.M. Pérez-Pardo
Geometrical Quantum MechanicsH ⌘ L2(�) | i ! hx | i = (x)
h |xi =
⇤(x)
Immersion of a measure space manifold into ⇥ PH (x|✓) =
pp(x|✓)ei↵(x|✓)
d ln (x|✓) = 1
2d ln p(x|✓) + id↵(x|✓)
h = g � iwg = 1
4Ep[(d ln p)⌦ (d ln p)] + Ep[d↵⌦ d↵]� Ep(d↵)⌦ Ep(d↵)
w = Ep[d ln p ^ d↵]
F = 14Ep(d ln p⌦ d ln p) Fisher-Rao metric
Ep(df) :=R�(df)pdx
J.M. Pérez-Pardo
Statistical Manifold:Each point represents a probability density
Fisher-Rao metric, a Riemannian metric:Infinitesimal “distinguishability” of probability densities
Dualistic Structure:
Usually defined by perturbation of the hermitean connection
Information Geometry
Mp 2 M
g
(g,r,r⇤)
X,Y, Z 2 X(M)
�ijk = g�ijk ± Tijk
X(g(Y, Z)) = g(rXY, Z) + g(Y,r⇤XZ)
J.M. Pérez-Pardo
Divergence Function
Statistical manifold:Divergence function:
Oriented distance-like function If is differentiable
is a generating functionGenerates tensors intrinsically
M
S : M⇥M ! R
S(x, y) 6= S(y, x)
S
@S
@x
i
���x=y
=@S
@y
i
���x=y
= 0
S(x,y) � 0
S(x,y) = 0 x = y,
S@
2S
@x
i@x
j= gij
J.M. Pérez-Pardo
Outline
Introduction and motivation
Generating Potentials for arbitrary tensors
Hamilton-Jacobi approach to contrast functions
J.M. Pérez-Pardo
Some examplesHessian Riemannian Structures[1]:
Kähler Potential
Generating Potentials for arbitrary Tensors
! =@2f
@z@z̄
f : U ! RConvex function:
Local coordinates on U ⇢ Mx = {xi}
gij(x) :=@
2f
@x
i@x
j
This is not a tensor!
Quote from Wikipedia: “There is no comparable way of describing a general Riemannian metric in terms of a single function.”
[1]: J.J. Duistermaat. On Hessian Riemannian Structures. Asian J. Math, 5 (1) (2001), pp. 79-92
J.M. Pérez-Pardo
Two copies of the manifold to define a two-point function
Provides several canonical structures:
Contrast Functions as Generating Potentials I
S : M⇥M ! R
⇡R : M⇥M ! M⇡L : M⇥M ! M
X 2 X(M)
Left-lift:Right-lift:
XL 2 X(M⇥M)
XR 2 X(M⇥M)
Diagonal embedding:d : M ,! M⇥M
p 7! (p, p)
Restriction to the diagonal:S|d := d⇤S
J.M. Pérez-Pardo
Consider the contrast function
It is clearly f-linear inLet us show that it is also f-linear in
Contrast Functions as Generating Potentials II
S : M⇥M ! R
(LXLLYLS)��d:= g(X,Y )
X
Y
LXLL(fY )LS = LXL(fLYLS)
= fLXLLYLS + LXLf · LYLS
J.M. Pérez-Pardo
Consider the contrast function
It is clearly f-linear inLet us show that it is also f-linear in
Contrast Functions as Generating Potentials II
S : M⇥M ! R
(LXLLYLS)��d:= g(X,Y )
X
Y
LXLL(fY )LS = LXL(fLYLS)
= fLXLLYLS + LXLf · LYLS
( )��d ( )
��d
( )��d
( )��d
0
= fg(X,Y )
J.M. Pérez-Pardo
Consider the contrast function
It is clearly f-linear inLet us show that it is also f-linear in
Contrast Functions as Generating Potentials II
S : M⇥M ! R
(LXLLYLS)��d:= g(X,Y )
X
Y
LXLL(fY )LS = LXL(fLYLS)
= fLXLLYLS + LXLf · LYLS
( )��d ( )
��d
( )��d
( )��d
0
= fg(X,Y )
One can define more rank 2 tensors:
(LXRLYRS)��d= (LXLLYLS)
��d= �(LXRLYLS)
��d= g(X,Y )
J.M. Pérez-Pardo
Contrast Functions as Generating Potentials III
This procedure can be used for tensors of order 3
One needs to find suitable linear combinations to cancel the non-tensorial terms
There are 8 possible combinations that lead to tensorial quantities.
They define the same symmetric tensor up to a signIn information geometry this is used to define pairs of dual connections
(LXLLYRLZRS � LXRLYLLZLS)��d= T (X,Y, Z)
or
(LXLLYLLZRS � LXRLYRLZLS)��d= T (X,Y, Z)
J.M. Pérez-Pardo
Can one go to higher order tensors?For instance the Riemann tensor
Can one generate non-symmetric tensors?
J.M. Pérez-Pardo
Outline
Introduction and motivation
Generating Potentials for arbitrary tensors
Hamilton-Jacobi approach to contrast functions
J.M. Pérez-Pardo
Find a canonical contrast functionGivenConstruct a two point function with vanishing derivatives
Hamilton’s Principal function
Solution of the equations of motion
Hamilton-Jacobi Approach to Contrast Functions
g T
�
�(0) = x �(1) = y
S[x,y] =
Z 1
0L(�(t), �̇(t))dt
S[�] =
Z tfin
tin
L(�(t), �̇(t))dt
)x = y
pi(x,y)|x=y
= 0�(t) = x
dS = pi(x,y)dxi � Pi(x,y)dy
i
Pi(x,y)|x=y
= 0
J.M. Pérez-Pardo
Hamilton-Jacobi Approach to Contrast Functions II
Choose the Lagrangian
Compute derivatives
We need the relation between andTaylor expansion
L(x, v) =1
2gijv
iv
j +1
6Tijkv
iv
jv
j
@S
@yj= gjk(y)v
kfin +
1
2Tjkl(y)v
kfinv
lfin
x,yvfin
�(t) = �(1) + vfin(1� t) +1
2v̇fin(1� t)2 +
1
6v̈fin(1� t)3
vfin = y � x+1
2v̇fin � 1
6v̈fin
Eq. of motion)v̇l = �T l
jkvkv̇j � �l
jkvjvk � 1
6glrArjksv
jvkvs
J.M. Pérez-Pardo
Assuming that is analytic in
Hamilton-Jacobi Approach to Contrast Functions III
v̇ v
�k := y
k � x
k
@S
@yj= g(y)jk�
k � 1
2�jkl(y)�
k�l +1
2Tjkl(y)�
k�l
+1
6�jkr(y)�
rls(y)�
k�l�s � 1
12Ajkls(y)�
k�l�s
@
2S
@x
i@y
j
���d= �gij(x)
@
3S
@x
l@x
k@y
j
���d= ��jkl + Tjkl
@
3S
@y
l@y
k@x
j
���d= ��jkl � TjklBy similar calculations:
J.M. Pérez-Pardo
Hamilton-Jacobi Approach to Contrast Functions IV
Find a combination of 4th order derivatives that is f-linear
Two possible linear combinations:LXLLYRLZLLWRS
��d=: LRLR
Q1(X,Y, Z,W ) =(LRRL�RLLR) + (LRRR�RLLL)
+ (LRLL�RLRR) + (LRLR�RLRL)
Q2(X,Y, Z,W ) =(LLLL�RRRR) + (LLLR�RRRL)
+ (LLRL�RRLR) + (LLRR�RRLL)
J.M. Pérez-Pardo
Hamilton-Jacobi Approach to Contrast Functions IV
Find a combination of 4th order derivatives that is f-linear
Two possible linear combinations:LXLLYRLZLLWRS
��d=: LRLR
Q1(X,Y, Z,W ) =(LRRL�RLLR) + (LRRR�RLLL)
+ (LRLL�RLRR) + (LRLR�RLRL)
Q2(X,Y, Z,W ) =(LLLL�RRRR) + (LLLR�RRRL)
+ (LLRL�RRLR) + (LLRR�RRLL)
Q1 = Q2 = 0
After a lengthy calculation:
J.M. Pérez-Pardo
Hamilton-Jacobi Approach to Contrast Functions V
Choose a lagrangian
L(x, v) =1
2gijv
iv
j +1
6Tijkv
iv
jv
j +1
24Cijklv
iv
jv
jv
l
J.M. Pérez-Pardo
Hamilton-Jacobi Approach to Contrast Functions V
Choose a lagrangian
L(x, v) =1
2gijv
iv
j +1
6Tijkv
iv
jv
j +1
24Cijklv
iv
jv
jv
l
Q1 = Q2 = 0
After a lengthy calculation:
J.M. Pérez-Pardo
Hamilton-Jacobi Approach to Contrast Functions V
Choose a lagrangian
L(x, v) =1
2gijv
iv
j +1
6Tijkv
iv
jv
j +1
24Cijklv
iv
jv
jv
l
Q1 = Q2 = 0
After a lengthy calculation:
What about non-symmetric tensors?
J.M. Pérez-Pardo
References
F.M. Ciaglia, F. di Cosmo, D. Felice, S. Mancini, G. Marmo and J.M. Pérez-Pardo. Hamilton-Jacobi approach to Potential Functions in Information Geometry. J. Math. Phys. 58, 063506. (2017).
P. Facchi, R. Kulkarni, V.I. Man’ko, G. Marmo, E.C.G. Sudarshan and F. Ventriglia. Classical and quantum Fisher information in the geometrical formulation of quantum mechanics. Phys. Lett. A 374, (2010), 4801-4803