Lecture 5: Wavefront reconstruction and prediction

Rufus Fraanje

[email protected]
TU Delft Delft Center for Systems and Control

May 30, 2011

1 / 37

Exercise
Study the relation between the Strehl ratio, the Fried parameter r0 and the
telescope diameter D.
Do this by means of a computer program, that:
1

an

Generates random wavefront phase aberrations satisfying the von Karm

model, choose L0 = 10m, r0 = 0.2, . . . , 1m, and D = 2, . . . , 20m.
Make sure the grid is of sufficient size and resolution.

Compute the corresponding point spread function;

Compute the Strehl ratio

S =

I (0, 0)
Io (0, 0)

where I (0, 0) and Io (0, 00 the intensity at the intersection of the optical axis
with the image plane with and without phase aberration.
4

Average your results over sufficient amount of realizations of the wavefront

phase.

Make graphs of the Strehl ratio S versus the Fried parameter r0 for several
choices of the telescope diameter D. How can you explain the results?
2 / 37

Matlab example
L0
= 10;
D list = [2:2:20];
r0 list = [0.1:0.1:1];
nD
nr0
Nav
Strehl

=
=
=
=

%[m] o u t e r s c a l e o f t u r b u l e n c e
% diameter telescope
% F r i e d parameter

length ( D l i s t ) ;
length ( r 0 l i s t ) ;
500; % number o f r e a l i z a t i o n f o r a v e r a g i n g
zeros ( nr0 , nD , Nav ) ; % a r r a y f o r S t r e h l r a t i o s

f o r iD = 1 : nD ,
D = D l i s t ( iD ) ;
f o r i r 0 = 1 : nr0 ,
r0 = r 0 l i s t ( i r 0 ) ;

% telescope diameter
% F r i e d parameter

r a d i u s = min (max( 1 6 , c e i l (2 D/ r 0 ) ) , 4 0 ) ; % # p o i n t s along r a d i u s

...
3 / 37

Matlab example
Corr = vonkarmancorr space ( x , y , r0 , L0 ) ; % compute c o r r . c o e f f s
% Cholesky f a c t o r i z . n
C o r r s q r t = chol ( Corr ) ;
...
% Compute u n d i s t o r t e d wave over c i r c u l a r t e l e s c o p e g r i d :
Af = c i r c s h i f t ( f f t 2 (A) , ( p +1) r a d i u s [ 1 1 ] ) ;
% i n t e n s i t y at center :
I 0 = abs ( Af ( ( p +1) r a d i u s + 1 , ( p +1) r a d i u s + 1 ) ) 2 ;
f o r i a v = 1 : Nav ,
% Generate a square random phase screen
p h i = C o r r s q r t randn ( ( 2 r a d i u s + 1 ) 2 , 1 ) ;
. . . % place phi at center of grid , i . e .

f i l l Phi

4 / 37

Matlab example
% Compute t h e wave over t h e c i r c u l a r t e l e s c o p e g r i d :
Ap = A. exp ( s q r t ( 1) Phi ) ;
% perform t h e f o u r i e r t r a n s f o r m o f t h e d i s t o r t e d wave :
Apf = c i r c s h i f t ( f f t 2 ( Ap) , ( p +1) r a d i u s [ 1 1 ] ) ;
% i n t e n s i t y at center :
I p = abs ( Apf ( ( p +1) r a d i u s+1av : ( p +1) r a d i u s +1+av , . . .
( p +1) r a d i u s+1av : ( p +1) r a d i u s +1+av ) ) 2 ;
S t r e h l ( i r 0 , iD , i a v ) = I p / I 0 ;
end ; % number o f averages Nav
end ; % number o f F r i e d parameters nr0
end ; % number o f t e l e s c o p e d i a m e t e r s nD
mean( S t r e h l , 3 ) , % average over r e a l i z a t i o n s
5 / 37

Matlab example
0

10

D=2m
D=4m
D=6m
D=8m
D=10m
D=12m
D=14m
D=16m
D=18m
D=20m

Strehl ratio [.]

10

10

10

10

10

Fried parameter r [m]

10

6 / 37

Matlab example
Strehl ratio

Strehl ratio tip-tilt compensated

10

Strehl ratio [.]

Strehl ratio [.]

10

10

10

D=0.10m
D=0.20m
D=0.30m
D=0.40m
D=0.50m
D=0.60m
D=0.70m
D=0.80m
D=0.90m
D=1.00m

10

10

Fried parameter r0 [m]

10

10

10

Fried parameter r0 [m]

10

7 / 37

Outline

Preliminaries: derivatives of functions wrt. matrices;

Wavefront reconstruction from pixel intensities;

Wavefront reconstruction from local gradients;

Wavefront prediction;

Current research.

8 / 37

Outline

Preliminaries: derivatives of functions wrt. matrices;

Wavefront reconstruction from pixel intensities;

Wavefront reconstruction from local gradients;

Wavefront prediction;

Current research.

8 / 37

Outline

Preliminaries: derivatives of functions wrt. matrices;

Wavefront reconstruction from pixel intensities;

Wavefront reconstruction from local gradients;

Wavefront prediction;

Current research.

8 / 37

Outline

Preliminaries: derivatives of functions wrt. matrices;

Wavefront reconstruction from pixel intensities;

Wavefront reconstruction from local gradients;

Wavefront prediction;

Current research.

8 / 37

Outline

Preliminaries: derivatives of functions wrt. matrices;

Wavefront reconstruction from pixel intensities;

Wavefront reconstruction from local gradients;

Wavefront prediction;

Current research.

8 / 37

Preliminaries
Lemma (1)

d
tr AXB T = AT B
dX

Proof.
Note that
tr AXB T

XXX
k

aki xij bkj

hence

d
tr AXB T
dXij
where A =

a1

an

,B=

b1

d X
aki bkj
dXij
k

= aiT bj

bm

9 / 37

Preliminaries
Lemma (1)

d
tr AXB T = AT B
dX

Proof.
Note that
tr AXB T

XXX
k

aki xij bkj

hence

d
tr AXB T
dXij
where A =

a1

an

,B=

b1

d X
aki bkj
dXij
k

= aiT bj

bm

9 / 37

Preliminaries
Lemma (1)

d
tr AXB T = AT B
dX

Proof.
Note that
tr AXB T

XXX
k

aki xij bkj

hence

d
tr AXB T
dXij
where A =

a1

an

,B=

b1

d X
aki bkj
dXij
k

= aiT bj

bm

9 / 37

Preliminaries
Lemma (1)

d
tr AXB T = AT B
dX

Proof.
Note that
tr AXB T

XXX
k

aki xij bkj

hence

d
tr AXB T
dXij
where A =

a1

an

,B=

b1

d X
aki bkj
dXij
k

= aiT bj

bm

9 / 37

Preliminaries
Lemma (1)

d
tr AXB T = AT B
dX

Proof.
Note that
tr AXB T

XXX
k

aki xij bkj

hence

d
tr AXB T
dXij
where A =

a1

an

,B=

b1

d X
aki bkj
dXij
k

= aiT bj

bm

9 / 37

Preliminaries
Lemma (2)
d T T
tr AX B
= BT A
dX

Lemma (3)

d
tr AXBX T C = AT C T XB T + CAXB
dX

Proof.
Use product rule:
dtr AXBX T C

= tr AdXBX T C + tr AXBdX T C

then evaluate each derivative using previous two lemmas.

10 / 37

Preliminaries
Lemma (2)
d T T
tr AX B
= BT A
dX

Lemma (3)

d
tr AXBX T C = AT C T XB T + CAXB
dX

Proof.
Use product rule:
dtr AXBX T C

= tr AdXBX T C + tr AXBdX T C

then evaluate each derivative using previous two lemmas.

10 / 37

Preliminaries
Lemma (2)
d T T
tr AX B
= BT A
dX

Lemma (3)

d
tr AXBX T C = AT C T XB T + CAXB
dX

Proof.
Use product rule:
dtr AXBX T C

= tr AdXBX T C + tr AXBdX T C

then evaluate each derivative using previous two lemmas.

10 / 37

Preliminaries
Theorem (Weighted least squares)
Let V = V T and W = W T and

J (X ) = tr V C AXB W C AXB
then

dJ (X )
= 2AT V C AXB WB T
dX

and dJ (X )/dX = 0 if and only if X satisfies the (generalized) normal equations:

AT VA X BWB T

AT V C WB T

Proof.
The proof follows by straightforward use of the previous lemmas.
Exercise: verify this.
11 / 37

Preliminaries
Theorem (Weighted least squares)
Let V = V T and W = W T and

J (X ) = tr V C AXB W C AXB
then

dJ (X )
= 2AT V C AXB WB T
dX

and dJ (X )/dX = 0 if and only if X satisfies the (generalized) normal equations:

AT VA X BWB T

AT V C WB T

Proof.
The proof follows by straightforward use of the previous lemmas.
Exercise: verify this.
11 / 37

Preliminaries
Theorem (Weighted least squares)
Let V = V T and W = W T and

J (X ) = tr V C AXB W C AXB
then

dJ (X )
= 2AT V C AXB WB T
dX

and dJ (X )/dX = 0 if and only if X satisfies the (generalized) normal equations:

AT VA X BWB T

AT V C WB T

Proof.
The proof follows by straightforward use of the previous lemmas.
Exercise: verify this.
11 / 37

Preliminaries
Theorem (Weighted least squares)
Let V = V T and W = W T and

J (X ) = tr V C AXB W C AXB
then

dJ (X )
= 2AT V C AXB WB T
dX

and dJ (X )/dX = 0 if and only if X satisfies the (generalized) normal equations:

AT VA X BWB T

AT V C WB T

Proof.
The proof follows by straightforward use of the previous lemmas.
Exercise: verify this.
11 / 37

Wavefront reconstruction from pixel intensities

Problem of wavefront reconstruction: how to get (, ) from i (x , y )?
Spatial impulse response given by:

s(x , y ) = F 1 a(, )ej (,)

Obtained image (incoherent imaging case):

i (x , y ) = o(x , y ) |s(x , y )|2 + n(x , y )

12 / 37

Wavefront reconstruction from pixel intensities

Problem of wavefront reconstruction: how to get (, ) from i (x , y )?
Spatial impulse response given by:

s(x , y ) = F 1 a(, )ej (,)

Obtained image (incoherent imaging case):

i (x , y ) = o(x , y ) |s(x , y )|2 + n(x , y )

12 / 37

Wavefront reconstruction from pixel intensities

Problem of wavefront reconstruction: how to get (, ) from i (x , y )?
Spatial impulse response given by:

s(x , y ) = F 1 a(, )ej (,)

Obtained image (incoherent imaging case):

i (x , y ) = o(x , y ) |s(x , y )|2 + n(x , y )

12 / 37

Wavefront reconstruction from pixel intensities

Problem of wavefront reconstruction: how to get (, ) from i (x , y )?
Spatial impulse response given by:

s(x , y ) = F 1 a(, )ej (,)

Obtained image (incoherent imaging case):

i (x , y ) = o(x , y ) |s(x , y )|2 + n(x , y )

Pupil plane

F 1 {.}
F{.}

Image plane
x

y
z

12 / 37

Wavefront reconstruction from pixel intensities

Problem of wavefront reconstruction: how to get (, ) from |s(x , y )| knowing
a(, )?
Wavefront (, )

PSF |s(x , y )| = F 1 a(, )ej (,)

phase-retrieval problem
13 / 37

Wavefront reconstruction from pixel intensities

Problem of wavefront reconstruction: how to get (, ) from |s(x , y )| knowing
a(, )?
Wavefront (, )

PSF |s(x , y )| = F 1 a(, )ej (,)

phase-retrieval problem
13 / 37

Wavefront reconstruction from pixel intensities

Methods for point objects:
Method 1: Solve nonlinear optimization problem, e.g., using
Gerchberg-Saxton iterations1 , 2 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
peak-detection or center of mass;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.
Other: wavefront sensorless methods3 , . . .
1 Gerchberg&Saxton,

Optik, 237 (1972)

Applied Optics, 21(15), (1982)
3 c.f. Song et al, Optics Express, 23(18), (2010)
2 Fienup,

14 / 37

Wavefront reconstruction from pixel intensities

Methods for point objects:
Method 1: Solve nonlinear optimization problem, e.g., using
Gerchberg-Saxton iterations1 , 2 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
peak-detection or center of mass;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.
Other: wavefront sensorless methods3 , . . .
1 Gerchberg&Saxton,

Optik, 237 (1972)

Applied Optics, 21(15), (1982)
3 c.f. Song et al, Optics Express, 23(18), (2010)
2 Fienup,

14 / 37

Wavefront reconstruction from pixel intensities

Methods for point objects:
Method 1: Solve nonlinear optimization problem, e.g., using
Gerchberg-Saxton iterations1 , 2 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
peak-detection or center of mass;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.
Other: wavefront sensorless methods3 , . . .
1 Gerchberg&Saxton,

Optik, 237 (1972)

Applied Optics, 21(15), (1982)
3 c.f. Song et al, Optics Express, 23(18), (2010)
2 Fienup,

14 / 37

Wavefront reconstruction from pixel intensities

Methods for point objects:
Method 1: Solve nonlinear optimization problem, e.g., using
Gerchberg-Saxton iterations1 , 2 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
peak-detection or center of mass;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.
Other: wavefront sensorless methods3 , . . .
1 Gerchberg&Saxton,

Optik, 237 (1972)

Applied Optics, 21(15), (1982)
3 c.f. Song et al, Optics Express, 23(18), (2010)
2 Fienup,

14 / 37

Wavefront reconstruction from pixel intensities

Methods for point objects:
Method 1: Solve nonlinear optimization problem, e.g., using
Gerchberg-Saxton iterations1 , 2 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
peak-detection or center of mass;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.
Other: wavefront sensorless methods3 , . . .
1 Gerchberg&Saxton,

Optik, 237 (1972)

Applied Optics, 21(15), (1982)
3 c.f. Song et al, Optics Express, 23(18), (2010)
2 Fienup,

14 / 37

Wavefront reconstruction from pixel intensities

Methods for point objects:
Method 1: Solve nonlinear optimization problem, e.g., using
Gerchberg-Saxton iterations1 , 2 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
peak-detection or center of mass;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.
Other: wavefront sensorless methods3 , . . .
1 Gerchberg&Saxton,

Optik, 237 (1972)

Applied Optics, 21(15), (1982)
3 c.f. Song et al, Optics Express, 23(18), (2010)
2 Fienup,

14 / 37

Wavefront reconstruction from pixel intensities

For extended objects:
Method 1: Solve nonlinear optimization problem to determine both wavefront
phase and object, e.g., using phase diversity4 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
correlation tracking;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.

4 Gonsalves,

Opt. Eng., 21, (1982)

15 / 37

Wavefront reconstruction from pixel intensities

For extended objects:
Method 1: Solve nonlinear optimization problem to determine both wavefront
phase and object, e.g., using phase diversity4 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
correlation tracking;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.

4 Gonsalves,

Opt. Eng., 21, (1982)

15 / 37

Wavefront reconstruction from pixel intensities

For extended objects:
Method 1: Solve nonlinear optimization problem to determine both wavefront
phase and object, e.g., using phase diversity4 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
correlation tracking;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.

4 Gonsalves,

Opt. Eng., 21, (1982)

15 / 37

Wavefront reconstruction from pixel intensities

For extended objects:
Method 1: Solve nonlinear optimization problem to determine both wavefront
phase and object, e.g., using phase diversity4 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
correlation tracking;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.

4 Gonsalves,

Opt. Eng., 21, (1982)

15 / 37

Wavefront reconstruction from pixel intensities

For extended objects:
Method 1: Solve nonlinear optimization problem to determine both wavefront
phase and object, e.g., using phase diversity4 ;
Method 2: Use wavefront sensor, e.g., Shack-Hartmann, and determine
wavefront by:
For each lenslet/sub-aperture determine local tip-tilt by
correlation tracking;
For each lenslet/sub-aperture determine local wavefront by
Method 1;
Determine global wavefront by Method 1 using information on
geometry of lenslet array.

4 Gonsalves,

Opt. Eng., 21, (1982)

15 / 37

Wavefront reconstruction from pixel intensities

Spot shift by center of mass:

RR
x I (x , y )dxdy
RR
I (x , y )dxdy
RR
y I (x , y )dxdy
RR
I (x , y )dxdy

Wavefront phase (small angles sin() tan() ):

x
y

x /f
y /f

where f the focal distance, such that

(, )

x + y

16 / 37

Wavefront reconstruction from pixel intensities

Spot shift by center of mass:

RR
x I (x , y )dxdy
RR
I (x , y )dxdy
RR
y I (x , y )dxdy
RR
I (x , y )dxdy

Wavefront phase (small angles sin() tan() ):

x
y

x /f
y /f

where f the focal distance, such that

(, )

x + y

16 / 37

Wavefront reconstruction from pixel intensities

Spot shift by center of mass:

RR
x I (x , y )dxdy
RR
I (x , y )dxdy
RR
y I (x , y )dxdy
RR
I (x , y )dxdy

Wavefront phase (small angles sin() tan() ):

x
y

x /f
y /f

where f the focal distance, such that

(, )

x + y

16 / 37

Wavefront reconstruction from pixel intensities

Spot shift by center of mass:

RR
x I (x , y )dxdy
RR
I (x , y )dxdy
RR
y I (x , y )dxdy
RR
I (x , y )dxdy

Wavefront phase (small angles sin() tan() ):

x
y

x /f
y /f

where f the focal distance, such that

(, )

x + y

16 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
Input: |s(x , y )| and a(, )

)
Output: s(x , y ) and (,
1:

s(x , y ) |s(x , y )|

2:

while not converged do

3:

(, ) F (s(x , y ))
S

4:

(, ) a(, )ej S (,)

S

5:

(, )
s(x , y ) F 1 S

6:

s(x , y ) |s(x , y )|ej s(x ,y )

7:

end while

8:

(, )
) = S
return (,
17 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
Input: |s(x , y )| and a(, )

)
Output: s(x , y ) and (,
1:

s(x , y ) |s(x , y )|

2:

while not converged do

3:

(, ) F (s(x , y ))
S

4:

(, ) a(, )ej S (,)

S

5:

(, )
s(x , y ) F 1 S

6:

s(x , y ) |s(x , y )|ej s(x ,y )

7:

end while

8:

(, )
) = S
return (,
17 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
Input: |s(x , y )| and a(, )

)
Output: s(x , y ) and (,
1:

s(x , y ) |s(x , y )|

2:

while not converged do

3:

(, ) F (s(x , y ))
S

4:

(, ) a(, )ej S (,)

S

5:

(, )
s(x , y ) F 1 S

6:

s(x , y ) |s(x , y )|ej s(x ,y )

7:

end while

8:

(, )
) = S
return (,
17 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
Input: |s(x , y )| and a(, )

)
Output: s(x , y ) and (,
1:

s(x , y ) |s(x , y )|

2:

while not converged do

3:

(, ) F (s(x , y ))
S

4:

(, ) a(, )ej S (,)

S

5:

(, )
s(x , y ) F 1 S

6:

s(x , y ) |s(x , y )|ej s(x ,y )

7:

end while

8:

(, )
) = S
return (,
17 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
Input: |s(x , y )| and a(, )

)
Output: s(x , y ) and (,
1:

s(x , y ) |s(x , y )|

2:

while not converged do

3:

(, ) F (s(x , y ))
S

4:

(, ) a(, )ej S (,)

S

5:

(, )
s(x , y ) F 1 S

6:

s(x , y ) |s(x , y )|ej s(x ,y )

7:

end while

8:

(, )
) = S
return (,
17 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
Input: |s(x , y )| and a(, )

)
Output: s(x , y ) and (,
1:

s(x , y ) |s(x , y )|

2:

while not converged do

3:

(, ) F (s(x , y ))
S

4:

(, ) a(, )ej S (,)

S

5:

(, )
s(x , y ) F 1 S

6:

s(x , y ) |s(x , y )|ej s(x ,y )

7:

end while

8:

(, )
) = S
return (,
17 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
Input: |s(x , y )| and a(, )

)
Output: s(x , y ) and (,
1:

s(x , y ) |s(x , y )|

2:

while not converged do

3:

(, ) F (s(x , y ))
S

4:

(, ) a(, )ej S (,)

S

5:

(, )
s(x , y ) F 1 S

6:

s(x , y ) |s(x , y )|ej s(x ,y )

7:

end while

8:

(, )
) = S
return (,
17 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
Input: |s(x , y )| and a(, )

)
Output: s(x , y ) and (,
1:

s(x , y ) |s(x , y )|

2:

while not converged do

3:

(, ) F (s(x , y ))
S

4:

(, ) a(, )ej S (,)

S

5:

(, )
s(x , y ) F 1 S

6:

s(x , y ) |s(x , y )|ej s(x ,y )

7:

end while

8:

(, )
) = S
return (,
17 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
an
turbulence: L0 = 10m, D = 0.05m
Simulation with 10 realizations of von Karm
and r0 = 0.60m, with Hartmann-mask to get more amplitude constraints
40
35
0.05

30

0.045

2
Error ||S S||

0.04

Yaxis [m]

0.035
0.03
0.025

zerophase initial condition

25
20
center of mass determined
intial condition

15

0.02

10

0.015
0.01

0.005
0
0

0.01

0.02
0.03
Xaxis [m]

0.04

0.05

0
0

10

15
Iteration nr. [.]

20

25

30

18 / 37

Wavefront reconstruction from pixel intensities

Gerchberg-Saxton algorithm:
an
turbulence: L0 = 10m, D = 0.05m
Simulation with 10 realizations of von Karm
and r0 = 0.60m, with Hartmann-mask to get more amplitude constraints
Gerchberg-Saxton

Center of mass

0.05

0.05

0.045

0.045

0.045

0.04

0.04

0.04

0.035

0.035

0.035

0.025
0.02
0.015

0.03

Yaxis [m]

0.03

Yaxis [m]

Yaxis [m]

True wavefront
0.05

0.025
0.02
0.015

0.03
0.025
0.02
0.015

0.01

0.01

0.01

0.005

0.005

0.005

0
0

0.01

0.02
0.03
Xaxis [m]

0.04

0.05

0
0

0.01

0.02
0.03
Xaxis [m]

0.04

0.05

0
0

0.01

0.02
0.03
Xaxis [m]

0.04

0.05

18 / 37

Wavefront reconstruction from pixel intensities

Comparison Center-of-mass (COM) & Gerchberg-Saxton (GS)

COM

GS

Complexity:

O (N )

O (N 2 log(N )) per iteration

Accuracy:

only tip-tilt

depending on initialization

Application:

zonal & global

zonal & global

19 / 37

Wavefront reconstruction from local gradients

Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j ) + nx (i , j )
= (i , j + 1) (i , j ) + ny (i , j )

and all stacked in vectors

t

= G + n

Reconstruction:
1

Solve by line integration (neglecting or filtering noise),

Solve by inversion (or estimation) in Fourier domain,

3
4

Least squares: minimize kt Gk22 ,

Linear estimation:
n random variable with known statistics,
n and
`
random variables with known statistics.
20 / 37

Wavefront reconstruction from local gradients

Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j ) + nx (i , j )
= (i , j + 1) (i , j ) + ny (i , j )

and all stacked in vectors

t

= G + n

Reconstruction:
1

Solve by line integration (neglecting or filtering noise),

Solve by inversion (or estimation) in Fourier domain,

3
4

Least squares: minimize kt Gk22 ,

Linear estimation:
n random variable with known statistics,
n and
`
random variables with known statistics.
20 / 37

Wavefront reconstruction from local gradients

Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j ) + nx (i , j )
= (i , j + 1) (i , j ) + ny (i , j )

and all stacked in vectors

t

= G + n

Reconstruction:
1

Solve by line integration (neglecting or filtering noise),

Solve by inversion (or estimation) in Fourier domain,

3
4

Least squares: minimize kt Gk22 ,

Linear estimation:
n random variable with known statistics,
n and
`
random variables with known statistics.
20 / 37

Wavefront reconstruction from local gradients

Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j ) + nx (i , j )
= (i , j + 1) (i , j ) + ny (i , j )

and all stacked in vectors

t

= G + n

Reconstruction:
1

Solve by line integration (neglecting or filtering noise),

Solve by inversion (or estimation) in Fourier domain,

3
4

Least squares: minimize kt Gk22 ,

Linear estimation:
n random variable with known statistics,
n and
`
random variables with known statistics.
20 / 37

Wavefront reconstruction from local gradients

Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j ) + nx (i , j )
= (i , j + 1) (i , j ) + ny (i , j )

and all stacked in vectors

t

= G + n

Reconstruction:
1

Solve by line integration (neglecting or filtering noise),

Solve by inversion (or estimation) in Fourier domain,

3
4

Least squares: minimize kt Gk22 ,

Linear estimation:
n random variable with known statistics,
n and
`
random variables with known statistics.
20 / 37

Wavefront reconstruction from local gradients

Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j ) + nx (i , j )
= (i , j + 1) (i , j ) + ny (i , j )

and all stacked in vectors

t

= G + n

Reconstruction:
1

Solve by line integration (neglecting or filtering noise),

Solve by inversion (or estimation) in Fourier domain,

3
4

Least squares: minimize kt Gk22 ,

Linear estimation:
n random variable with known statistics,
n and
`
random variables with known statistics.
20 / 37

Wavefront reconstruction from local gradients

Fourier domain method
Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j )
= (i , j + 1) (i , j )

after fourier transform (FFT)a :

Tx (x , y )
Ty (x , y )

e j x 1
e j y 1

(x , y )

Solve for each (discrete) frequency (x , y ) 6= (0, 0):

x , y ) =
(
a Poyneer

et al, JOSA, 19(10), (2002)

21 / 37

Wavefront reconstruction from local gradients

Fourier domain method
Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j )
= (i , j + 1) (i , j )

after fourier transform (FFT)a :

Tx (x , y )
Ty (x , y )

e j x 1
e j y 1

(x , y )

Solve for each (discrete) frequency (x , y ) 6= (0, 0):

x , y ) =
(
a Poyneer

et al, JOSA, 19(10), (2002)

21 / 37

Wavefront reconstruction from local gradients

Fourier domain method
Local tilt measurements:
tx (i , j )
ty (i , j )

= (i + 1, j ) (i , j )
= (i , j + 1) (i , j )

after fourier transform (FFT)a :

Tx (x , y )
Ty (x , y )

e j x 1
e j y 1

(x , y )

Solve for each (discrete) frequency (x , y ) 6= (0, 0):

x , y )
(
a Poyneer

e j x 1
e j y 1

e j x 1
e j y 1

!1

e j x 1
e j y 1

Tx (x , y )
Ty (x , y )

et al, JOSA, 19(10), (2002)

21 / 37

Wavefront reconstruction from local gradients

Fourier domain method
Local tilt measurements:
tx (i , j )

= (i + 1, j ) (i , j )

ty (i , j )

= (i , j + 1) (i , j )

after fourier transform (FFT)a :

Tx (x , y )
Ty (x , y )

e j x 1
e j y 1

(x , y )

Solve for each (discrete) frequency (x , y ) 6= (0, 0):

x , y ) =
(
a Poyneer

(ej x 1) (ej y 1)
4(sin2 (x /2) + sin2 (y /2))

Tx (x , y )
Ty (x , y )

et al, JOSA, 19(10), (2002)

21 / 37

Wavefront reconstruction from local gradients

Least squares
Recall stacked tilt measurements
t = G + n
Determine by minimizing squared errors

J ()

By theorem on slide 11:

T (t G)

(t G)

t G)
T
= tr (t G)(
=

d = 0
dJ ()/

GT G = GT t

If GT G is invertible, solution is given by:

= (GT G)1 GT t
22 / 37

Wavefront reconstruction from local gradients

Least squares
Recall stacked tilt measurements
t = G + n
Determine by minimizing squared errors

J ()

By theorem on slide 11:

T (t G)

(t G)

t G)
T
= tr (t G)(
=

d = 0
dJ ()/

GT G = GT t

If GT G is invertible, solution is given by:

= (GT G)1 GT t
22 / 37

Wavefront reconstruction from local gradients

Least squares
Recall stacked tilt measurements
t = G + n
Determine by minimizing squared errors

J ()

By theorem on slide 11:

T (t G)

(t G)

t G)
T
= tr (t G)(
=

d = 0
dJ ()/

GT G = GT t

If GT G is invertible, solution is given by:

= (GT G)1 GT t
22 / 37

Wavefront reconstruction from local gradients

Least squares
Recall stacked tilt measurements
t = G + n
Determine by minimizing squared errors

J ()

By theorem on slide 11:

T (t G)

(t G)

t G)
T
= tr (t G)(
=

d = 0
dJ ()/

GT G = GT t

If GT G is invertible, solution is given by:

= (GT G)1 GT t
22 / 37

Wavefront reconstruction from local gradients

Least squares
Recall stacked tilt measurements
t = G + n
Determine by minimizing squared errors

J ()

By theorem on slide 11:

T (t G)

(t G)

t G)
T
= tr (t G)(
=

d = 0
dJ ()/

GT G = GT t

If GT G is invertible, solution is given by:

= (GT G)1 GT t
22 / 37

Wavefront reconstruction from local gradients

Least squares
However, usually G1 = 0 such that GT G is singular!
Then there are infinite solutions, one possible solution:

= G+ t
with G+ the Moore-Penrose pseudoinverse of G (Exercise: verify this.)

Moore-Penrose pseudoinverse
Let the singular value decomposition of G be given by
G =

0
0

VT
T
V

with a diagonal matrix with strictly positive elements on its diagonal.

Then the Moore-pseudoinverse G+ of G is given by
G+ = V 1 U T

23 / 37

Wavefront reconstruction from local gradients

Least squares
However, usually G1 = 0 such that GT G is singular!
Then there are infinite solutions, one possible solution:

= G+ t
with G+ the Moore-Penrose pseudoinverse of G (Exercise: verify this.)

Moore-Penrose pseudoinverse
Let the singular value decomposition of G be given by
G =

0
0

VT
T
V

with a diagonal matrix with strictly positive elements on its diagonal.

Then the Moore-pseudoinverse G+ of G is given by
G+ = V 1 U T

23 / 37

Wavefront reconstruction from local gradients

Least squares
However, usually G1 = 0 such that GT G is singular!
Then there are infinite solutions, one possible solution:

= G+ t
with G+ the Moore-Penrose pseudoinverse of G (Exercise: verify this.)

Moore-Penrose pseudoinverse
Let the singular value decomposition of G be given by
G =

0
0

VT
T
V

with a diagonal matrix with strictly positive elements on its diagonal.

Then the Moore-pseudoinverse G+ of G is given by
G+ = V 1 U T

23 / 37

Wavefront reconstruction from local gradients

Least squares
However, usually G1 = 0 such that GT G is singular!
Then there are infinite solutions, one possible solution:

= G+ t
with G+ the Moore-Penrose pseudoinverse of G (Exercise: verify this.)

Moore-Penrose pseudoinverse
Let the singular value decomposition of G be given by
G =

0
0

VT
T
V

with a diagonal matrix with strictly positive elements on its diagonal.

Then the Moore-pseudoinverse G+ of G is given by
G+ = V 1 U T

23 / 37

Wavefront reconstruction from local gradients

Least squares
However, usually G1 = 0 such that GT G is singular!
Then there are infinite solutions, one possible solution:

= G+ t
with G+ the Moore-Penrose pseudoinverse of G (Exercise: verify this.)

Moore-Penrose pseudoinverse
Let the singular value decomposition of G be given by
G =

0
0

VT
T
V

with a diagonal matrix with strictly positive elements on its diagonal.

Then the Moore-pseudoinverse G+ of G is given by
G+ = V 1 U T

23 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume n is zero-mean random variable with Gaussian distribution
p(n) =

T 1
1
en Rn n/2
1
/
2
|2 Rn |

Hence, distribution of t = G + n given by

p (t ) =

T 1
1
e(t G) Rn (t G)/2
1
/
2
|2 Rn |

Maximize likelihood of measurement t, by minimizing:

J ()

By theorem on slide 11:

T Rn1 (t G)

(t G)

t G)
T
= tr Rn1 (t G)(
=

d = 0
dJ ()/

GT Rn1 G = GT Rn1 t

24 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume n is zero-mean random variable with Gaussian distribution
p(n) =

T 1
1
en Rn n/2
1
/
2
|2 Rn |

Hence, distribution of t = G + n given by

p (t ) =

T 1
1
e(t G) Rn (t G)/2
1
/
2
|2 Rn |

Maximize likelihood of measurement t, by minimizing:

J ()

By theorem on slide 11:

T Rn1 (t G)

(t G)

t G)
T
= tr Rn1 (t G)(
=

d = 0
dJ ()/

GT Rn1 G = GT Rn1 t

24 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume n is zero-mean random variable with Gaussian distribution
p(n) =

T 1
1
en Rn n/2
1
/
2
|2 Rn |

Hence, distribution of t = G + n given by

p (t ) =

T 1
1
e(t G) Rn (t G)/2
1
/
2
|2 Rn |

Maximize likelihood of measurement t, by minimizing:

J ()

By theorem on slide 11:

T Rn1 (t G)

(t G)

t G)
T
= tr Rn1 (t G)(
=

d = 0
dJ ()/

GT Rn1 G = GT Rn1 t

24 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume n is zero-mean random variable with Gaussian distribution
p(n) =

T 1
1
en Rn n/2
1
/
2
|2 Rn |

Hence, distribution of t = G + n given by

p (t ) =

T 1
1
e(t G) Rn (t G)/2
1
/
2
|2 Rn |

Maximize likelihood of measurement t, by minimizing:

J ()

By theorem on slide 11:

T Rn1 (t G)

(t G)

t G)
T
= tr Rn1 (t G)(
=

d = 0
dJ ()/

GT Rn1 G = GT Rn1 t

24 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume n is zero-mean random variable with Gaussian distribution
p(n) =

T 1
1
en Rn n/2
1
/
2
|2 Rn |

Hence, distribution of t = G + n given by

p (t ) =

T 1
1
e(t G) Rn (t G)/2
1
/
2
|2 Rn |

Maximize likelihood of measurement t, by minimizing:

J ()

By theorem on slide 11:

T Rn1 (t G)

(t G)

t G)
T
= tr Rn1 (t G)(
=

d = 0
dJ ()/

GT Rn1 G = GT Rn1 t

24 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume is also zero-mean random variable with Gaussian distribution
p() =

T 1
1
e R /2
1
/
2
|2 R |

and independent of n.
Suppose is given by a linear estimator

= Lt
where L minimizes
J (L)

T
= E tr( )(
)

= E tr( L(G + n))( L(G + n))T

T nT
I LG L
I LG
= trE
n nnT

25 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume is also zero-mean random variable with Gaussian distribution
p() =

T 1
1
e R /2
1
/
2
|2 R |

and independent of n.
Suppose is given by a linear estimator

= Lt
where L minimizes
J (L)

T
= E tr( )(
)

= E tr( L(G + n))( L(G + n))T

T nT
I LG L
I LG
= trE
n nnT

25 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume is also zero-mean random variable with Gaussian distribution
p() =

T 1
1
e R /2
1
/
2
|2 R |

and independent of n.
Suppose is given by a linear estimator

= Lt
where L minimizes
J (L)

T
= E tr( )(
)

= E tr( L(G + n))( L(G + n))T

T nT
I LG L
I LG
= trE
n nnT

25 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume is also zero-mean random variable with Gaussian distribution
p() =

T 1
1
e R /2
1
/
2
|2 R |

and independent of n.
Suppose is given by a linear estimator

= Lt
where L minimizes
J (L)

T
= E tr( )(
)

= E tr( L(G + n))( L(G + n))T

T nT
I LG L
I LG
= trE
n nnT

25 / 37

Wavefront reconstruction from local gradients

Linear estimation
Assume is also zero-mean random variable with Gaussian distribution
p() =

T 1
1
e R /2
1
/
2
|2 R |

and independent of n.
Suppose is given by a linear estimator

= Lt
where L minimizes
J (L)

T
= E tr( )(
)

= E tr( L(G + n))( L(G + n))T

T nT
I LG L
I LG
= trE
n nnT

25 / 37

Wavefront reconstruction from local gradients

Linear estimation
J (L)

T nT
I

LG

L
I LG L
= trE
n nnT

R 0

I 0 L G I
I 0 L G
= tr
0

Rn

By theorem on slide 11: dJ (L)/dL = 0 if and only if

L(GR GT + Rn )

= R GT

which is equivalent to

(R1 + GT Rn1 G)L

= GT Rn1

(also follows from maximizing likelihood of p(, t ) given t)

26 / 37

Wavefront reconstruction from local gradients

Linear estimation
J (L)

T nT
I

LG

L
I LG L
= trE
n nnT

R 0

I 0 L G I
I 0 L G
= tr
0

Rn

By theorem on slide 11: dJ (L)/dL = 0 if and only if

L(GR GT + Rn )

= R GT

which is equivalent to

(R1 + GT Rn1 G)L

= GT Rn1

(also follows from maximizing likelihood of p(, t ) given t)

26 / 37

Wavefront reconstruction from local gradients

Linear estimation
J (L)

T nT
I

LG

L
I LG L
= trE
n nnT

R 0

I 0 L G I
I 0 L G
= tr
0

Rn

By theorem on slide 11: dJ (L)/dL = 0 if and only if

L(GR GT + Rn )

= R GT

which is equivalent to

(R1 + GT Rn1 G)L

= GT Rn1

(also follows from maximizing likelihood of p(, t ) given t)

26 / 37

Wavefront reconstruction from local gradients

Linear estimation
J (L)

T nT
I

LG

L
I LG L
= trE
n nnT

R 0

I 0 L G I
I 0 L G
= tr
0

Rn

By theorem on slide 11: dJ (L)/dL = 0 if and only if

L(GR GT + Rn )

= R GT

which is equivalent to

(R1 + GT Rn1 G)L

= GT Rn1

(also follows from maximizing likelihood of p(, t ) given t)

26 / 37

Wavefront reconstruction from local gradients

Linear estimation
J (L)

T nT
I

LG

L
I LG L
= trE
n nnT

R 0

I 0 L G I
I 0 L G
= tr
0

Rn

By theorem on slide 11: dJ (L)/dL = 0 if and only if

L(GR GT + Rn )

= R GT

which is equivalent to

(R1 + GT Rn1 G)L

= GT Rn1

(also follows from maximizing likelihood of p(, t ) given t)

26 / 37

Wavefront reconstruction from local gradients

Linear estimation
Hence can be solved from

(R1 + GT Rn1 G) = GT Rn1 t

For Kolmogorov turbulence R1 appears to be sparse and fast iterative solvers
can be used, e.g., (Multigrid/Preconditioned) Conjugate Gradient.a .
a Ellerbroek,

JOSA, 19(9), (2002)

27 / 37

Wavefront reconstruction from local gradients

Overview wavefront reconstruction from local gradients
1

Line integration: depending on line integration method

(fast-inaccurate, slow-accurate);

Fourier domain inversion: errors because boundary effects of fourier

transform (noise can be weighted);

Least squares:

= G+ t = (GT G)+ GT t
4

Linear estimation:

= (GT Rn1 G)+ GT Rn1 t

or

= (R1 + GT Rn1 G)+ GT Rn1 t

28 / 37

Wavefront reconstruction from local gradients

Overview wavefront reconstruction from local gradients
1

Line integration: depending on line integration method

(fast-inaccurate, slow-accurate);

Fourier domain inversion: errors because boundary effects of fourier

transform (noise can be weighted);

Least squares:

= G+ t = (GT G)+ GT t
4

Linear estimation:

= (GT Rn1 G)+ GT Rn1 t

or

= (R1 + GT Rn1 G)+ GT Rn1 t

28 / 37

Wavefront reconstruction from local gradients

Overview wavefront reconstruction from local gradients
1

Line integration: depending on line integration method

(fast-inaccurate, slow-accurate);

Fourier domain inversion: errors because boundary effects of fourier

transform (noise can be weighted);

Least squares:

= G+ t = (GT G)+ GT t
4

Linear estimation:

= (GT Rn1 G)+ GT Rn1 t

or

= (R1 + GT Rn1 G)+ GT Rn1 t

28 / 37

Wavefront reconstruction from local gradients

Overview wavefront reconstruction from local gradients
1

Line integration: depending on line integration method

(fast-inaccurate, slow-accurate);

Fourier domain inversion: errors because boundary effects of fourier

transform (noise can be weighted);

Least squares:

= G+ t = (GT G)+ GT t
4

Linear estimation:

= (GT Rn1 G)+ GT Rn1 t

or

= (R1 + GT Rn1 G)+ GT Rn1 t

28 / 37

Wavefront reconstruction from local gradients

Wavefront reconstruction over time
Measurements over time given by:
t (k )

= G(k ) + n(k )

Vector stacking over time-window k = 1, , N:

t1:N = (IN G)1:N + n1:N
with E[1:N T1:N ] and E[n1:N n1T:N ] may be given.
Usually not the way to go:
For large N extremely complex problems;
Usually (k + 1) needs to be estimated from t (1), , t (k ).

Time-recursive methods!
29 / 37

Wavefront reconstruction from local gradients

Wavefront reconstruction over time
Measurements over time given by:
t (k )

= G(k ) + n(k )

Vector stacking over time-window k = 1, , N:

t1:N = (IN G)1:N + n1:N
with E[1:N T1:N ] and E[n1:N n1T:N ] may be given.
Usually not the way to go:
For large N extremely complex problems;
Usually (k + 1) needs to be estimated from t (1), , t (k ).

Time-recursive methods!
29 / 37

Wavefront reconstruction from local gradients

Wavefront reconstruction over time
Measurements over time given by:
t (k )

= G(k ) + n(k )

Vector stacking over time-window k = 1, , N:

t1:N = (IN G)1:N + n1:N
with E[1:N T1:N ] and E[n1:N n1T:N ] may be given.
Usually not the way to go:
For large N extremely complex problems;
Usually (k + 1) needs to be estimated from t (1), , t (k ).

Time-recursive methods!
29 / 37

Wavefront reconstruction from local gradients

Wavefront reconstruction over time
Measurements over time given by:
t (k )

= G(k ) + n(k )

Vector stacking over time-window k = 1, , N:

t1:N = (IN G)1:N + n1:N
with E[1:N T1:N ] and E[n1:N n1T:N ] may be given.
Usually not the way to go:
For large N extremely complex problems;
Usually (k + 1) needs to be estimated from t (1), , t (k ).

Time-recursive methods!
29 / 37

Wavefront reconstruction from local gradients

Wavefront reconstruction over time
Measurements over time given by:
t (k )

= G(k ) + n(k )

Vector stacking over time-window k = 1, , N:

t1:N = (IN G)1:N + n1:N
with E[1:N T1:N ] and E[n1:N n1T:N ] may be given.
Usually not the way to go:
For large N extremely complex problems;
Usually (k + 1) needs to be estimated from t (1), , t (k ).

Time-recursive methods!
29 / 37

Wavefront prediction
Problem formulation:
Given measurements:
t (k ) = G(k ) + n(k ),

k = 1, 2,

Given correlation coefficients:

E

(k )
n(k )

(k )
n(k )

T !

R ()
0

0
n2 I ()

k + 1|k ) of (k + 1) given t (k ), t (k 1), , t (1)

Determine the estimate (
such that

k + 1|k )) = E ((k + 1) (
k + 1|k ))T ((k + 1) (
k + 1|k ))
Jk ((
is minimized.

30 / 37

Wavefront prediction
Problem formulation:
Given measurements:
t (k ) = G(k ) + n(k ),

k = 1, 2,

Given correlation coefficients:

E

(k )
n(k )

(k )
n(k )

T !

R ()
0

0
n2 I ()

k + 1|k ) of (k + 1) given t (k ), t (k 1), , t (1)

Determine the estimate (
such that

k + 1|k )) = E ((k + 1) (
k + 1|k ))T ((k + 1) (
k + 1|k ))
Jk ((
is minimized.

30 / 37

Wavefront prediction
Problem formulation:
Given measurements:
t (k ) = G(k ) + n(k ),

k = 1, 2,

Given correlation coefficients:

E

(k )
n(k )

(k )
n(k )

T !

R ()
0

0
n2 I ()

k + 1|k ) of (k + 1) given t (k ), t (k 1), , t (1)

Determine the estimate (
such that

k + 1|k )) = E ((k + 1) (
k + 1|k ))T ((k + 1) (
k + 1|k ))
Jk ((
is minimized.

30 / 37

Wavefront prediction
Problem formulation:
Given measurements:
t (k ) = G(k ) + n(k ),

k = 1, 2,

Given correlation coefficients:

E

(k )
n(k )

(k )
n(k )

T !

R ()
0

0
n2 I ()

k + 1|k ) of (k + 1) given t (k ), t (k 1), , t (1)

Determine the estimate (
such that

k + 1|k )) = E ((k + 1) (
k + 1|k ))T ((k + 1) (
k + 1|k ))
Jk ((
is minimized.

30 / 37

Wavefront prediction
Solutiona :
a Fraanje

et al., JOSA, 27(11), (2010)

k + 1|k ) = Ak t(k )
(
where
t(k )

t (1)T ,

t (k )T

A1

= R (1)GT (GR (0)GT + n2 I )1

Ak

= linear function of R (0), , R (k ), G, n2

Warning: memory grows with time index k !

31 / 37

Wavefront prediction
Solutiona :
a Fraanje

et al., JOSA, 27(11), (2010)

k + 1|k ) = Ak t(k )
(
where
t(k )

t (1)T ,

t (k )T

A1

= R (1)GT (GR (0)GT + n2 I )1

Ak

= linear function of R (0), , R (k ), G, n2

Warning: memory grows with time index k !

31 / 37

Wavefront prediction
Solutiona :
a Fraanje

et al., JOSA, 27(11), (2010)

k + 1|k ) = Ak t(k )
(
where
t(k )

t (1)T ,

t (k )T

A1

= R (1)GT (GR (0)GT + n2 I )1

Ak

= linear function of R (0), , R (k ), G, n2

Warning: memory grows with time index k !

31 / 37

Wavefront prediction
AR predictor
Suppose auto-regressive model of wavefront is given:

(k + 1) =

p
X

Ai (k i ) + e(k + 1)

i =0

A1

Ap

(k )
..
.

(k n)

= Ap p (k ) + e(k + 1)

+ e(k + 1)

where e(k ) is a zero-mean white-noise process.

Multiplying with from the right with Tp (k ) and taking expected value yields:
E (k + 1)p (k )T

= Ap E p (k )p (k )T
= Ap p

1 T
and Re = E e(k )e(k )T satisfies Re = R (0) p
p p

32 / 37

Wavefront prediction
AR predictor
Suppose auto-regressive model of wavefront is given:

(k + 1) =

p
X

Ai (k i ) + e(k + 1)

i =0

A1

Ap

(k )
..
.

(k n)

= Ap p (k ) + e(k + 1)

+ e(k + 1)

where e(k ) is a zero-mean white-noise process.

Multiplying with from the right with Tp (k ) and taking expected value yields:
E (k + 1)p (k )T

= Ap E p (k )p (k )T
= Ap p

1 T
and Re = E e(k )e(k )T satisfies Re = R (0) p
p p

32 / 37

Wavefront prediction
AR predictor
Suppose auto-regressive model of wavefront is given:

(k + 1) =

p
X

Ai (k i ) + e(k + 1)

i =0

A1

Ap

(k )
..
.

(k n)

= Ap p (k ) + e(k + 1)

+ e(k + 1)

where e(k ) is a zero-mean white-noise process.

Multiplying with from the right with Tp (k ) and taking expected value yields:
E (k + 1)p (k )T

= Ap E p (k )p (k )T
= Ap p

1 T
and Re = E e(k )e(k )T satisfies Re = R (0) p
p p

32 / 37

Wavefront prediction
AR predictor
Suppose auto-regressive model of wavefront is given:

(k + 1) =

p
X

Ai (k i ) + e(k + 1)

i =0

A1

Ap

(k )
..
.

(k n)

= Ap p (k ) + e(k + 1)

+ e(k + 1)

where e(k ) is a zero-mean white-noise process.

Multiplying with from the right with Tp (k ) and taking expected value yields:
E (k + 1)p (k )T

= Ap E p (k )p (k )T
= Ap p

1 T
and Re = E e(k )e(k )T satisfies Re = R (0) p
p p

32 / 37

Wavefront prediction
AR predictor
Suppose auto-regressive model of wavefront is given:

(k + 1) =

p
X

Ai (k i ) + e(k + 1)

i =0

A1

Ap

(k )
..
.

(k n)

= Ap p (k ) + e(k + 1)

+ e(k + 1)

where e(k ) is a zero-mean white-noise process.

Multiplying with from the right with Tp (k ) and taking expected value yields:
E (k + 1)p (k )T

= Ap E p (k )p (k )T
= Ap p

1 T
and Re = E e(k )e(k )T satisfies Re = R (0) p
p p

32 / 37

Wavefront prediction
AR predictor
Suppose auto-regressive model of wavefront is given:

(k + 1) =

p
X

Ai (k i ) + e(k + 1)

i =0

A1

Ap

(k )
..
.

(k n)

= Ap p (k ) + e(k + 1)

+ e(k + 1)

where e(k ) is a zero-mean white-noise process.

Multiplying with from the right with Tp (k ) and taking expected value yields:
E (k + 1)p (k )T

= Ap E p (k )p (k )T
= Ap p

1 T
and Re = E e(k )e(k )T satisfies Re = R (0) p
p p

32 / 37

Wavefront prediction
AR predictor
Suppose auto-regressive model of wavefront is given:

(k + 1) =

p
X

Ai (k i ) + e(k + 1)

i =0

A1

Ap

(k )
..
.

(k n)

= Ap p (k ) + e(k + 1)

+ e(k + 1)

where e(k ) is a zero-mean white-noise process.

Multiplying with from the right with Tp (k ) and taking expected value yields:
E (k + 1)p (k )T

= Ap E p (k )p (k )T
= Ap p

1 T
and Re = E e(k )e(k )T satisfies Re = R (0) p
p p

32 / 37

Wavefront prediction
AR predictor
k + 1|k ) is estimated by
Assume (
k + 1|k ) =
(

p
X

Li t (k i )

i =0

L0

t (k )

..
Lp

.
t (k p)

G(k ) + n(k )

..
Lp

.
G(k p) + n(k p)

= Lp (Ip G)p (k ) + Lp np (k )

33 / 37

Wavefront prediction
AR predictor
k + 1|k ) is estimated by
Assume (
k + 1|k ) =
(

p
X

Li t (k i )

i =0

L0

t (k )

..
Lp

.
t (k p)

G(k ) + n(k )

..
Lp

.
G(k p) + n(k p)

= Lp (Ip G)p (k ) + Lp np (k )

33 / 37

Wavefront prediction
AR predictor
k + 1|k ) is estimated by
Assume (
k + 1|k ) =
(

p
X

Li t (k i )

i =0

L0

t (k )

..
Lp

.
t (k p)

G(k ) + n(k )

..
Lp

.
G(k p) + n(k p)

= Lp (Ip G)p (k ) + Lp np (k )

33 / 37

Wavefront prediction
AR predictor
k + 1|k ) is estimated by
Assume (
k + 1|k ) =
(

p
X

Li t (k i )

i =0

L0

t (k )

..
Lp

.
t (k p)

G(k ) + n(k )

..
Lp

.
G(k p) + n(k p)

= Lp (Ip G)p (k ) + Lp np (k )

33 / 37

Wavefront prediction
AR predictor
(k + 1) = Ap p (k ) + e(k + 1)

(k + 1|k ) = Lp (Ip G)p (k ) + Lp np (k )

such that

k + 1|k ) = (Ap Lp (Ip G))p (k ) + e(k + 1) Lp np (k )

(k + 1) (
Objective is to minimize:
Jk (Lp )

k + 1|k ))T ((k + 1) (

k + 1|k ))
= E ((k + 1) (

= tr (Ap Lp (Ip G))p (Ap Lp (Ip G))T + Re + n2 Lp Lp T

Hence by theorem on slide 11 Lp is a solution if and only if:

Lp (Ip G)p (Ip GT ) + n2 I

= Ap p (Ip GT )

34 / 37

Wavefront prediction
AR predictor
(k + 1) = Ap p (k ) + e(k + 1)

(k + 1|k ) = Lp (Ip G)p (k ) + Lp np (k )

such that

k + 1|k ) = (Ap Lp (Ip G))p (k ) + e(k + 1) Lp np (k )

(k + 1) (
Objective is to minimize:
Jk (Lp )

k + 1|k ))T ((k + 1) (

k + 1|k ))
= E ((k + 1) (

= tr (Ap Lp (Ip G))p (Ap Lp (Ip G))T + Re + n2 Lp Lp T

Hence by theorem on slide 11 Lp is a solution if and only if:

Lp (Ip G)p (Ip GT ) + n2 I

= Ap p (Ip GT )

34 / 37

Wavefront prediction
AR predictor
(k + 1) = Ap p (k ) + e(k + 1)

(k + 1|k ) = Lp (Ip G)p (k ) + Lp np (k )

such that

k + 1|k ) = (Ap Lp (Ip G))p (k ) + e(k + 1) Lp np (k )

(k + 1) (
Objective is to minimize:
Jk (Lp )

k + 1|k ))T ((k + 1) (

k + 1|k ))
= E ((k + 1) (

= tr (Ap Lp (Ip G))p (Ap Lp (Ip G))T + Re + n2 Lp Lp T

Hence by theorem on slide 11 Lp is a solution if and only if:

Lp (Ip G)p (Ip GT ) + n2 I

= Ap p (Ip GT )

34 / 37

Wavefront prediction
AR predictor
(k + 1) = Ap p (k ) + e(k + 1)

(k + 1|k ) = Lp (Ip G)p (k ) + Lp np (k )

such that

k + 1|k ) = (Ap Lp (Ip G))p (k ) + e(k + 1) Lp np (k )

(k + 1) (
Objective is to minimize:
Jk (Lp )

k + 1|k ))T ((k + 1) (

k + 1|k ))
= E ((k + 1) (

= tr (Ap Lp (Ip G))p (Ap Lp (Ip G))T + Re + n2 Lp Lp T

Hence by theorem on slide 11 Lp is a solution if and only if:

Lp (Ip G)p (Ip GT ) + n2 I

= Ap p (Ip GT )

34 / 37

Wavefront prediction
State-space predictors
State-space model for wavefront phase:

(k + 1) = A(k ) + Ke(k )
(k ) = C (k ) + e(k )
measurement equation:
t (k )

= G(k ) + n(k )

Solutiona given by Kalman filter:

k |k 1)
t (k |k 1) = GC (
k + 1|k ) = A(
k |k 1) + Kt (t (k ) t (k ))
(
k + 1|k ) = C (
k + 1|k )
(
where Kt the Kalman gain.
a c.f.

Fraanje et al., JOSA, 27(11), (2010)

35 / 37

Wavefront prediction
State-space predictors
State-space model for wavefront phase:

(k + 1) = A(k ) + Ke(k )
(k ) = C (k ) + e(k )
measurement equation:
t (k )

= G(k ) + n(k )

Solutiona given by Kalman filter:

k |k 1)
t (k |k 1) = GC (
k + 1|k ) = A(
k |k 1) + Kt (t (k ) t (k ))
(
k + 1|k ) = C (
k + 1|k )
(
where Kt the Kalman gain.
a c.f.

Fraanje et al., JOSA, 27(11), (2010)

35 / 37

Wavefront prediction
State-space predictors
State-space model for wavefront phase:

(k + 1) = A(k ) + Ke(k )
(k ) = C (k ) + e(k )
measurement equation:
t (k )

= G(k ) + n(k )

Solutiona given by Kalman filter:

k |k 1)
t (k |k 1) = GC (
k + 1|k ) = A(
k |k 1) + Kt (t (k ) t (k ))
(
k + 1|k ) = C (
k + 1|k )
(
where Kt the Kalman gain.
a c.f.

Fraanje et al., JOSA, 27(11), (2010)

35 / 37

Wavefront prediction
State-space predictors
State-space model for wavefront phase:

(k + 1) = A(k ) + Ke(k )
(k ) = C (k ) + e(k )
measurement equation:
t (k )

= G(k ) + n(k )

Solutiona given by Kalman filter:

k |k 1)
t (k |k 1) = GC (
k + 1|k ) = A(
k |k 1) + Kt (t (k ) t (k ))
(
k + 1|k )
k + 1|k ) = C (
(
where Kt the Kalman gain.
a c.f.

Fraanje et al., JOSA, 27(11), (2010)

35 / 37

Research issues
Phase reconstruction from pixels for extended sources;
Phase reconstruction from single pixel for point sources;
Efficient distributed wavefront reconstruction / prediction (implementation on

GPU / n-core CPU);

Fast wavefront reconstuction for nonstationary turbulence;
Phase reconstruction over multiple layers (tomography);
Data-driven methods (one-button operation).

36 / 37

Research issues
Phase reconstruction from pixels for extended sources;
Phase reconstruction from single pixel for point sources;
Efficient distributed wavefront reconstruction / prediction (implementation on

GPU / n-core CPU);

Fast wavefront reconstuction for nonstationary turbulence;
Phase reconstruction over multiple layers (tomography);
Data-driven methods (one-button operation).

36 / 37

Research issues
Phase reconstruction from pixels for extended sources;
Phase reconstruction from single pixel for point sources;
Efficient distributed wavefront reconstruction / prediction (implementation on

GPU / n-core CPU);

Fast wavefront reconstuction for nonstationary turbulence;
Phase reconstruction over multiple layers (tomography);
Data-driven methods (one-button operation).

36 / 37

Research issues
Phase reconstruction from pixels for extended sources;
Phase reconstruction from single pixel for point sources;
Efficient distributed wavefront reconstruction / prediction (implementation on

GPU / n-core CPU);

Fast wavefront reconstuction for nonstationary turbulence;
Phase reconstruction over multiple layers (tomography);
Data-driven methods (one-button operation).

36 / 37

Research issues
Phase reconstruction from pixels for extended sources;
Phase reconstruction from single pixel for point sources;
Efficient distributed wavefront reconstruction / prediction (implementation on

GPU / n-core CPU);

Fast wavefront reconstuction for nonstationary turbulence;
Phase reconstruction over multiple layers (tomography);
Data-driven methods (one-button operation).

36 / 37

Research issues
Phase reconstruction from pixels for extended sources;
Phase reconstruction from single pixel for point sources;
Efficient distributed wavefront reconstruction / prediction (implementation on

GPU / n-core CPU);

Fast wavefront reconstuction for nonstationary turbulence;
Phase reconstruction over multiple layers (tomography);
Data-driven methods (one-button operation).

36 / 37

Overview
Wavefront reconstruction from pixel intensities:
Gerchberg-Saxton;
Center-of-mass algorithm.

Wavefront reconstruction from local gradients:

Line integration;
Fourier domain inversion;
Least squares;
Linear estimation (exploit statistical information).

Wavefront prediction:
Time-recursive methods needed;
AR-predictors;
Kalman filter (state-space models).
37 / 37

Overview
Wavefront reconstruction from pixel intensities:
Gerchberg-Saxton;
Center-of-mass algorithm.

Wavefront reconstruction from local gradients:

Line integration;
Fourier domain inversion;
Least squares;
Linear estimation (exploit statistical information).

Wavefront prediction:
Time-recursive methods needed;
AR-predictors;
Kalman filter (state-space models).
37 / 37

Overview
Wavefront reconstruction from pixel intensities:
Gerchberg-Saxton;
Center-of-mass algorithm.

Wavefront reconstruction from local gradients:

Line integration;
Fourier domain inversion;
Least squares;
Linear estimation (exploit statistical information).

Wavefront prediction:
Time-recursive methods needed;
AR-predictors;
Kalman filter (state-space models).
37 / 37