statistical analysis of multi source calibration - astron.nl · the scientific goals of new radio...
TRANSCRIPT
Statistical analysis of multiStatistical analysis of multi‐‐source source CalibrationCalibrationCalibrationCalibration
SS. . KazemiKazemi
KapteynKapteyn Astronomical Institute Astronomical Institute [email protected]@astro.rug.nl
Presentation in CALIM, August Presentation in CALIM, August 20102010
Here comes your footer Page 1Page 1
(Under supervision of S. (Under supervision of S. ZaroubiZaroubi, and , and S.YatawattaS.Yatawatta))
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
CalibrationNoise in solutions
ww
w.lo
faIllustrative exampleConclusions and future work
IntroductionIntroductionThe scientific goals of new radio arrays (e.g. LOFAR & SKA) require extreme sensitivity and dynamic range (e.g. LOFAR EOR)
Calibration
E i i k i d h k d i h b f i i
Calibration challenges:• Accuracy of calibration algorithms becomes crucial for achievement of scientific goals
Estimating unknown instrument and the sky parameters and correcting them before imaging.
• Accuracy of calibration algorithms becomes crucial for achievement of scientific goals•The large number of data requires fast algorithms
Input noise output noise“I t i ”
p“Instrumental noise”
“Sky noise”“RFI”
“Input noise”+
Noise of the calibration algorithm“solver noise”
Calibration process
Here comes your footerPresentation in CALIM August 2010 Page 2
The aim is to devise new calibration methods which minimize the solver noiseThe aim is to devise new calibration methods which minimize the solver noise
ar.o
rg/
ww
w.lo
fa
2 Measurement equation
3 The LS, EM, and SAGE algorithms
4 Solver noise
6 Conclusions and future work
5 Illustrative example
6 Conclusions and future work
Here comes your footer Page 3Presentation in CALIM August 2010
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
Initial assumptionsCalibration problem modeling
ww
w.lo
faIllustrative exampleConclusions and future work
Measurement equationMeasurement equationInitial assumptions Every antenna
has dual l i d f d
X
YRadio frequency sky is decomposed to K separated
sources far away from the array
polarized feeds Y
Baseline pqCalibration problem modeling Hamaker‐Bregman‐Sault, 1996.
Coherency: describes polarization state of the source i
Jones matrix: describes amplifier gains, beam shapes, ionospheric effects, and etc
V i d f
Noise matrix of
Vectorized form
Here comes your footer Page 4
Noise matrix of the baseline pq
Measured visibility
Presentation in CALIM August 2010
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
The Least Squares (LS, Normal) calibration methodExpectation Maximization (EM) algorithmSpace Alternating Generalized Expectation Maximization algorithm (SAGE)
ww
w.lo
faIllustrative exampleConclusions and future work
The LS, EM, and SAGE algorithmsThe LS, EM, and SAGE algorithmsThe Least Squares (LS, Normal) calibration method ( AIPS, AIPS++, CASA)
Levenberg Marquardt
•Easy to program But,
g q
technique
•Heavy computational cost of O((KN)2 ) •Slow rate of convergence
Expectation Maximization (EM) algorithm
Y Ob d d t
Feder and Weinstein, 1988.
Y : Observed dataX : Complete data: Parameter value
Jensen’s inequality: this expected value will be decreased for all
Here comes your footer Page 5
will be decreased for all
Presentation in CALIM August 2010
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
The Least Squares (LS, Normal) calibration methodExpectation Maximization (EM) algorithmSpace Alternating Generalized Expectation Maximization algorithm (SAGE)
ww
w.lo
faIllustrative exampleConclusions and future work
Expectation‐Step: Compute the expected value of the log‐likelihood function, with respect to conditional distribution of the complete data X given observed data y under the current
Maximization‐Step: Maximize the log‐likelihood with respect to
conditional distribution of the complete data X given observed data y under the current parameter estimate
Assuming that the contribution of source i depends only on a subset of parameters and partitioning the parameter vector as
Yatawatta et al. 2009, Kazemi et al. in prep.
E‐ Step:
M‐ Step: min
Fessler and Herro, 1994.
M Step: min
Space Alternating Generalized Expectation Maximization (SAGE) algorithm
• Dedicating whole the noise to only one sourcel h l h
Fessler and Herro, 1994.
Here comes your footer Page 6
• Utilizing the EM algorithm
Presentation in CALIM August 2010
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
The Least Squares (LS, Normal) calibration methodExpectation Maximization (EM) algorithmSpace Alternating Generalized Expectation Maximization algorithm (SAGE)
ww
w.lo
faIllustrative exampleConclusions and future work
Yatawatta et al. 2009, Kazemi et al. in prep.
E StE‐ Step:
M‐ Step: min
Advantages of the EM algorithm over the Normal algorithm:•Breaking the likelihood maximization into smaller computational steps•Breaking the likelihood maximization into smaller computational steps•Computational cost equal to KO(N2 )
•Increasing the likelihood at each iteration step•Increasing the likelihood at each iteration step•Improving the speed of convergence compared with LS calibrationBut still,•Possibility of converging to a local optimum•Implementation of the algorithm is complicated compared with LS method•Implementation of the algorithm is complicated compared with LS method
Additional advantages of using the SAGE algorithm:•Improving the speed of convergence
Here comes your footer Page 7
Improving the speed of convergence•Improving the accuracy of calibration results
Presentation in CALIM August 2010
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
Initial assumptionsKullback ‐Leibler divergence (KLD)Likelihood Ratio Test (LRT)
ww
w.lo
faIllustrative exampleConclusions and future work
Solver noiseSolver noiseInitial assumptions
The goal is to minimize the “distance” between the real gainsThe goal is to minimize the distance between the real gains and calculated solutions = Minimizing the solver noise
x := gain solutions for the source s at the antenna q
• Sky noise for two different directions is uncorrelated• Thermal noise for any number of directions from the same station is the same
xq,s:= gain solutions for the source s at the antenna q
• Solver noise is the same for all stations and directions
Kullback‐Leibler Divergence (KLD)
The higher KLD between two different sources solutions at the same antenna, the solver noise is lessW fit G i i t d l t th l ti b EM l ith
Here comes your footer Page 8
•We fit a Gaussian mixture model to the solutions by EM algorithm•KLD is calculated via Monte‐Carlo algorithm
Presentation in CALIM August 2010
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
Initial assumptionsKullback ‐Leibler Divergence (KLD)Likelihood Ratio Test (LRT)
ww
w.lo
faIllustrative exampleConclusions and future work
Likelihood Ratio Test (LRT)Likelihood Ratio Test (LRT)
Null model: Independent solutions for different sources for the same antennaAlternative model: Dependent solutions for different sources for the same antenna
Probability of accepting null modelrather than alternative model
The less LRT between two different sources solutions at the same antenna, the solver noise is less
Here comes your footer Page 9Presentation in CALIM August 2010
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
Comparison between the SAGE and the Normal calibration algorithms
ww
w.lo
faIllustrative exampleConclusions and future work
Illustrative exampleIllustrative exampleOne channel simulated observation of three sources by fourteen antennas
Clean image Image having gain errors and
Here comes your footer Page 10Presentation in CALIM August 2010
Clean image Image having gain errors and white Gaussian noise
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
Comparison between the SAGE and the Normal calibration algorithms
ww
w.lo
faIllustrative exampleConclusions and future work
Residual ErrorsSAGE, 9 iterations NORMAL, 9 iterations
=1.2186
=1.2179
STD
STD
Here comes your footer Page 11Presentation in CALIM August 2010
RMSE= 3.9259 RMSE= 3.9505
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
Comparison between the SAGE and the Normal calibration algorithms
ww
w.lo
faIllustrative exampleConclusions and future work
LOFAR CS1 images from CasA and CygA
Here comes your footer Page 12Presentation in CALIM August 2010
Yatawatta et al. 2009.
ar.o
rg/
C l i d f t kC l i d f t k
ww
w.lo
faConclusions and future workConclusions and future work
Conclusions:•The SAGE algorithm is superior in terms of accuracy and speed of convergence•The non‐linear problem requires suitable regularization •The KLD and the LRT could be used to study the influence of solver noise
Future work:A l th l ith t l d t f LOFAR•Apply the algorithms to real data of LOFAR
•Study different regularization methods•Implement different noise models
Here comes your footer Page 13Presentation in CALIM August 2010
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
LOFARCalibrationSolver noise
Th kTh k ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
LOFARLOFAR Thank youThank yourai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Here comes your footerPresentation at YERAC July 2010 Page 14
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
SimulationDirection dependent gainsDetecting solver noise via KLD and LRT
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Illustrative exampleIllustrative examplerai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Detection solver noise via KLD and LRTSolutions of antenna seven are given sin(t) noise where t is integration time
Normal calibrationThe SAGE calibrationg ( ) g
Real Solutions Solution added sin(t) noise
Here comes your footerPresentation at YERAC July 2010 Page 15
The KLD of the noisy antenna becomes zero
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
SimulationDirection dependent gainsDetecting solver noise via KLD and LRT
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Illustrative exampleIllustrative examplerai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Detection solver noise via KLD and LRT Normal calibrationThe SAGE calibration
Real Solutions Solution added sin(t) noise
Here comes your footerPresentation at YERAC July 2010 Page 16
The LRT of the noisy antenna becomes higher The LRT of the noisy antenna becomes higher
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
SimulationDirection dependent gainsDetecting solver noise via KLD and LRT
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Illustrative exampleIllustrative examplerai.ncsa.uiuc.edu/RAI_Projects_SKA.html
KLD comparison LRT comparisonThe Normal and the SAGE algorithms with 16 number of iterations
Here comes your footerPresentation at YERAC July 2010 Page 17
The SAGE algorithm solutions have superior accuracy
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
SimulationDirection dependent gainsDetecting solver noise via KLD and LRT
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Illustrative exampleIllustrative examplerai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Real Solutions Solution added sin(t) noise
Here comes your footerPresentation at YERAC July 2010 Page 18
The LRT of noisy antenna is highly increased
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
SimulationDirection dependent gainsDetecting solver noise via KLD and LRT
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Illustrative exampleIllustrative examplerai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Real Solutions Solution added sin(t) noise
Here comes your footerPresentation at YERAC July 2010 Page 19
The KLD of noisy antenna becomes almost zero
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
Initial assumptionsKullback ‐Leibler divergence (KLD)Likelihood Ratio Test (LRT)
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Noise in solutionsNoise in solutionsrai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Initial assumptions
• Sky noise for two different directions is uncorrelated• Thermal noise for any number of directions from the same station is the same• Solver noise is the same for all stations and directions
The goal is to minimize the “distance” between the real gains d l l d l i Mi i i i h l i
Kullback‐Leibler Divergence (KLD)
and calculated solutions = Minimizing the solver noise
The higher KLD between two different sources solutions at the same antenna, the solver noise is lessW fit G i i t d l t th l ti b EM l ith
Here comes your footerPresentation at YERAC July 2010 Page 20
•We fit a Gaussian mixture model to the solutions by EM algorithm•KLD is calculated via Monte‐Carlo algorithm
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
CalibrationSolver noise
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
IntroductionIntroduction
Calibration
rai.ncsa.uiuc.edu/RAI_Projects_SKA.html
•Estimating unknown instrument and the sky parameters and correcting them before imaging.•Finding the Maximum Likelihood estimate of the sky and the Instrument Unknown parameters.
Calibration challenges:•The scientific goals of LOFAR require extreme sensitivity and dynamic range (LOFAR EOR)•Intrinsically polarized feeds (dipoles) ‐ polarized measurement equation Wid fi ld f i•Wide fields of view
•Pronounced direction‐dependent effects
Input noise “instrumental noise”
output noiseNoise in calibration solutions
Solver noise
Here comes your footerPresentation at YERAC July 2010 Page 21
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
SimulationDirection dependent gainsDetecting solver noise via KLD and LRT
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Illustrative exampleIllustrative examplerai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Detection solver noise via KLD and LRT
Solutions of antenna seven are added sin(t) noise where t is integration time:
Here comes your footerPresentation at YERAC July 2010 Page 22
ar.o
rg/
IntroductionMeasurement equation
The LM, EM, and SAGE algorithmsNoise in solutions
Illustrative example
SimulationDirection dependent gainsDetecting solver noise via KLD and LRT
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Illustrative exampleIllustrative examplerai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Direction dependent gains
Jy Jy
Here comes your footerPresentation at YERAC July 2010 Page 23
Random solutions Sin(t) solutions
Xq,s
Here comes your footer Page 24
Statistical analysis of multiStatistical analysis of multi‐‐sourcesourceStatistical analysis of multiStatistical analysis of multi source source CalibrationCalibration
axie
n/zu
kunf
t.htm
bonn
.mpg
.de/
.../g
ala
S. S. KazemiKazemi
ww
w.m
pifr-
b
KapteynKapteyn Astronomical Institute Astronomical Institute [email protected]@astro.rug.nl
Presentation in CALIM, August Presentation in CALIM, August 20102010(Under supervision of S(Under supervision of S ZaroubiZaroubi andand S YatawattaS Yatawatta))(Under supervision of S. (Under supervision of S. ZaroubiZaroubi, and , and S.YatawattaS.Yatawatta))
ar.o
rg/
IntroductionMeasurement equation
The LS, EM, and SAGE algorithmsSolver noise
Illustrative example
Comparison between the SAGE and the Normal calibration algorithms
ww
w.lo
faIllustrative exampleConclusions and future work
i i d /RAI P j t SKA ht l
Illustrative exampleIllustrative examplerai.ncsa.uiuc.edu/RAI_Projects_SKA.html
Here comes your footer Page 26Presentation in CALIM August 2010