O UTLINE I NTRODUCTION M ETHOD R ESULT
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION Irina Udaltsova, UC Davis
Samuel Schmidt, Jiming Jiang, UC Davis October 28, 2011
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
O UTLINE
I NTRODUCTION
M ETHOD
R ESULT
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
R EDSHIFT ▶
Redshift: Due to expansion of the Universe, light spectrum from distant galaxies shifts towards longer (redder) wavelengths proportional to each galaxy’s distance
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
P HOTOMETRY & C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
C OLORS
.
.
.
.
.
.
O UTLINE I NTRODUCTION M ETHOD R ESULT
F EW METHODS OF PHOTOMETRIC REDSHIFT ESTIMATION
Author Connolly et.al. (1995) Wang et.al. (2007) Ball et.al. (2006) Collister et al. (2004) Wadadekar (2004) Mobasher et.al. (1999) Benitez (2000), Bender (2001) Feldmann et al. (2006)
Method Empirical training set / Regression Empirical training set / Kernel Regression K-nearest-neighbor (KNN) Neural networks Support vector machines SED template fitting / MLE SED template fitting / Bayesian SED template fitting / MLE/ Bayesian
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
C LASSIC K ERNEL REGRESSION
Observations (Xi , Yi ), i = 1, . . . , n are iid. Estimate regression function m(x) = E(Y|X = x). Kernel regression estimator ∑n K[(Xi − x)/h]Yi ˆ m(x) = ∑i=1 n i=1 K[(Xi − x)/h] where K is kernel, for example, Gaussian kernel.
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
A DAPTIVE K ERNEL M ETHOD
F IGURE : (a) Kernel regression, (b) Adapted-Kernel regression ▶ ▶
Takeda et.al. (2007) Kernel regression for image processing and reconstruction Automated procedure selects importance of local pixels for estimation of neighbors. .
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
A DAPTIVE K ERNEL M ETHOD
KH (Xi − x) =
1
−1
{ } 1 ′ −2 exp − (Xi − x) H (Xi − x) 2
|H| (2π)d/2 Xi = d-dim data in local analysis window, i = 1, ..., N x = center location of interest 1
H = hC− 2 = bandwidth matrix h = bandwidth C = inverse ’covariance’ matrix built from SVD of local gradient matrix
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
S PECIFICATION OF G RADIENT ▶
Gradient matrix ( ) G = z′1 (xj ) z′2 (xj ) . . . z′d (xj ) N×d z′ = first order derivatives along d coordinate directions for xj points in local analysis window
▶
SVD:G = USV ′ , where s1 0 . . . 0 0 s . . . 0 2 . ( ) .. s d , V = v1 |v2 | . . . |vd S = 0 0 0 0 ... 0 .. .. .. .. . . . . .
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
S PECIFICATION OF STEERING MATRIX H 1
H = hC− 2
C = γUθ ΛUθ′
where si d Λ = diag{σi }, σi = ∑ = function of eigenvalues of G ( ) sj Uθ = v1 |v2 | . . . |vd = eigenvectors of G for rotation angle θ γ = ˆf 1/d , where ˆf = kernel density estimate
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
D ATA 2-D C OLOR S PACE
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
A DAPTIVE K ERNEL M ETHOD - S AMPLE R ESULT
.
.
.
.
.
.
O UTLINE I NTRODUCTION M ETHOD R ESULT
R ESULTS : SIMULATED REDSHIFT DATA d 2 3 4 5
MSEK 0.0082 (0.0058) 0.0026 (0.0003) 0.0016 (0.0015) 0.0019 (0.0014)
MSEAdaptK 0.0053 (0.0016) 0.0011 (0.0007) 0.0009 (0.0006) 0.0011 (0.0008)
MSEKNN 0.0351 (0.0084) 0.0016 (0.0005) 0.0021 (0.0008) 0.003 (0.0024)
30000 galaxy observations: ugrizy magnitudes based on six
CWWSB templates & spectroscopic redshifts (zs ∼ 0 - 4). Stopping criteria: |yiter − yiter−1 | is mimimum for iter = 1, ..., 10, usually convergence reached in 3-7 iterations.
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
R ESULTS : SDSS DATA
d 2 3 4
MSEK 0.0081 (0.0010) 0.0077 (0.0028) 0.0099 (0.0032)
MSEAdaptK 0.0042 (0.0010) 0.0022 (0.0010) 0.0032 (0.0012)
MSEKNN 0.0065 (0.0010) 0.0090 (0.0024) 0.0039 (0.0015)
16000 galaxy observations: ugriz magnitudes & spectroscopic
redshifts (zs ∼ 0 - 0.5).
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
O UTLINE I NTRODUCTION M ETHOD R ESULT
C OMMENTS
▶
Photometric redshift require excellent photometry and colors (S/N ∼ 30)
▶
Training set must be representative and spanning in order to avoid introducing bias
▶
Bandwidth choice is an issue - no unique optimal solution across various iterations
THANK YOU! .
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
References:
▶ Takeda et.al. (2007) Kernel regression for image processing and reconstruction. IEEE transactions on image processing 16, 349-366. ▶ Vieu (1991) Nonparametric regression: optimal local bandwidth choice. JRSSB 53, 453-464. ▶ Silverman (1986) Density estimation for statistics and data analysis. Chapman & Hall. ▶ Connolly et.al. (1995), Slicing through multicolor space: galaxy redshifts from broadband photometry. Astronomical Journal 110, 2655-2664. ▶ Wang et.al. (2007) Kernel regression for determining photometric redshifts from Sloan broad-band photometry. Monthly notices of Royal Astronomical society 382, 1601-1606. ▶ Ball, Brunner, Myers, et.al. (2006), Robust machine learning applied to astronomical datasets III. Astrophysical Journal 650, 497-506. ▶ Collister et al. (2004) ANNz: estimating photometric redshifts using artificial neural networks. Publications of the Astronomical Society of the Pacific 116, 345-351. ▶ Wadadekar (2004) Estimating photometric redshifts using support vector machines. Publications of the Astronomical Society of the Pacific 117, 79-85. ▶ Mobasher et al. (1999) A self-consistent method for estimating photometric redshifts. ASP Conference Series 191, 37. ▶ Benitez et al. (2000) Bayesian photometric redshift estimation. Astrophysical Journa 536, 571-583. ▶ Feldmann et al. (2006) The Zurich extragalactic bayesian redshift analyzer and its first application: COSMOS. Monthly Notices of the Royal Astronomical Society 372, 565-577.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
O UTLINE I NTRODUCTION M ETHOD R ESULT
I MAGE PROCESSING
▶
Kernel regression for image processing and reconstruction Takeda et.al. (2007)
▶
2-D set of pixel locations 1-D set of pixel grayscale values Goal: estimate grayscale pixel value based on local grayscale values, adapting to local features of image
▶ ▶
.
I RINA U DALTSOVA , UC D AVIS
.
.
.
.
.
P HOTOMETRIC R EDSHIFTS W ITH A DAPTIVE K ERNEL R EGRESSION
F IGURE : Examples of steering weights
.
.
.
.
.
.