The retrieval algorithms in remote sensing generally involve complex physical forward models that are nonlinear and computationally expensive to evaluate. Statistical emulation provides an alternative with cheap computation and can be used to calibrate model parameters and to improve computational efficiency of the retrieval algorithms. We introduce a framework of combining dimension reduction of input and output spaces and Gaussian process emulation technique. The functional principal component analysis (FPCA) is chosen to reduce to the output space of thousands of dimensions by orders of magnitude. In addition, instead of making restrictive assumptions regarding the correlation structure of the high-dimensional input space, we identity and exploit the most important directions of this space and thus construct a Gaussian process emulator with feasible computation. We will present preliminary results obtained from applying our method to OCO-2 data, and discuss how our framework can be generalized in distributed systems.
Применение машинного обучения для навигации и управления роботами
Similar to CLIM Program: Remote Sensing Workshop, Statistical Emulation with Dimension Reduction for Complex Physical Forward Models - Emily Kang, Feb 12, 2018
11.polynomial regression model of making cost prediction in mixed cost analysisAlexander Decker
Similar to CLIM Program: Remote Sensing Workshop, Statistical Emulation with Dimension Reduction for Complex Physical Forward Models - Emily Kang, Feb 12, 2018 (20)
A Critique of the Proposed National Education Policy Reform
CLIM Program: Remote Sensing Workshop, Statistical Emulation with Dimension Reduction for Complex Physical Forward Models - Emily Kang, Feb 12, 2018
1. Statistical Emulation with Dimension Reduction for
Complex Physical Forward Models
Emily L. Kang
Department of Mathematical Sciences
University of Cincinnati
SAMSI-JPL Workshop “Remote Sensing, Uncertainty Quantification, and the
Theory of Data Systems” @Caltech , February 2018
This research is joint with Jon Hobbs (JPL), Alex Konomi (Univ. of Cincinnati), Pulong Ma
(Univ. of Cincinnati), and Anirban Mondal (Case Western Reserve Univ.), and Joon Jin Song
(Baylor Univ.)
Emulator Group February 2018 1 / 25
3. Introduction
Introduction
Complex physical models often require large resources in time and
memory to produce realistic results.
To characterize the impact of the uncertainties in the conditions or
the parameterization of these models (including computer models,
simulators), a sufficient number of simulations is required. However,
this can become extremely costly or even prohibitive.
One prevailing way to overcome this hurdle is to construct statistical
surrogates, namely emulators, to approximate the complex physical
models in a probabilistic way.
Emulator Group February 2018 3 / 25
4. Introduction
Introduction
Complex physical models often require large resources in time and
memory to produce realistic results.
To characterize the impact of the uncertainties in the conditions or
the parameterization of these models (including computer models,
simulators), a sufficient number of simulations is required. However,
this can become extremely costly or even prohibitive.
One prevailing way to overcome this hurdle is to construct statistical
surrogates, namely emulators, to approximate the complex physical
models in a probabilistic way.
Emulator Group February 2018 3 / 25
5. Introduction
Introduction
Complex physical models often require large resources in time and
memory to produce realistic results.
To characterize the impact of the uncertainties in the conditions or
the parameterization of these models (including computer models,
simulators), a sufficient number of simulations is required. However,
this can become extremely costly or even prohibitive.
One prevailing way to overcome this hurdle is to construct statistical
surrogates, namely emulators, to approximate the complex physical
models in a probabilistic way.
Emulator Group February 2018 3 / 25
6. Introduction
Emulating the Forward Model in Remote Sensing
X Y
retrieval
algorithm
N(0, ⌃✏)
F(X, B) + ✏ R(Y, B, F)
ˆX X
ˆX
radiancestate vector retrieval
Marginal
distribution
Radiative transfer (RF) computation within the forward model
F(X, B) is usually the rate-limiting step in many remote sensing
applications.
Emulator Group February 2018 4 / 25
7. Introduction
Emulating the Forward Model in Remote Sensing
X Y
retrieval
algorithm
N(0, ⌃✏)
F(X, B) + ✏ R(Y, B, F)
ˆX X
ˆX
radiancestate vector retrieval
Marginal
distribution
Radiative transfer (RF) computation within the forward model
F(X, B) is usually the rate-limiting step in many remote sensing
applications.
Emulator Group February 2018 4 / 25
9. Introduction
Goal and Motivation
We would like to build a statistical emulator ˆF(X, B) to reproduce
radiances, which does not comprise much on accuracy while
enhancing computational efficiency.
We can employ this computationally more efficient emulator for any
subsequent purposes such as uncertainty propagation, sensitivity
analysis, and calibration.
The emulator can also be potentially used to speedup the retrieval
algorithm process.
Emulator Group February 2018 6 / 25
10. Introduction
Challenges
The outputs of F(X, B) are high-dimensional. In OCO-2, the outputs
are radiances at hundreds of wavelengths from three bands, the O2
band, the weak CO2 band, and the strong CO2 band.
Directly emulating the relationship between high-dimensional outputs
and inputs can be complicated, in terms of both computation and
modeling.
F(X, B) is a nonlinear complex function with high-dimensional inputs
X, called the state vector. In OCO-2, X is m-dimensional with
m = 62.
Theoretical results in the literature have shown that the goodness of
the approximation of a target function using a statistical emulator
such as a Gaussian process (GP) deteriorates as the dimension m
increases.
Emulator Group February 2018 7 / 25
11. Introduction
Challenges
The outputs of F(X, B) are high-dimensional. In OCO-2, the outputs
are radiances at hundreds of wavelengths from three bands, the O2
band, the weak CO2 band, and the strong CO2 band.
Directly emulating the relationship between high-dimensional outputs
and inputs can be complicated, in terms of both computation and
modeling.
F(X, B) is a nonlinear complex function with high-dimensional inputs
X, called the state vector. In OCO-2, X is m-dimensional with
m = 62.
Theoretical results in the literature have shown that the goodness of
the approximation of a target function using a statistical emulator
such as a Gaussian process (GP) deteriorates as the dimension m
increases.
Emulator Group February 2018 7 / 25
12. Introduction
Challenges
The outputs of F(X, B) are high-dimensional. In OCO-2, the outputs
are radiances at hundreds of wavelengths from three bands, the O2
band, the weak CO2 band, and the strong CO2 band.
Directly emulating the relationship between high-dimensional outputs
and inputs can be complicated, in terms of both computation and
modeling.
F(X, B) is a nonlinear complex function with high-dimensional inputs
X, called the state vector. In OCO-2, X is m-dimensional with
m = 62.
Theoretical results in the literature have shown that the goodness of
the approximation of a target function using a statistical emulator
such as a Gaussian process (GP) deteriorates as the dimension m
increases.
Emulator Group February 2018 7 / 25
13. Method
Emulation with Dimension Reduction
We propose a framework to perform dimension reduction for both inputs
and outputs and then build a GP emulator on the low-dimensional spaces.
Step 1: Given a set of n available pairs of state vectors and soundings
(X1, Y1), . . . , (X2, Yn), perform dimension reduction for X and Y:
Dimension reduction for Y: Performing functional principal component
analysis (FPCA) to obtain the functional principal component scores ξ
Dimension reduction for X: Using the active subspace (AS) approach
to project X to a low-dimensional vector S = PX
Step 2: Build a GP emulator using the low-dimensional pairs (Si , ξi ),
i = 1, . . . , n.
Emulator Group February 2018 8 / 25
14. Method
Emulation with Dimension Reduction
We propose a framework to perform dimension reduction for both inputs
and outputs and then build a GP emulator on the low-dimensional spaces.
Step 1: Given a set of n available pairs of state vectors and soundings
(X1, Y1), . . . , (X2, Yn), perform dimension reduction for X and Y:
Dimension reduction for Y: Performing functional principal component
analysis (FPCA) to obtain the functional principal component scores ξ
Dimension reduction for X: Using the active subspace (AS) approach
to project X to a low-dimensional vector S = PX
Step 2: Build a GP emulator using the low-dimensional pairs (Si , ξi ),
i = 1, . . . , n.
Emulator Group February 2018 8 / 25
15. Method
Emulation with Dimension Reduction
We propose a framework to perform dimension reduction for both inputs
and outputs and then build a GP emulator on the low-dimensional spaces.
Step 1: Given a set of n available pairs of state vectors and soundings
(X1, Y1), . . . , (X2, Yn), perform dimension reduction for X and Y:
Dimension reduction for Y: Performing functional principal component
analysis (FPCA) to obtain the functional principal component scores ξ
Dimension reduction for X: Using the active subspace (AS) approach
to project X to a low-dimensional vector S = PX
Step 2: Build a GP emulator using the low-dimensional pairs (Si , ξi ),
i = 1, . . . , n.
Emulator Group February 2018 8 / 25
16. Method
Emulation with Dimension Reduction
We propose a framework to perform dimension reduction for both inputs
and outputs and then build a GP emulator on the low-dimensional spaces.
Step 1: Given a set of n available pairs of state vectors and soundings
(X1, Y1), . . . , (X2, Yn), perform dimension reduction for X and Y:
Dimension reduction for Y: Performing functional principal component
analysis (FPCA) to obtain the functional principal component scores ξ
Dimension reduction for X: Using the active subspace (AS) approach
to project X to a low-dimensional vector S = PX
Step 2: Build a GP emulator using the low-dimensional pairs (Si , ξi ),
i = 1, . . . , n.
Emulator Group February 2018 8 / 25
17. Method
Emulation with Dimension Reduction
We propose a framework to perform dimension reduction for both inputs
and outputs and then build a GP emulator on the low-dimensional spaces.
Step 1: Given a set of n available pairs of state vectors and soundings
(X1, Y1), . . . , (X2, Yn), perform dimension reduction for X and Y:
Dimension reduction for Y: Performing functional principal component
analysis (FPCA) to obtain the functional principal component scores ξ
Dimension reduction for X: Using the active subspace (AS) approach
to project X to a low-dimensional vector S = PX
Step 2: Build a GP emulator using the low-dimensional pairs (Si , ξi ),
i = 1, . . . , n.
Emulator Group February 2018 8 / 25
18. Method
The OCO-2 Radiance Data
SCO2
O2 WCO2
0 250 500 750 1000
0 250 500 750 1000
0.0e+00
2.5e+19
5.0e+19
7.5e+19
1.0e+20
0.0e+00
2.5e+19
5.0e+19
7.5e+19
1.0e+20
Wavelength Index
Radiance
Sounding
1
2
3
4
5
6
7
8
9
10
The wavelengths vary from sounding to sounding.
Many data are missing.
Emulator Group February 2018 9 / 25
19. Method Dimension Reduction for Y
Define Yijk as the radiance at wavelength ωijk for the ith sounding and
jth spectral band, where i = 1, . . . , n, j = 1, 2, 3, and k = 1, . . . , Nij .
Here, Nij can vary across both soundings (indexed by i) and spectral
bands (indexed by j).
Since we have radiances at irregular wavelengths, we perform
functional principal component analysis (FPCA) on the radiances to
reduce dimension.
Emulator Group February 2018 10 / 25
20. Method Dimension Reduction for Y
Dimension Reduction for Radiances
We model the radiances as realizations of a random function,
Yijk = µj (ωijk) +
∞
l=1
ξijl φjl (ωijk);
µj (·) denotes the mean function for the jth spectral band, j = 1, 2, 3;
φjl (·) denotes the lth eigenfunction for the jth spectral band,
corresponding to nonincreasing eigenvalues λjl ;
{ξijl } are random variables with mean 0 and var(ξijl ) = λjl .
{ξijl } are called the functional principal component (FPC) scores.
Emulator Group February 2018 11 / 25
21. Method Dimension Reduction for Y
Dimension Reduction for Radiances
We model the radiances as realizations of a random function,
Yijk = µj (ωijk) +
∞
l=1
ξijl φjl (ωijk);
µj (·) denotes the mean function for the jth spectral band, j = 1, 2, 3;
φjl (·) denotes the lth eigenfunction for the jth spectral band,
corresponding to nonincreasing eigenvalues λjl ;
{ξijl } are random variables with mean 0 and var(ξijl ) = λjl .
{ξijl } are called the functional principal component (FPC) scores.
Emulator Group February 2018 11 / 25
22. Method Dimension Reduction for Y
Dimension Reduction for Radiances
We model the radiances as realizations of a random function,
Yijk = µj (ωijk) +
∞
l=1
ξijl φjl (ωijk);
µj (·) denotes the mean function for the jth spectral band, j = 1, 2, 3;
φjl (·) denotes the lth eigenfunction for the jth spectral band,
corresponding to nonincreasing eigenvalues λjl ;
{ξijl } are random variables with mean 0 and var(ξijl ) = λjl .
{ξijl } are called the functional principal component (FPC) scores.
Emulator Group February 2018 11 / 25
23. Method Dimension Reduction for Y
Dimension Reduction for Radiances
We model the radiances as realizations of a random function,
Yijk = µj (ωijk) +
∞
l=1
ξijl φjl (ωijk);
µj (·) denotes the mean function for the jth spectral band, j = 1, 2, 3;
φjl (·) denotes the lth eigenfunction for the jth spectral band,
corresponding to nonincreasing eigenvalues λjl ;
{ξijl } are random variables with mean 0 and var(ξijl ) = λjl .
{ξijl } are called the functional principal component (FPC) scores.
Emulator Group February 2018 11 / 25
24. Method Dimension Reduction for Y
Dimension Reduction for Radiances
We model the radiances as realizations of a random function,
Yijk = µj (ωijk) +
∞
l=1
ξijl φjl (ωijk);
µj (·) denotes the mean function for the jth spectral band, j = 1, 2, 3;
φjl (·) denotes the lth eigenfunction for the jth spectral band,
corresponding to nonincreasing eigenvalues λjl ;
{ξijl } are random variables with mean 0 and var(ξijl ) = λjl .
{ξijl } are called the functional principal component (FPC) scores.
Emulator Group February 2018 11 / 25
25. Method Dimension Reduction for Y
Computational Methods for FPCA
Various methods have been proposed for FPCA, including Ramsay
and Silverman (2005), Yao et al. (2005), and Li and Hsing (2010).
As in the previous work, we carry out FPCA for the jth spectral band:
Estimate the mean curve µj using a local linea smoother (Fan and
Gijbels, 2996);
Estimate the covariance function Gj (ω1, ω2) = l λjl φjl (ω1)φjl (ω2)
using a local linear smoother;
Obtain estimates of the eigenfunctions and eigenvalues, ˆφjl and ˆλjl ;
Estimate the FPC score ξijl = (Yij (ω) − µj (ω))φjl (ω)dω by numerical
integration:
ˆξijl =
Nij
k=1
(Yijk − ˆµj (ωijk ))ˆφjl (ωijk )(ωijk − ωij(k−1)).
Emulator Group February 2018 12 / 25
26. Method Dimension Reduction for Y
Computational Methods for FPCA
Various methods have been proposed for FPCA, including Ramsay
and Silverman (2005), Yao et al. (2005), and Li and Hsing (2010).
As in the previous work, we carry out FPCA for the jth spectral band:
Estimate the mean curve µj using a local linea smoother (Fan and
Gijbels, 2996);
Estimate the covariance function Gj (ω1, ω2) = l λjl φjl (ω1)φjl (ω2)
using a local linear smoother;
Obtain estimates of the eigenfunctions and eigenvalues, ˆφjl and ˆλjl ;
Estimate the FPC score ξijl = (Yij (ω) − µj (ω))φjl (ω)dω by numerical
integration:
ˆξijl =
Nij
k=1
(Yijk − ˆµj (ωijk ))ˆφjl (ωijk )(ωijk − ωij(k−1)).
Emulator Group February 2018 12 / 25
27. Method Dimension Reduction for Y
Computational Methods for FPCA
Various methods have been proposed for FPCA, including Ramsay
and Silverman (2005), Yao et al. (2005), and Li and Hsing (2010).
As in the previous work, we carry out FPCA for the jth spectral band:
Estimate the mean curve µj using a local linea smoother (Fan and
Gijbels, 2996);
Estimate the covariance function Gj (ω1, ω2) = l λjl φjl (ω1)φjl (ω2)
using a local linear smoother;
Obtain estimates of the eigenfunctions and eigenvalues, ˆφjl and ˆλjl ;
Estimate the FPC score ξijl = (Yij (ω) − µj (ω))φjl (ω)dω by numerical
integration:
ˆξijl =
Nij
k=1
(Yijk − ˆµj (ωijk ))ˆφjl (ωijk )(ωijk − ωij(k−1)).
Emulator Group February 2018 12 / 25
28. Method Dimension Reduction for Y
Computational Methods for FPCA
Various methods have been proposed for FPCA, including Ramsay
and Silverman (2005), Yao et al. (2005), and Li and Hsing (2010).
As in the previous work, we carry out FPCA for the jth spectral band:
Estimate the mean curve µj using a local linea smoother (Fan and
Gijbels, 2996);
Estimate the covariance function Gj (ω1, ω2) = l λjl φjl (ω1)φjl (ω2)
using a local linear smoother;
Obtain estimates of the eigenfunctions and eigenvalues, ˆφjl and ˆλjl ;
Estimate the FPC score ξijl = (Yij (ω) − µj (ω))φjl (ω)dω by numerical
integration:
ˆξijl =
Nij
k=1
(Yijk − ˆµj (ωijk ))ˆφjl (ωijk )(ωijk − ωij(k−1)).
Emulator Group February 2018 12 / 25
29. Method Dimension Reduction for Y
Computational Methods for FPCA
Various methods have been proposed for FPCA, including Ramsay
and Silverman (2005), Yao et al. (2005), and Li and Hsing (2010).
As in the previous work, we carry out FPCA for the jth spectral band:
Estimate the mean curve µj using a local linea smoother (Fan and
Gijbels, 2996);
Estimate the covariance function Gj (ω1, ω2) = l λjl φjl (ω1)φjl (ω2)
using a local linear smoother;
Obtain estimates of the eigenfunctions and eigenvalues, ˆφjl and ˆλjl ;
Estimate the FPC score ξijl = (Yij (ω) − µj (ω))φjl (ω)dω by numerical
integration:
ˆξijl =
Nij
k=1
(Yijk − ˆµj (ωijk ))ˆφjl (ωijk )(ωijk − ωij(k−1)).
Emulator Group February 2018 12 / 25
30. Method Dimension Reduction for Y
For the jth spectral band, by choosing Kj principal components, we
transform the original Nij -dimensional outputs Yij = (Yij1, . . . , YijNij
)
to a Kj -dimensional output ξij = (ξij1, . . . , ξijKj
) , i = 1, . . . , n.
The number of principal components, Kj is chosen using
leave-one-curve-out or K-fold cross validations, and can be potentially
different from the three bands.
Let ξi ≡ (ξi1, ξi2, ξi3) . It has K = 3
j=1 Kj elements, and is lower
dimensional vector compared to the original ith sounding
Yi ≡ (Yi1, Yi2, Yi3) , for i = 1, . . . , n.
Our preliminary results show that with Kj = 2 or 3, we can achieve
nearly lossless representationtion with at least 99% of variation is
preserved.
Emulator Group February 2018 13 / 25
31. Method Dimension Reduction for Y
For the jth spectral band, by choosing Kj principal components, we
transform the original Nij -dimensional outputs Yij = (Yij1, . . . , YijNij
)
to a Kj -dimensional output ξij = (ξij1, . . . , ξijKj
) , i = 1, . . . , n.
The number of principal components, Kj is chosen using
leave-one-curve-out or K-fold cross validations, and can be potentially
different from the three bands.
Let ξi ≡ (ξi1, ξi2, ξi3) . It has K = 3
j=1 Kj elements, and is lower
dimensional vector compared to the original ith sounding
Yi ≡ (Yi1, Yi2, Yi3) , for i = 1, . . . , n.
Our preliminary results show that with Kj = 2 or 3, we can achieve
nearly lossless representationtion with at least 99% of variation is
preserved.
Emulator Group February 2018 13 / 25
32. Method Dimension Reduction for Y
For the jth spectral band, by choosing Kj principal components, we
transform the original Nij -dimensional outputs Yij = (Yij1, . . . , YijNij
)
to a Kj -dimensional output ξij = (ξij1, . . . , ξijKj
) , i = 1, . . . , n.
The number of principal components, Kj is chosen using
leave-one-curve-out or K-fold cross validations, and can be potentially
different from the three bands.
Let ξi ≡ (ξi1, ξi2, ξi3) . It has K = 3
j=1 Kj elements, and is lower
dimensional vector compared to the original ith sounding
Yi ≡ (Yi1, Yi2, Yi3) , for i = 1, . . . , n.
Our preliminary results show that with Kj = 2 or 3, we can achieve
nearly lossless representationtion with at least 99% of variation is
preserved.
Emulator Group February 2018 13 / 25
33. Method Dimension Reduction for Y
For the jth spectral band, by choosing Kj principal components, we
transform the original Nij -dimensional outputs Yij = (Yij1, . . . , YijNij
)
to a Kj -dimensional output ξij = (ξij1, . . . , ξijKj
) , i = 1, . . . , n.
The number of principal components, Kj is chosen using
leave-one-curve-out or K-fold cross validations, and can be potentially
different from the three bands.
Let ξi ≡ (ξi1, ξi2, ξi3) . It has K = 3
j=1 Kj elements, and is lower
dimensional vector compared to the original ith sounding
Yi ≡ (Yi1, Yi2, Yi3) , for i = 1, . . . , n.
Our preliminary results show that with Kj = 2 or 3, we can achieve
nearly lossless representationtion with at least 99% of variation is
preserved.
Emulator Group February 2018 13 / 25
34. Method Dimension Reduction for X
Inputs X
The inputs of the forward function is an m-dimensional state vector
X. In OCO-2, m = 62.
An intuitive way to reduce the dimension is just to apply the principal
components analysis (PCA). However, PCA only considers the
correlation structure within X and ignores the role of outputs Y.
We call R(X) ∈ Rd where d < m a sufficient dimension reduction if
p(Y|X) = p(Y|R(X)),
where p(Y|X) and p(Y|R(X)) are conditional probability density
functions with respect to X and R(X), respectively.
Emulator Group February 2018 14 / 25
35. Method Dimension Reduction for X
Inputs X
The inputs of the forward function is an m-dimensional state vector
X. In OCO-2, m = 62.
An intuitive way to reduce the dimension is just to apply the principal
components analysis (PCA). However, PCA only considers the
correlation structure within X and ignores the role of outputs Y.
We call R(X) ∈ Rd where d < m a sufficient dimension reduction if
p(Y|X) = p(Y|R(X)),
where p(Y|X) and p(Y|R(X)) are conditional probability density
functions with respect to X and R(X), respectively.
Emulator Group February 2018 14 / 25
36. Method Dimension Reduction for X
Inputs X
The inputs of the forward function is an m-dimensional state vector
X. In OCO-2, m = 62.
An intuitive way to reduce the dimension is just to apply the principal
components analysis (PCA). However, PCA only considers the
correlation structure within X and ignores the role of outputs Y.
We call R(X) ∈ Rd where d < m a sufficient dimension reduction if
p(Y|X) = p(Y|R(X)),
where p(Y|X) and p(Y|R(X)) are conditional probability density
functions with respect to X and R(X), respectively.
Emulator Group February 2018 14 / 25
37. Method Dimension Reduction for X
Therefore, we want to detect subspace Rd ⊂ Rm that the
loglikelihood of Y, f , is most sensitive to.
1. Compute the gradient f ≡ Xf (X).
2. Compute eigenvalue decomposition of the m × m matrix
C ≡ Eπ(X)[( f )( f ) ]:
C = WΛWT
,
where W is the orthogonal matrix and Λ = diag{λ1, . . . , λm} is the
diagonal matrix of ordered eigenvalues with λ1 ≥ λ2 ≥ · · · ≥ λm ≥ 0.
3. Let W ≡ [W1, W2], and Λ ≡ blockdiag{Λ1, Λ2}, where Λ1 contains
the largest d eigenvalues and W1 contains the corresponding d
eigenvectors.
The column space of W1 is called active subspace.
S = W1X is called active variable.
U = W2X is called inactive variable.
X = W1S + W2U.
Emulator Group February 2018 15 / 25
38. Method Dimension Reduction for X
Therefore, we want to detect subspace Rd ⊂ Rm that the
loglikelihood of Y, f , is most sensitive to.
1. Compute the gradient f ≡ Xf (X).
2. Compute eigenvalue decomposition of the m × m matrix
C ≡ Eπ(X)[( f )( f ) ]:
C = WΛWT
,
where W is the orthogonal matrix and Λ = diag{λ1, . . . , λm} is the
diagonal matrix of ordered eigenvalues with λ1 ≥ λ2 ≥ · · · ≥ λm ≥ 0.
3. Let W ≡ [W1, W2], and Λ ≡ blockdiag{Λ1, Λ2}, where Λ1 contains
the largest d eigenvalues and W1 contains the corresponding d
eigenvectors.
The column space of W1 is called active subspace.
S = W1X is called active variable.
U = W2X is called inactive variable.
X = W1S + W2U.
Emulator Group February 2018 15 / 25
39. Method Dimension Reduction for X
Therefore, we want to detect subspace Rd ⊂ Rm that the
loglikelihood of Y, f , is most sensitive to.
1. Compute the gradient f ≡ Xf (X).
2. Compute eigenvalue decomposition of the m × m matrix
C ≡ Eπ(X)[( f )( f ) ]:
C = WΛWT
,
where W is the orthogonal matrix and Λ = diag{λ1, . . . , λm} is the
diagonal matrix of ordered eigenvalues with λ1 ≥ λ2 ≥ · · · ≥ λm ≥ 0.
3. Let W ≡ [W1, W2], and Λ ≡ blockdiag{Λ1, Λ2}, where Λ1 contains
the largest d eigenvalues and W1 contains the corresponding d
eigenvectors.
The column space of W1 is called active subspace.
S = W1X is called active variable.
U = W2X is called inactive variable.
X = W1S + W2U.
Emulator Group February 2018 15 / 25
40. Method Dimension Reduction for X
Therefore, we want to detect subspace Rd ⊂ Rm that the
loglikelihood of Y, f , is most sensitive to.
1. Compute the gradient f ≡ Xf (X).
2. Compute eigenvalue decomposition of the m × m matrix
C ≡ Eπ(X)[( f )( f ) ]:
C = WΛWT
,
where W is the orthogonal matrix and Λ = diag{λ1, . . . , λm} is the
diagonal matrix of ordered eigenvalues with λ1 ≥ λ2 ≥ · · · ≥ λm ≥ 0.
3. Let W ≡ [W1, W2], and Λ ≡ blockdiag{Λ1, Λ2}, where Λ1 contains
the largest d eigenvalues and W1 contains the corresponding d
eigenvectors.
The column space of W1 is called active subspace.
S = W1X is called active variable.
U = W2X is called inactive variable.
X = W1S + W2U.
Emulator Group February 2018 15 / 25
41. Method Dimension Reduction for X
Some More Details
Given data (X1, Y1), . . . , (Xn, Yn), we can use Monte Carlo
approximation to identify the active subspace by
calculating fi ≡ Xi
f (Xi ) for i = 1, . . . , n;
computing ˆC ≡ 1/n
n
i=1 fi fi .
The input space is then changed from the original Rm to the active
subspace within Rd .
The value of d can be determined based on the gap of eigenvalues
and the summation d
j=1 λj / m
j=1 λj .
Emulator Group February 2018 16 / 25
42. Method Dimension Reduction for X
Some More Details
Given data (X1, Y1), . . . , (Xn, Yn), we can use Monte Carlo
approximation to identify the active subspace by
calculating fi ≡ Xi
f (Xi ) for i = 1, . . . , n;
computing ˆC ≡ 1/n
n
i=1 fi fi .
The input space is then changed from the original Rm to the active
subspace within Rd .
The value of d can be determined based on the gap of eigenvalues
and the summation d
j=1 λj / m
j=1 λj .
Emulator Group February 2018 16 / 25
43. Method Dimension Reduction for X
Some More Details
Given data (X1, Y1), . . . , (Xn, Yn), we can use Monte Carlo
approximation to identify the active subspace by
calculating fi ≡ Xi
f (Xi ) for i = 1, . . . , n;
computing ˆC ≡ 1/n
n
i=1 fi fi .
The input space is then changed from the original Rm to the active
subspace within Rd .
The value of d can be determined based on the gap of eigenvalues
and the summation d
j=1 λj / m
j=1 λj .
Emulator Group February 2018 16 / 25
44. Method Dimension Reduction for X
Results of Dimension Reduction of X based on PCA
Figure: Eigenvalues based on PCA
Emulator Group February 2018 17 / 25
45. Method Dimension Reduction for X
Results of Dimension Reduction of X based on AS
Figure: Eigenvalues based on AS
Emulator Group February 2018 18 / 25
46. Method Dimension Reduction for X
Results based on PCA, cont’d
Figure: Eigenvectors based on PCA
Emulator Group February 2018 19 / 25
47. Method Dimension Reduction for X
Results based on AS, cont’d
Figure: Eigenvectors based on AS
Emulator Group February 2018 20 / 25
48. Method GP Emulation
Gaussian Process (GP) Emulation
Recall that ξi = (ξi1, ξi2, ξi3) denotes the K = 3
j=1 Kj FPC scores
for the ith sounding, i = 1, . . . , n.
The active variable corresponding to Xi is denoted by Si = W1Xi .
Let B denote a b-dimensional vector of other parameters. Here, for
OCO-2 emulation problem, it includes solar zenith angle, solar
azimuth angle, instrument zenith angle, instrument azimuth angle.
We assume a GP model on ξ(S; B):
ξ(S; B) ∼ GP{µ, Σ(·, ·)}
µ ∈ RK
is the mean vector;
Σ(·, ·) : Rd+b
× Rd+b
→ RK×K
is the corresponding cross-covariance
matrix function.
We will begin with commonly used parametric cross-covariance
functions discussed in Genton and Kleiber (2015).
Emulator Group February 2018 21 / 25
49. Method GP Emulation
Gaussian Process (GP) Emulation
Recall that ξi = (ξi1, ξi2, ξi3) denotes the K = 3
j=1 Kj FPC scores
for the ith sounding, i = 1, . . . , n.
The active variable corresponding to Xi is denoted by Si = W1Xi .
Let B denote a b-dimensional vector of other parameters. Here, for
OCO-2 emulation problem, it includes solar zenith angle, solar
azimuth angle, instrument zenith angle, instrument azimuth angle.
We assume a GP model on ξ(S; B):
ξ(S; B) ∼ GP{µ, Σ(·, ·)}
µ ∈ RK
is the mean vector;
Σ(·, ·) : Rd+b
× Rd+b
→ RK×K
is the corresponding cross-covariance
matrix function.
We will begin with commonly used parametric cross-covariance
functions discussed in Genton and Kleiber (2015).
Emulator Group February 2018 21 / 25
50. Method GP Emulation
Gaussian Process (GP) Emulation
Recall that ξi = (ξi1, ξi2, ξi3) denotes the K = 3
j=1 Kj FPC scores
for the ith sounding, i = 1, . . . , n.
The active variable corresponding to Xi is denoted by Si = W1Xi .
Let B denote a b-dimensional vector of other parameters. Here, for
OCO-2 emulation problem, it includes solar zenith angle, solar
azimuth angle, instrument zenith angle, instrument azimuth angle.
We assume a GP model on ξ(S; B):
ξ(S; B) ∼ GP{µ, Σ(·, ·)}
µ ∈ RK
is the mean vector;
Σ(·, ·) : Rd+b
× Rd+b
→ RK×K
is the corresponding cross-covariance
matrix function.
We will begin with commonly used parametric cross-covariance
functions discussed in Genton and Kleiber (2015).
Emulator Group February 2018 21 / 25
51. Method GP Emulation
Gaussian Process (GP) Emulation
Recall that ξi = (ξi1, ξi2, ξi3) denotes the K = 3
j=1 Kj FPC scores
for the ith sounding, i = 1, . . . , n.
The active variable corresponding to Xi is denoted by Si = W1Xi .
Let B denote a b-dimensional vector of other parameters. Here, for
OCO-2 emulation problem, it includes solar zenith angle, solar
azimuth angle, instrument zenith angle, instrument azimuth angle.
We assume a GP model on ξ(S; B):
ξ(S; B) ∼ GP{µ, Σ(·, ·)}
µ ∈ RK
is the mean vector;
Σ(·, ·) : Rd+b
× Rd+b
→ RK×K
is the corresponding cross-covariance
matrix function.
We will begin with commonly used parametric cross-covariance
functions discussed in Genton and Kleiber (2015).
Emulator Group February 2018 21 / 25
52. Ongoing Work and Discussion
Ongoing Work
We will carry out empirical studies to investigate how well the
approximations (dimension reductions in both Y and X and the
emulator) retain most of the information about the input-output
behavior.
We will compare our method with other techniques in terms of
efficiency and accuracy.
We are also investigating the theoretical properties of the proposed
approach.
Our method can be applied to set up simulation studies for
uncertainty quantification of retrieval algorithms, sensitivity analysis
and calibration.
Emulator Group February 2018 22 / 25
53. Ongoing Work and Discussion
Ongoing Work
We will carry out empirical studies to investigate how well the
approximations (dimension reductions in both Y and X and the
emulator) retain most of the information about the input-output
behavior.
We will compare our method with other techniques in terms of
efficiency and accuracy.
We are also investigating the theoretical properties of the proposed
approach.
Our method can be applied to set up simulation studies for
uncertainty quantification of retrieval algorithms, sensitivity analysis
and calibration.
Emulator Group February 2018 22 / 25
54. Ongoing Work and Discussion
Ongoing Work
We will carry out empirical studies to investigate how well the
approximations (dimension reductions in both Y and X and the
emulator) retain most of the information about the input-output
behavior.
We will compare our method with other techniques in terms of
efficiency and accuracy.
We are also investigating the theoretical properties of the proposed
approach.
Our method can be applied to set up simulation studies for
uncertainty quantification of retrieval algorithms, sensitivity analysis
and calibration.
Emulator Group February 2018 22 / 25
55. Ongoing Work and Discussion
Ongoing Work
We will carry out empirical studies to investigate how well the
approximations (dimension reductions in both Y and X and the
emulator) retain most of the information about the input-output
behavior.
We will compare our method with other techniques in terms of
efficiency and accuracy.
We are also investigating the theoretical properties of the proposed
approach.
Our method can be applied to set up simulation studies for
uncertainty quantification of retrieval algorithms, sensitivity analysis
and calibration.
Emulator Group February 2018 22 / 25
56. Ongoing Work and Discussion
Discussion
The methods we used in this emulation problem can be used to
answer some questions in order to solve the inverse problem in the
retrieval algorithms as well.
When we substitute the emulator for the forward function in the
retrieval algorithm, the associated inverse problem can be solved much
more efficiently, which can be used to warm start the original optimal
estimation procedure.
We will also apply the dimension reduction techniques directly for the
inverse problem which will potentially improve the computational
efficiency.
Incorporating spatial dependence: spatial X/spatial Y
Distributed computation: Model globally, fit locally
Emulator Group February 2018 23 / 25
57. Ongoing Work and Discussion
Discussion
The methods we used in this emulation problem can be used to
answer some questions in order to solve the inverse problem in the
retrieval algorithms as well.
When we substitute the emulator for the forward function in the
retrieval algorithm, the associated inverse problem can be solved much
more efficiently, which can be used to warm start the original optimal
estimation procedure.
We will also apply the dimension reduction techniques directly for the
inverse problem which will potentially improve the computational
efficiency.
Incorporating spatial dependence: spatial X/spatial Y
Distributed computation: Model globally, fit locally
Emulator Group February 2018 23 / 25
58. Ongoing Work and Discussion
Discussion
The methods we used in this emulation problem can be used to
answer some questions in order to solve the inverse problem in the
retrieval algorithms as well.
When we substitute the emulator for the forward function in the
retrieval algorithm, the associated inverse problem can be solved much
more efficiently, which can be used to warm start the original optimal
estimation procedure.
We will also apply the dimension reduction techniques directly for the
inverse problem which will potentially improve the computational
efficiency.
Incorporating spatial dependence: spatial X/spatial Y
Distributed computation: Model globally, fit locally
Emulator Group February 2018 23 / 25
59. Ongoing Work and Discussion
Discussion
The methods we used in this emulation problem can be used to
answer some questions in order to solve the inverse problem in the
retrieval algorithms as well.
When we substitute the emulator for the forward function in the
retrieval algorithm, the associated inverse problem can be solved much
more efficiently, which can be used to warm start the original optimal
estimation procedure.
We will also apply the dimension reduction techniques directly for the
inverse problem which will potentially improve the computational
efficiency.
Incorporating spatial dependence: spatial X/spatial Y
Distributed computation: Model globally, fit locally
Emulator Group February 2018 23 / 25
60. Ongoing Work and Discussion
Discussion
The methods we used in this emulation problem can be used to
answer some questions in order to solve the inverse problem in the
retrieval algorithms as well.
When we substitute the emulator for the forward function in the
retrieval algorithm, the associated inverse problem can be solved much
more efficiently, which can be used to warm start the original optimal
estimation procedure.
We will also apply the dimension reduction techniques directly for the
inverse problem which will potentially improve the computational
efficiency.
Incorporating spatial dependence: spatial X/spatial Y
Distributed computation: Model globally, fit locally
Emulator Group February 2018 23 / 25
61. Acknowledgements
This research was partially supported by the National Science
Foundation under Grant DMS-1638521 to the Statistical and Applied
Mathematical Sciences Institute (SAMSI). Any opinions, findings, and
conclusions or recommendations expressed in this material do not
necessarily reflect the views of the National Science Foundation.
We would like to thank other members of the Working Group on
Remote Sensing in the SAMSI Program on Mathematical and
Statistical Methods for Climate and the Earth System(CLIM) for
helpful discussion and suggestions.
Emulator Group February 2018 24 / 25