Sharad D. Silwal [email protected].

edu

Research Statement
My research areas include partial differential equations, Fourier analysis and statistics, of which I
am interested in both pure and applied sides. On the pure side, I am interested in analysis on spaces
of homogeneous type in the context of Moser’s and Krylov-Safonov’s theories on the regularity
of solutions to elliptic PDEs. On the applied side, my broad interest lies in mathematical and
statistical methods in applied sciences, particularly, image processing. I am currently involved in
interdisciplinary research in image quality assessment using tools from Fourier analysis, Statistics
as well as programming languages MATLAB and R. I have a Masters degree in Statistics (2009)
and am expecting a PhD in Mathematics in May 2012. I have delivered over a dozen research
talks in the areas of my research and have been awarded the Outstanding Graduate Student
Academic Award for the academic year 2011/2012 based on my academic accomplishments. I
have published a research article titled “A non-parametric-test-based structural similarity measure
for digital images” in Computational Statistics and Data Analysis and have a few others in the
pipeline, two of which are already submitted. I am very keen on mentoring students as well as
developing curricula on the topics of my research.
I believe in professional development and career advancement through hard work and personal
efforts. I am fully committed to research in an environment where maintaining excellence in un-
dergraduate teaching is of primary importance. I love establishing close rapport with my students
and would love to engage them in my research activities as much as possible.

1. Applied Mathematics and Statistics

I am mainly interested in pursuing research in applied mathematics and statistics. I developed

a passion for interdisciplinary research while I was working on my Masters report in Statistics
alongside my PhD studies in Mathematics. The report was titled “Bayesian inference and wavelet
methods in image processing”, and it featured two image processing techniques, namely, Bayesian
image classification and wavelet denoising along with experimentation using MATLAB. Working
on this report was a true motivation for me to do research in image processing and this has
resulted in a publication and a submitted article on the topic of image quality assessment. I am also
interested in doing research in other areas of image processing as well as exploring interdisciplinary
and collaborative research in other applied sciences.

1.1. Mathematical and Statistical methods in image processing. My research in image

processing mainly extends classical as well as modern tools in Fourier analysis and Statistics to
develop new image analysis methods. A digital image is a two-dimensional finite signal, or, in other
words, a matrix. Fourier analysis is the study of decomposition and reconstruction of signals and
thus provides tools to represent images in alternate domains that allow for a better analysis than
when dealt with in the raw pixel domain. For example, various filtering techniques are applied on
the Fourier domain for various effects such as blurring, deblurring, denoising and sharpening. The
wavelet domain allows for a multi-resolution approximation of an image, the analysis of details
with respect to orientation and a feature of local-adaptiveness. Statistics, on the other hand, plays
a central role in analyzing image data. As with any applied discipline, image processing makes
ample use of Statistics in both theory and practice. For example, in order to measure structural
similarity between images, Pearson’s correlation coefficient was one of the factors of the image
similarity index developed in [16]. The research articles [17] and [14], on the other hand, employ a
statistical hypothesis test for the same purpose, with the former applying it in the spatial domain
and the latter in the wavelet domain.

1.2. Image quality assessment. My current interdisciplinary research is focused on image qual-
ity assessment. My research can be explained by considering the following situation. Suppose, we
have a large database of images, out of which we need to single out the one image which matches
some given reference image best. The most common way of doing this is to choose the image
that is the least Euclidean distance away from the reference image. This distance is given by
Mean Square Error (MSE), which, despite being most convenient mathematically, does not take
into account a redundancy present in images encountered in our daily lives and acquired from
nature itself as opposed to the images which are generated from random matrices having no visual
significance. These images, known as the natural images, are believed to make up a very thin set
inside the space of all possible images. MSE, being based on the Euclidean norm, cannot differ-
entiate between a random matrix and a matrix corresponding to a natural image and thus fails
to provide an accurate measure of visual distance between natural images. Another image quality
assessment method which shares the limitations of MSE is Total Variation (TV) which is more
popular among the PDE-based image processing community. This limitation of MSE and TV is
illustrated in Figure 1. It shows four images which are approximately the same MSE distances
away from a reference image, yet, to the human eye, they clearly appear quite far apart.

MSE= 1600 MSE= 1608

SSIM=.8544 SSIM=.6133
TV=5 TV=812

MSE= 1592 MSE= 1596

SSIM=.2301 SSIM=.1175
TV=6274 TV=9189

Figure 1. Images approximately the same MSE distance (1600 ± 8) away from a
reference image (center). However, they are not visually equidistant.

Thus, the challenge lies in designing a method which is implementable on a computer and at the
same time is as close as possible to the human visual system (HVS) in performance. The HVS is
a highly complicated system that has evolved over millions of years and attempting to mimic it
on a computer poses all sorts of difficulties. For example, one of the features of the HVS is the
ability to see through low-level random noises in images that occur naturally over time or through
normal processes of image acquisition or storage. But such small alterations which are almost
negligible to the human eye (for example, the bottom images in Figure 1) can appear huge to a
computer which is only following an algorithm that does not take into account such properties of

2
the HVS. The seminal work [16] introduces Structure Similarity Index (SSIM) for image quality
assessment where SSIM=1 corresponds to identical images and SSIM close to zero corresponds to
images with significant dissimilarity. A great deal of experimentation has established that SSIM
hugely outperforms MSE in capturing the features of the HVS. However, it can be seen from
Figure 1 that SSIM overly penalizes random noises (bottom images) and is not sensitive enough
to deterministic noises involving local smoothing such as blurring (top right image).

MSE=116 MSE=438
TV=10529 TV=7424
SSIM=.8129 SSIM=.738
PSSIM=.9754 PSSIM=.1042

MSE=118 MSE=105
TV=3731 TV=12878
SSIM=.8604 SSIM=.8534
PSSIM=.996 PSSIM=.9836

Figure 2. Comparison of a reference image (center) with three random noisy and
one blurred versions.

There have been two articles as an outcome of my research in image quality assessment. The first
one [17] is based on a rank-based non-parametric test developed in [19] applied in the pixel domain
and proposes a P-value-based Structure Similarity Index (PSSIM). The second one [14] (submitted)
is based on a different rank-based non-parametric test developed in [18] applied in the wavelet
domain and proposes a Wavelet-based Non-parametric Structure Similarity Index (WNPSSIM).
We propose a method of measuring structural similarity between a reference image X and a test
image Y by applying a statistical hypothesis test to detect the general dependence between X and
the error Z = Y − X. In [17], the null hypothesis is that the conditional distribution of Z given
X is independent of the marginal distribution of X. In [14], first we divide Z into a convenient
collection of vectors Z1 , Z2 , . . . , Zm , denoting Zi = (zij ) for each i and then set the null hypothesis
to be
H0 : F1 (x) = F2 (x) = · · · = Fm (x), eij ∼ Fi (x).

Our method of using hypothesis testing to measure structural similarity between images provides
certain advantages over the method of using correlation, which is the case for SSIM. In fact, the
hypothesis test employed in [17] has advantages not only over correlation-based approaches includ-
ing Pearson, Spearman or Kendall’s correlations, but also over likelihood-based methods such as
linear models, generalized additive models that combine a likelihood approach with local smooth-
ing developed in [20], a rank-based method for linear models proposed in [15], and copula-based
tests of independence given in [5]. More details about these claims can be found in [19]. Below
we provide some key features of our method of hypothesis testing for image quality assessment.

3
(1) We apply the hypothesis test to local blocks of a fixed window that slides horizontally and
vertically across the entire image and then make the decision based on the proportion of
non-rejections.
(2) Due to our interest in non-rejection rather than rejection, a Bonferroni correction or a false
discovery rate control is rendered unnecessary.
(3) The flexibility of the response variable being either continuous or discrete makes the method
valid for all sorts of distortions.
(4) The method allows for heteroscedastic variations among rows or columns which are not possible
with classical regression methods.
(5) The whole procedure is distribution free and resistant to outliers.
Using visual perception as a benchmark, PSSIM has been shown to provide a better evaluation of
image similarity than several popular similarity assessment methods such as SSIM and MSE. It
is also shown to be robust for both random noises and deterministic alterations such as blurring
and watermarking. Figure 2 illustrates the performance of PSSIM vis-à-vis other popular image
quality assessment methods, where PSSIM is seen to assign more realistic similarity indices than
SSIM. This is particularly evident in the case of the blurred image (the top right). WNPSSIM,
on the other hand, is especially designed to capture the seeing-through ability of the HVS when
low-level random noises are present in images.

2. Pure mathematics

My research on pure mathematics is focused on the Harnack theory of elliptic partial differential
equations in spaces of homogeneous type. Having no a priori smooth structure, spaces of homoge-
neous type generalize many spaces encountered in mathematics. Heinonen’s book [7] popularized
the topic of analysis in spaces of homogeneous type by developing a very rich theory on these
spaces including concepts from first-order calculus. Below I will provide a brief description of my
research on this topic.

2.1. Harnack’s inequality in spaces of homogeneous type. In 1887, Carl Gustav Axel Har-
nack in his book [6] on potential theory, introduced Harnack’s inequality as a property verified by
non-negative harmonic functions in the case of n = 2. A version of this inequality in the form it
is currently used in the theory of partial differential equations reads:
Theorem 1. Let Ω ⊂ Rn be a domain. There exists a constant CH = CH (n, Ω) > 1 such that for
any non-negative harmonic function u in Ω and every ball B2R (x0 ) ⊂ Ω, we have
sup u ≤ CH inf u.
BR (x0 ) BR (x0 )

A triad (X, d, µ) is said to be a space of homogeneous type if X is a non-empty set; d is a quasi-

metric on X, i.e., it relaxes the triangle inequality of a metric: there exists a constant K > 1
(K = 1 if d is a metric) such that
d(x, y) ≤ K(d(x, z) + d(z, y)) for every x, y, z ∈ X;
and µ is a measure compatible with the quasi-metric d in the sense that µ is doubling with respect
to the d-balls: there exists a constant Cµ > 1 such that for every d-ball B2r ⊂ X, we have
µ(B2r ) ≤ Cµ µ(Br ).
Examples of spaces of homogeneous type include: (1) Euclidean spaces with a non-isotropic metric,
(2) Boundaries of Lipchitz domains endowed with harmonic measure, (3) Connected Riemannian

4
manifolds with a non-negative Ricci curvature, (4) Real connected Lie groups with polynomial
volume growth, and (5) Graphs with a uniform bound on the number of neighbors of vertices.
Moser’s and Krylov-Safonov’s techniques in the context of divergence form PDEs [12, 13] and
non-divergence form PDEs [9, 10] respectively are regarded as ground-breaking in the study of
regularity problems of solutions to elliptic PDEs. Let Ω ⊂ Rn be a domain, u a real-valued
function on Ω, and A a matrix-valued function on Ω having a uniform ellipticity, i.e., it verifies
for some 0 < λ < Λ,
λ|ξ|2 ≤ hA(x)ξ, ξi ≤ Λ|ξ|2 , x ∈ Ω, ξ ∈ Rn .
In [12, 13], Moser verified Harnack’s inequality for non-negative solutions to divergence-form el-
liptic operators defined as
n
X
Lu := (aij (x)ui )j = div(A(x)∇u), x ∈ Ω.
i,j=1

The key step in Moser’s approach is to get the following estimate:

Z 1q
1 q
sup u ≤ C(n, q, λ, Λ) u , 2B ⊂ Ω, ∀q > 0,
B |B| 2B
for a supersolution u which is accompolished through Moser’s celebrated iterative scheme. An-
other breakthrough in this direction came along when Krylov and Safonov [9, 10] introduced
new measure-theoretic and probabilistic tools to establish Harnack’s inequality for non-negative
solutions to non-divergence-form elliptic operators defined as
n
X
Lu := aij (x)uij = tr (A(x)D2 u), x ∈ Ω.
i,j=1

The central idea in Krylov-Safonov’s technique is a measure-theoretic property about a control on

the size of a function in terms of a prescribed probability. Given M > 1 and ε ∈ (0, 1), a function
u ≥ 0 defined on Ω is said to have the critical density property with constants M and ε if for every
B2r ⊂ Ω, the following holds true:
|{x ∈ B2r : u(x) > M }| ≥ ε|B2r | ⇒ inf u > 1.
Br

Since the introduction Moser’s and Krylov-Safonov’s techniques, several authors have been able
to adapt, extend and generalize these techniques in several contexts. My research in this topic has
been toward an axiomatization of Krylov-Safonov’s approach to Harnack’s inequality in spaces
of homogeneous type, which, combined with some other collaborative work, has resulted in a
submitted article [8].
Theorem 2 gives an axiomatic approach to Harnack’s inequality in spaces of homogeneous type,
the proof of which is based on Krylov-Safonov’s techniques.
Theorem 2 ([8]). Let (X, d, µ) be a doubling quasi-metric space. Let Ω ⊂ X be open. Let KΩ be
a set of non-negative functions defined on X which is closed under multiplication by constants in
(0, 1). Further denote KΩ (BR (x0 )) := {u ∈ KΩ | BR (x0 ) ⊂ dom(u)}. Assume the following:

(i) The continuous functions are dense in L1 (X, dµ).

(ii) KΩ satisfies the critical density property with constants M and ε.

5
(iii) There exist constants % > 0 and CD ≥ 1 such that whenever u ∈ KΩ , u% is a doubling weight
with constant CD , that is,
Z Z
%
u dµ ≤ CD u% dµ,
B2r (x) Br (x)

for every Br (x) with B8Kr (x) ⊂⊂ Ω.

(iv) Whenever u ∈ KΩ (BR (x0 )) and λ ≥ u in BR (x0 ) then λ − u ∈ KΩ (BR (x0 )).
Then KΩ satisfies the Harnack property with constant CH depending only on ε, M , CD , and the
doubling quasi-metric constants K and Cµ .

3. Future work

As I am interested in both pure and applied mathematics, I would like to be engaged in research in
both fields. On the applied side, I will continue to pursue my research in image quality assessment
and will also work on other areas of image processing. An image, being a matrix, after all, is
just data. Consequently, all the data processing tools and theoretical developments in Statistics
naturally have great potential in the realm of research in image processing, which is what I will
explore. On the pure side, I will continue to explore real analysis in spaces of homogeneous type,
especially, in the context of Moser’s and Krylov-Safonov’s theories which stand as cornerstones in
the study of regularity properties of solutions to elliptic PDEs. Their techniques are still being
exploited by many researchers to obtain new results or generalizations in many different contexts
(for example, see [1, 2] for Moser’s and [3, 4] for Krylov-Safonov’s).
I hope to design special courses in the topics of my research interests so as to bring my students on
board with my research activities. I would also be very interested to pursue collaborative research
in other areas of applied mathematics related to partial differential equations, Fourier analysis
and Statistics. I am very willing to learn new material in order to expand my research potential.
Some specific research projects that I intend to pursue in the near future are as follows:
(1) Traditional image reconstruction methods are based on the Euclidean distance metric, or
the MSE error. However, as shown in Figure 1, MSE is a poor measure of image similarity
relative to the HVS. I wish to develop image reconstruction methods based on our image
quality criteria. This amounts to solving an optimization problems expressed in terms of
image similarity indices PSSIM and WNPSSIM.
(2) Liao and Akritas [11] proposed a strong statistical classification methodology which do not
require the data to be normally distributed. I am interested in developing an image segmenta-
tion method based on the discrete wavelet transform and the test-based classification method
in [11].
(3) Parts of my applied projects such as experimentation and performance evaluation are quite
suitable to be carried out as undergraduate research. Undergraduate students can also be
recruited to conduct psycho-visual experiments to come up with true image similarity indices
based on the HVS for a standard image database that could be used as reference by the
research community in image quality assessment (discussed in [14]).
(4) Parabolic PDEs are evolution equations involving an additional time variable. Yet, many
results for elliptic PDEs have been extended to the case of parabolic PDEs by adapting the
role played by the time variable in terms of a time lag, i.e., having a bound for the value of
some function at a given time in terms of its value at a later time. Motivated by the results
in the elliptic case, my future research in this regard will be to work on an axiomatization of
Krylov-Safonov’s approach to parabolic Harnack’s inequality in spaces of homogeneous type.

6
(5) Harnack’s inequality for harmonic functions on graphs was proved by Delmotte [2] using
Moser’s iteration techniques. I wish to employ Krylov-Safonov’s techniques on graphs in
order to extend the result in [2].
R 1/s
1 s
(6) Given a function w > 0, a ball B ⊂ Ω and s ∈ [−∞, ∞], define w(s, B) := |B| B w as
the s-mean of w over B, whenever it exists. Many well-known classes such as the Harnack
class, Muckenhoupt’s Ap (Ω), p > 1 classes and reverse Hölder classes RHs (Ω), s > 1 are
functions satisfying reverse inqualities of the s-mean for various values of s. I wish to study
the properties exhibited by such functions by adopting a visual approach using diagrams for
better exposition.

References
[1] A. Björn and N. Marola, Moser iteration for (quasi)minimizers on metric spaces, Manuscripta Math., 121,
(2006), 339–366.
[2] T. Delmotte, Inégalité de Harnack elliptique sur les graphes, Colloq. Math. 72, (1997), no. 1, 19-37.
[3] G. Di Fazio, C. Gutiérrez, and E. Lanconelli, Covering theorems, inequalities on metric spaces and applications
to PDE’s, Math. Ann. 341, (2008), 255–291.
[4] L. Forzani and D. Maldonado, A mean-value inequality for non-negative solutions to the linearized Monge-
Ampère equation, Potential Anal. 30 (3), (2009), 251-270.
[5] C. Genest and B. Rémillard, Tests of independence and Randomness Based on the Empirical Copula Process,
Test, 13 (2004), 335–369.
[6] C. G. A. Harnack, Die Grundlagen der Theorie des logarithmischen Potentiales und der eindeutigen Poten-
tialfunktion in der Ebene, Teubner, Leipzig, Germany, 1887.
[7] J. Heinonen, Lectures on Analysis on Metric Spaces. 1st edition. Springer Verlag. 2000.
[8] S. Indratno, D. Maldonado and S. Silwal, On the axiomatic approach to Harnack’s inequality in doubling
quasi-metric spaces, submitted.
[9] N. Krylov and M. Safonov, An estimate on the probability that a diffusion hits a set of positive measure, Soviet
Math. 20, (1979), 253–256.
[10] N. Krylov and M. Safonov, A property of the solutions of parabolic equations with measurable coefficients, Izv.
Akad. Nauk SSR Ser. Mat. 44 (1980), no.1, 161-175.
[11] S. Liao, M. Akritas, Test-based classification: A linkage between classification and statistical testing, Statistics
& Probability Letters, 77 (2007), 1269–1281.
[12] J. Moser, On Harnack’s theorem for elliptic differential equations, Comm. Pure and Appl. Math. 14, (1961),
577–591.
[13] J. Moser, On a pointwise estimate for parabolic differential equations, Comm. Pure Appl. Math. 24 (1971),
727–740.
[14] S. Silwal, H. Wang and D. Maldonado, Random-noise contamination assessment of images via testing on
wavelet coefficients, submitted.
[15] J. T. Terpstra and J.W. Mckean, Rank-Based Analyses of Linear Models Using R, Journal of Statistical
Software, 14 (2005), http://www.jstatsoft.org.
[16] Z. Wang, A. Bovik, H. Sheikh and E. Simoncelli, Image quality assessment: From error visibility to structural
similarity, IEEE Trans. Image Processing, 13(4) (2004), 600–612.
[17] H. Wang, D. Maldonado, S. Silwal, A non-parametric-test-based structural similarity measure for digital images,
Computational Statistics & Data Analysis, 55(11), (2011), 2925–2936.
[18] H. Wang, J. Neill and F. Miller, Nonparametric clustering of functional data, Statistics and Its Interface.
1(2008), 47–62 .
[19] H. Wang, S. Tolos and S. Wang, A distribution-free nonparametric test to detect dependence between a response
variable and covariate in presence of heteroscedastic treatment effects, Canadian Journal of Statistics. 38(3)
(2010), 408–433.
[20] N. Wood, Fast stable direct fitting and smoothness selection for generalized additive models, J. R. Stat. Soc.
Ser. B, 70 (2008), 495–518.