Lecture50

THREE LINEAR ALGEBRA APPLICATIONS
11/28/2018, MATH 121 GUEST LECTURE
Three Applications of Linear algebra
Rather than grinding through a laundry list of applications, we focus on three parts,
where linear algebra plays a role. The first is an unsolved problem in complexity
theory of arithmetic, the second is a short overview how data structures and data
storage rely on notions put forward by linear algebra. The third is a spectral problem
in graph theory which is related to networks. The lecture will conclude with a slide
show showing off some applications without going into details.
1. Fast multiplication
1.1. An important unsolved problem in computer science is the question, how fast one
can multiply two integers. The grade-school multiplication of two n-digit numbers uses
about n2
computation steps. Modern computer algebra frame works use more sophis-
ticated methods like the GNU multiple precision arithmetic library. These methods
chose between different methods.
1.2. The Karatsuba method uses the idea that the one can write (a + b)(c + d) =
u+(u+w−v)+w where u = ac, w = bd, v = (b−a)(d−c). Instead of 4 multiplications,
we need only 3. This cuts the O(n2
) complexity to O(nlog2(3)
) = O(n1.5849
) which is
faster. For example, to multiply x = 1234 with y = 5577, build u = 12 ∗ 55 = 660
and w = 34 ∗ 77 = 2618, and v = (33 − 12)(77 − 55) = 22 ∗ 22 = 484, then compute
u ∗ 10000 + v100 + w = 6600000 + (660 + 2618 − 484) ∗ 100 + 2618 = 6882018.
1.3. I learned the following method from a talk of Albrecht Beutelspacher. If
numbers are close to a round number like 1000, there is a neat trick to make the
computation. Assume for example, we have two 3 digit numbers x, y so that the
product of 1000−a and 1000−b is smaller than 1000. Then use (1000−a)(1000−b) =
1000(1000 − a − b) + ab = 1000(x − b) + ab. For example, to compute 991 ∗ 899 build
x − b = 991 − 101 = 890 and ab = 9 ∗ 101 = 909. The result is 890909.
1.4. The fastest multiplication methods are based on linear algebra. Given again two
integers like x = 1234 and y = 5577. If n is the number of digits in x and y, then we
need about n2
operations to get the product. It turns out that there is a faster way
to do that. And this involves linear algebra. The idea is to write the numbers x, y as
functions, then multiply the functions using Fourier theory. How does this work?

Linear Algebra Applications
1.5. If xk and yk are the integer coeﬃcients of x and y so that x = n−1
k=0 xk10k
and
y = n−1
k=0 yk10k
, we look at the functions f(z) = n−1
k=0 xkzk
and g(z) = n
k=0 ykzk
.
We assume n is the number of digits in xy and that xk, yk are zero for k larger than
the number of digits of x or y. Now, if h(z) = f(z)g(z), then h(10) is the product.
The point is that we can compute n data points of h with n log(n) operations and that
we can get from n data points of a polynomial the function back in n log(n) operators.
So, we can compute the product xy in n log(n) operations. How do we achieve this
miracle?
1.6. Given a polynomial f, we can look at the periodic function F(t) = f(e2πit
)/
√
n
and evaluate Xk = F(2πk/n) for k = 0, . . . , (n − 1). This produces a linear map
U : (x0, . . . , xn−1) → (X0, . . . , Xn−1) which is called the discrete Fourier trans-
form. There is a fast Fourier transform implementation which does this in n log(n)
computation steps. If U(y0, . . . , yn−1) = (Y0, . . . , Yn−1) then we have just to compute
Q = U−
1(XY ) and get xy = n
k=0 Qk10k
. Mathematica has the fast Fourier transform
and its inverse implemented. Here is the code:
FastMultiplication [ x , y ]:=Module[{X,Y,U,V, n ,Q} ,
X = Reverse [ IntegerDigits [ x ] ] ; Y = Reverse [ IntegerDigits [ y ] ] ;
n =Length [X]+Length [Y]+1; X=PadRight [X, n ] ; Y=PadRight [Y, n ] ;
U=InverseFourier [X] ; V=InverseFourier [Y] ;
Q=Round[Re[ Fourier [U∗V]∗ Sqrt [ n ] ] ] ;
Sum[Q[ [ k ] ] 10ˆ(k−1), {k , n } ] ] ;
x0 = 11234; y0 = 52342; FastMultiplication [ x0 , y0]==x0∗y0
The procedures Fourier and InverseFourier are implemented already in Mathematica
Here are emulations showing what they do:
X={4 ,2 ,3 ,2 ,1}; n =Length [X] ; InverseFourier [X]
f = Sum[X[ [ k ] ] t ˆ(k−1),{k , n } ] ;
N[ Table [ f /. t −> Exp[−I 2 Pi k/n ] ,{ k ,0 , n−1}]]/ Sqrt [ n ]
N[ ( Table [Exp[ I 2Pi k l /n ] ,{ k ,0 , n−1},{ l ,0 , n−1}]/Sqrt [ n ] ) .X]
Fourier [X]
2. Data structures
2.1. Information is commonly represented in the form of vectors or matrices. This
is sometimes not obvious as data can be complicated. Still, we usually get away with
matrices. For example, if we have names linked to addresses and telephone numbers,
we can make a spread sheet A(i, j), where A(i, 1) is the name and A(i, 2) is the
address and A(i, 3) is the telephone number. In a relational database, information
is stored in tables as matrices containing records as rows and attributes as columns.
2.2. Data can also be stored more geometrically, especially in the form of a graph. A
special case is the hierarchical database in which the graph is a tree, the Khipu knot
encoding used in the Incan empire is an example of a hierarchical database. It is
important to note however that also geometric information like a graph can be stored as
part of a relational data base. A graph with nodes can be encoded with an adjacency
matrix A(i, j) deﬁned as A(i, j) = 1 if i and j are connected and A(i, j) = 0 else.

Figure 1. Khipu from the Inka period. Image Source: Museo Chileno
De Arte Precolombino.
3. Media
3.1. Multimedia like images, sound or movies are stored in vector or matrix form. An
image is an array of pixels, a sound is an array of amplitudes, a movie is given by
an array of pictures and a sound.
3.2. Let us look at a color image A: It is a matrix where each entry A(i, j) contains
three numbers (r, g, b), where r, g, b are red, green and blue color values. Even so this
is a rectangular array of pixels, where each pixel is a vector of numbers, this can be
written as a single vector.
3.3. A sound ﬁle consists of two vectors (l, r), where l is the left channel and r is the
right channel. Each entry can be a signed integers from −127 to 128. A movie can be
seen as a vector of pictures. It is a curve in the large space of all pictures.
Figure 2. 6 Frames from the Movie Samsara (2011). Each frame is a
360 x 854 matrix, where each pixel has a Red,Green and Blue value.

Linear Algebra Applications
4. Networks
4.1. A graph G = (V, E) consists of a collection of nodes V which are connected by
edges collected in E. Graphs in which the direction of the edges matter are also called
digraphs. If one says graph, one usually does not specify directions. A graph can be
encoded as a matrix A, the adjacency matrix of A. It is deﬁned as A(i, j) = 1 if i is
connected to j and A(i, j) = 0 else. This matrix is self-adjoint, A = AT
. Therefore,
the eigenvalues of A are real.
4.2. The analogue of the Laplacian in calculus is the matrix L = B − A, where
B = Diag(d1, · · · , dn) is the diagonal matrix containing the vertex degrees of B in
the diagonal. The matrix L always has the eigenvalue 0, with eigenvector [1, 1, · · · , 1].
If the graph is connected, the rest of the eigenvalues are positive. They are the fre-
quencies one can hear, when one hits the graph. The eigenvectors tell something
about the distribution of electrons if the graph represents a molecule. In the case
of circular graphs, we need the same mathematics than for the fast multiplication of
integers. The circle is closed.
Figure 3. Examples of networks.
Oliver Knill, knill@math.harvard.edu, Math 121, Harvard College, 11/27/18

Lecture50

More Related Content

What's hot (19)

Similar to Lecture50 (20)

Recently uploaded (20)

Lecture50