Function and Data Approximation

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili

November 28, 2024

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 1 / 87
 Content

1 Interpolation

2 Introduction to Fourier Transform

3 Phase Space and the Fourier Transform

4 Conclusion

5 Least Square Approximation

6 Importance of Approximation in Real-World Applications

7 References

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 2 / 87
 Interpolation

Weierstrass Approximation Theorem

Theorem
Suppose that f is defined and continuous on [a, b]. For each ε > 0, there exists a
polynomial P (x) such that

|f (x) − P (x)| < ε, for all x ∈ [a, b].

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 3 / 87
 Interpolation

A function is said to interpolate a set of data points if it passes through those

points.
Definition
The function y = P (x) interpolates the data points (x1 , y1 ), . . . , (xn , yn ) if
P (xi ) = yi for each 1 ≤ i ≤ n.

Note that P is required to be a function; that is, each value x corresponds to a

single y.This puts a restriction on the set of data points {(xi , yi )}that can be
interpolated—the xi ’s must be all distinct in order for a function to pass through
them. There is no such restriction on the yi ’s.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 4 / 87
 Interpolation

Lagrange interpolation

There is an explicit formula, called the Lagrange interpolating formula, for writing
down a polynomial of degree d = n − 1 that interpolates the points. Suppose that
we are given three points (x1 , y1 ), (x2 , y2 ), (x3 , y3 ).Then the polynomial

(x − x2 )(x − x3 ) (x − x1 )(x − x3 ) (x − x1 )(x − x2 )

P2 (x) = y1 + y2 + y3
(x1 − x2 )(x1 − x3 ) (x2 − x1 )(x2 − x3 ) (x3 − x1 )(x3 − x2 )

is the Lagrange interpolating polynomial for these points.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 5 / 87
 Interpolation

In general, suppose that we are presented with n points (x1 , y1 ), ..., (xn , yn ). For
each k between 1 and n, define the degree n − 1 polynomial

(x − x1 ) · · · (x − xk−1 )(x − xk+1 ) · · · (x − xn )

Lk (x) =
(xk − x1 ) · · · (xk − xk−1 )(xk − xk+1 ) · · · (xk − xn )

Lk (xk ) = 1, while Lk (xj ) = 0, where xj is any of the other data points.

The degree n − 1 polynomial

Pn−1 (x) = y1 L1 (x) + · · · yn Ln (x)

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 6 / 87
 Interpolation

Main Theorem of Polynomial Interpolation

Theorem
Let (x1 , y1 ), . . . , (xn , yn ) be n points in the plane with distinct xi . Then there
exists one and only one polynomial P of degree n − 1 or less that satisfies

P (xi ) = yi for i = 1, . . . , n.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 7 / 87
 Interpolation

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 8 / 87
 Interpolation

Newton’s divided differences

Assume that the data points come from a function f (x), so that our goal is to
interpolate (x1 , f (x1 )), ..., (xn , f (xn )).

Definition
Denote by f [x1 , . . . , xn ] the coefficient of the xn−1 term in the (unique)
polynomial that interpolates (x1 , f (x1 )), . . . , (xn , f (xn )).

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 9 / 87
 Interpolation

Newton’s divided difference formula

P (x) = f [x1 ] + f [x1 , x2 ](x − x1 )
+ f [x1 , x2 , x3 ](x − x1 )(x − x2 )
+ f [x1 , x2 , x3 , x4 ](x − x1 )(x − x2 )(x − x3 )
+ ···
+ f [x1 , . . . , xn ](x − x1 ) · · · (x − xn−1 ).

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 10 / 87
 Interpolation

f [xk ] = f (xk )
f [xk+1 ] − f [xk ]
f [xk , xk+1 ] =
xk+1 − xk
f [xk+1 , xk+2 ] − f [xk , xk+1 ]
f [xk , xk+1 , xk+2 ] =
xk+2 − xk
f [xk+1 , xk+2 , xk+3 ] − f [xk , xk+1 , xk+2 ]
f [xk , xk+1 , xk+2 , xk+3 ] = .
xk+3 − xk

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 11 / 87
 Interpolation

The recursive definition of the Newton’s divided differences allows arrangement

into a convenient table. For three points the table has the form

For data points (0, 1), (2, 2), (3, 4)

1 1
P (x) = 1 + (x − 0) + (x − 0)(x − 2)
2 2

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 12 / 87
 Interpolation

How many degree d polynomials pass through n points?

if we have n points, from previous theorem there is one interpolating polynomial

of degree n − 1 or less.
But
how many degree n polynomials interpolate the same n points?
We can add a (n + 1)-th point. Extending the Newton’s divided difference sceme
gives a new top coefficient.
let add (1, 0) point to the previous example and we got
1
P3 (x) = P2 (x) − (x − 0)(x − 2)(x − 3)
2
So, there is at least one degree 3 polynomial passing through our three original
points.
Also,we can see that there are infinitely many degree n polynomial passing
through the given points depending on how we choose additional points.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 13 / 87
 Interpolation

Interpolation Error

Assume that P (x) is the (degree n − 1 or less) interpolating polynomial fitting the
n points (x1 , y1 ) , . . . , (xn , yn ). The interpolation error is

(x − x1 ) (x − x2 ) · · · (x − xn ) (n)
f (x) − P (x) = f (c),
n!
where c lies between the smallest and largest of the numbers x, x1 , . . . , xn .

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 14 / 87
 Interpolation

Let P (x) denote the (unique) polynomial that interpolates

(x1 , f (x1 )), . . . , (xn , f (xn )), denote by f [x1 , . . . , xn ] the degree n − 1 coefficient
of P (x). Thus,

P (x) = a0 + a1 x + a2 x2 + · · · + an−1 xn−1 ,

where an−1 = f [x1 , . . . , xn ], and two facts are readily apparent.

Fact
f [x1 , . . . , xn ] = f [σ(x1 ), . . . , σ(xn )] for any permutation σ of the xi .

Fact
P (x) can be written in the form:

P (x) = c0 +c1 (x−x1 )+c2 (x−x1 )(x−x2 )+· · ·+cn−1 (x−x1 )(x−x2 ) · · · (x−xn−1 ).

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 15 / 87
 Interpolation

Runge Phenomenon

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 16 / 87
 Interpolation

Chebyshev’s theorem

The motivation for Chebyshev interpolation is to improve control of the maximum

value of the interpolation error

(x − x1 )(x − x2 ) · · · (x − xn ) (n)
· f (c)
n!
on the interpolation interval.Let’s fix the interval to be [−1, 1]. The numerator of
the above interpolation is a degree n polynomial ans has some maximum value on
[−1, 1. We have to try and consider minimax problem.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 17 / 87
 Interpolation

Theorem
The choice of real numbers −1 ≤ x1 , . . . , xn ≤ 1 that makes the value of

max |(x − x1 ) · · · (x − xn )|
−1≤x≤1

as small as possible is

(2i − 1)π
xi = cos for i = 1, . . . , n
2n
and the minimum value is 1/2n−1 . In fact, the minimum is achieved by
1
(x − x1 ) · · · (x − xn ) = Tn (x)
2n−1
where Tn (x) denotes the degree n Chebyshev polynomial.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 18 / 87
 Interpolation

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 19 / 87
 Interpolation

Chebishev Polynomials

Define the nth Chebyshev polynomial by Tn (x) = cos(narccosx).

Fact
The Tn ’s are polynomials. Since T3 is a polynomial combination of T1 and T2 , T3
is also a polynomial. The same argument goes for all Tn . The first few Chebyshev
polynomials (see Figure 3.9) are

T0 (x) = 1
T1 (x) = x
T2 (x) = 2x2 − 1
T3 (x) = 4x3 − 3x

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 20 / 87
 Interpolation

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 21 / 87
 Interpolation

Fact
deg(Tn ) = n,and the leading coefficient is 2n−1 .
This is clear for n = 1 and n = 2, and the recursion relation extends the fact to all
n.

Fact
Tn (1) = 1 and Tn (−1) = (−1)n . Both are clear for n = 1 and 2. In general,
Tn+1 (1) = 2(1)Tn (1) − Tn−1 (1) = 2(1) − 1 = 1
Tn+1 (−1) = 2(−1)Tn (−1) − Tn−1 (−1)
= −2(−1)n − (−1)n−1
= (−1)n−1 (2 − 1) = (−1)n−1 = (−1)n+1

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 22 / 87
 Interpolation

Fact
The maximum absolute value of Tn (x) for −1 ≤ x ≤ 1 is 1. This follows
immediately from the fact that Tn (x) = cos(y) for some y.

Fact
All zeros of Tn (x) are located between -1 and 1 . See Figure 3.10. In fact, the
zeros are the solution of 0 = cos(n arccos x). Since cos y = 0 if and only if y =
odd integer ·(π/2), we find that

n arccos x = odd · π/2

odd · π
x = cos
2n

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 23 / 87
 Interpolation

Fact
Tn (x) alternates between −1  and 1a total of n + 1 times. In fact, this happens
at cos(0), cos n , . . . , cos (n−1)π
π


n , cos(π).

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 24 / 87
 Interpolation

Change of interval

(b−a)
Stretch the points by the factor 2 (the ratio of the two interval lengths)
(b+a)
Translate the points by 2 to move the center of mass from 0 to the
midpoint of [a, b].
On the interval [a, b],

b+a b−a (2i − 1)π

xi = + cos
2 2 2n
for i = 1, . . . , n. The inequality
b−a n


2
|(x − x1 ) · · · (x − xn )| ≤
2n−1
holds on [a, b].

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 25 / 87
 Interpolation

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 26 / 87
 Interpolation

Cubic Splines

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 27 / 87
 Interpolation

Properties of Splines

Assume that we are given the n data points (x1 , y1 ), ..., (xn , yn ), where the xi are
distinct and in increasing order.
A cubic spline S(x) through the data points (x1 , y1 ), ..., (xn , yn ) is a set of cubic
polynomials
2 3
S1 (x) = y1 + b1 (x − x1 ) + c1 (x − x1 ) + d1 (x − x1 ) on [x1 , x2 ]
2 3
S2 (x) = y2 + b2 (x − x2 ) + c2 (x − x2 ) + d2 (x − x2 ) on [x2 , x3 ]
..
.
2 3
Sn−1 (x) = yn−1 + bn−1 (x − xn−1 ) + cn−1 (x − xn−1 ) + dn−1 (x − xn−1 )

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 28 / 87
 Interpolation

Property 1: Si (xi ) = yi and Si (xi+1 ) = yi+1 for i = 1, . . . , n − 1.

′
Property 2: Si−1 (xi ) = Si′ (xi ) for i = 2, . . . , n − 1.
′′
Property 3: Si−1 (xi ) = Si′′ (xi ) for i = 2, . . . , n − 1.
Property 4a (Natural spline): S1′′ (x1 ) = 0 and Sn−1
′′
(xn ) = 0.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 29 / 87
 Interpolation

Radial Basis Functions

let’s start with a simple example:

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 30 / 87
 Interpolation

Definition

A radial basis function (RBF) is a real-valued function φ whose value depends

only on the distance between the input and some fixed point, either the origin, so
that

φ(x) = φ̂(∥x∥),
or some other fixed point c, called a center, so that

φ(x) = φ̂(∥x − c∥).

Any function φ that satisfies the property

φ(x) = φ̂(∥x∥),
is a radial function. The distance is usually Euclidean distance, although other
metrics are sometimes used. Radial basis functions are often used as a collection
{φk }k which forms a basis for some function space of interest, hence the name.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 31 / 87
 Interpolation

A radial function is a function φ : [0, ∞) → R. When paired with a norm on a

vector space ∥ · ∥ : V → [0, ∞), a function of the form

φc = φ(∥x − c∥)
is said to be a radial kernel centered at c ∈ V . A radial function and the
associated radial kernels are said to be radial basis functions if, for any finite set of
nodes

{xk }nk=1 ⊆ V,
all of the following conditions are true:
The kernels

φx1 , φx2 , . . . , φxn

are linearly independent (for example, φ(r) = r2 in V = R is not a radial basis
function).

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 32 / 87
 Interpolation

The kernels φx1 , φx2 , . . . , φxn form a basis for a Haar Space, meaning that the
interpolation matrix (given below) is non-singular:
 
φ(∥x1 − x1 ∥) φ(∥x2 − x1 ∥) . . . φ(∥xn − x1 ∥)
 φ(∥x1 − x2 ∥) φ(∥x2 − x2 ∥) . . . φ(∥xn − x2 ∥) 
 
 .. .. .. .. 
 . . . . 
φ(∥x1 − xn ∥) φ(∥x2 − xn ∥) . . . φ(∥xn − xn ∥)

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 33 / 87
 Interpolation

Examples
Commonly used types of radial basis functions include (writing r = ∥x − xi ∥ and
using ε to indicate a shape parameter that can be used to scale the input of the
radial kernel):
2
φ(r) = e−(εr)
1
φ(r) =
p
φ(r) = 1 + (εr)2 1 + (εr)2

Figure: Multiquadratic
Figure: Gaussian Radial Figure: Inverse Multiquadric
Radial Basis Function
Basis Function Radial Basis Function

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 34 / 87
 Interpolation

Radial basis functions are typically used to build up function approximations of the
form
N
X
y(x) = wi φ(∥x − xi ∥),
i=1

where the approximating function y(x) is represented as a sum of N radial basis

functions, each associated with a different center xi , and weighted by an
appropriate coefficient wi . The weights wi can be estimated using the matrix
methods of linear least squares, because the approximating function is linear in the
weights wi .

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 35 / 87
 Introduction to Fourier Transform

Euler’s Formula

Let’s revisit one of the most iconic and fundamental formulas in mathematics:
Euler’s formula, which elegantly connects complex exponentials with
trigonometric functions:
eiθ = cos θ + i sin θ
This relationship allows us to express complex numbers in polar form, linking their
magnitude and phase to trigonometric terms.
For example, the complex number z = eiπ corresponds to z = −1, which also
gives us the celebrated identity:

eiπ + 1 = 0

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 36 / 87
 Introduction to Fourier Transform

Primitive Roots of Unity

Now let’s go to special subset of complex numbers with magnitude 1. A complex

number z is an n-th root of unity if it satisfies:

zn = 1

On the real number line, there are only two roots of unity, −1 and 1. However, in
the complex plane, there are many more. For example, the imaginary unit i is a
4th root of unity, because:
i4 = (−1)2 = 1
An n-th root of unity is called *primitive* if it is not a k-th root of unity for any
k < n. For instance: −1 is a primitive second root of unity but a non-primitive
fourth root of unity.
For any integer n, the complex number:

ωn = e−i2π/n

is a primitive n-th root of unity.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 37 / 87
 Introduction to Fourier Transform

Key to the Discrete Fourier Transform

From previous slide we have:
1 + ω + ω 2 + ω 3 + · · · + ω n−1 = 0.
This identity follows from the telescoping sum:
(1 − ω)(1 + ω + ω 2 + ω 3 + · · · + ω n−1 ) = 1 − ω n = 0.
Since the first term on the left is not zero, the second must be zero.
A similar method of proof shows that:
1 + ω 2 + ω 4 + ω 6 + · · · + ω 2(n−1) = 0,
1 + ω 3 + ω 6 + ω 9 + · · · + ω 3(n−1) = 0,
..
.
1 + ω n−1 + ω (n−1)2 + ω (n−1)3 + · · · + ω (n−1)(n−1) = 0.
In contrast, for the following sum:
1 + ω n + ω 2n + ω 3n + · · · + ω n(n−1) = 1 + 1 + 1 + 1 + · · · + 1 = n.
Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 38 / 87
 Introduction to Fourier Transform

Primitive Roots of Unity

Lemma
Let ω be a primitive nth root of unity and k be an integer. Then
n−1
(
X
jk n if nk is an integer
ω =
j=0
0 otherwise

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 39 / 87
 Introduction to Fourier Transform

Discrete Fourier Transform (DFT)

The Discrete Fourier Transform (DFT) of x = [x0 , . . . , xn−1 ]T is the
n-dimensional vector
y = [y0 , . . . , yn−1 ],
where
n−1
1 X
yk = √ xj ω jk
n j=0

For example,
√ Lemma shows that the DFT of x = [1, 1, . . . , 1] is
y = [ n, 0, . . . , 0].
In matrix terms, this definition says
 0
ω0 ω0
     
y0 a0 + ib0 ω ··· x0
 y1   a1 + ib1  ω 0 ω1 ··· ω n−1   x1 
 
  
 y2   a2 + ib2 
 1 ω 0 ω 2
··· ω 2(n−1)  
= = √    x2  .

n  ..

 ..   ..  .. .. . .
..   .. 
 
 .   .   . . . 
yn−1 an−1 + ibn−1 ω 0 ω n−1 · · · ω (n−1)2 xn−1

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 40 / 87
 Introduction to Fourier Transform

Each yk = ak + ibk is a complex number. The n × n matrix above is called the

Fourier matrix:  0
ω0 ω0

ω ···
ω 0 ω1 ··· ω n−1 
1  0 2
· · · ω 2(n−1) 

Fn = √ ω ω .

n  .. .. .. .. 
 . . . . 
0 n−1 (n−1)2
ω ω ··· ω
Except for the top row, each row of the Fourier matrix adds to zero, and the same
is true for the columns since Fn is a symmetric matrix. The Fourier matrix has an
explicit inverse
 0
ω0 ··· ω0

ω
ω 0 ω −1 · · · ω −(n−1) 
1  0 −2
· · · ω −2(n−1) 

Fn−1 = √ ω ω

n  ..

.. .. .. 
 . . . . 
ω0 ω −(n−1) ··· ω −(n−1)2
and the inverse Discrete Fourier Transform of the vector y is x = Fn−1 y.
Inverse DFT is the matrix of complex conjugates of the entries of Fn :
Fn−1 = Fn∗ .
Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 41 / 87
 Introduction to Fourier Transform

Number of Operations
Definition
The magnitude of a complex vector v is the real number
√
∥v∥ = vT v.

A square complex matrix F is unitary if

FT F = I.
A unitary matrix, like the Fourier matrix, is the complex version of a real
orthogonal matrix. If F is unitary, then
∥Fv∥2 = vT FT Fv = vT v = ∥v∥2 .
Applying the Discrete Fourier Transform is a matter of multiplying by the n × n
matrix Fn , and therefore requires O(n2 ) operations (specifically n2 multiplications
and n(n − 1) additions). The inverse Discrete Fourier Transform, which is applied
by multiplication by Fn−1 , is also an O(n2 ) process.
Fast Fourier Transform, developed from DFT requires significantly fewer
operations.
Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 42 / 87
 Introduction to Fourier Transform

Introduction to Continuous Fourier Transform

The Fourier Transform is a fundamental tool in mathematics and engineering that

converts a function from the time (or spatial) domain into the frequency domain.

Definition (1D Fourier Transform)

Let f (t) ∈ L1 (R), the space of integrable functions. The Fourier Transform of f
is given by: Z ∞
F{f (t)} = F (ω) = f (t)e−iωt dt
−∞

where ω is the angular frequency and t represents time.

The inverse Fourier Transform reconstructs f (t) from F (ω):

Z ∞
1
f (t) = F (ω)eiωt dω
2π −∞

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 43 / 87
 Introduction to Fourier Transform

Parseval’s Theorem in Fourier Transform

Parseval’s Theorem states that the total energy of a signal in the time domain is
equal to its total energy in the frequency domain.

Parseval’s Theorem (Continuous FT)

Z ∞ Z ∞
2 1
|f (t)| dt = |F (ω)|2 dω
−∞ 2π −∞

f (t) is the signal in the time domain.

F (ω) is the Fourier transform of f (t), representing the signal in the
frequency domain.
Interpretation:
The energy of a signal is preserved whether you measure it in the time
domain or in the frequency domain.
This theorem is a form of energy conservation across the domains.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 44 / 87
 Introduction to Fourier Transform

Fourier Inversion Theorem

The Fourier Inversion Theorem states that if a function f (t) and its Fourier
Transform F (ω) exist, then f (t) can be recovered from F (ω).
Fourier Inversion Formula
Given the Fourier Transform of f (t):
Z ∞
F (ω) = f (t)e−iωt dt
−∞

the function f (t) can be recovered by the inverse Fourier transform:

Z ∞
1
f (t) = F (ω)eiωt dω
2π −∞

Interpretation:
The Fourier Transform and its inverse form a pair of operations.
The Fourier Inversion Theorem guarantees that no information is lost when
transforming a function into the frequency domain and then back to the time
domain.
Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 45 / 87
 Introduction to Fourier Transform

Convolution Theorem
The Convolution Theorem states that the Fourier transform of the convolution of
two functions is the product of their Fourier transforms.
Convolution of Two Functions
The convolution of two functions f (t) and g(t) is defined as:
Z ∞
(f ∗ g)(t) = f (τ )g(t − τ ) dτ
−∞

Convolution Theorem
If F (ω) = F{f (t)} and G(ω) = F{g(t)}, then:

F{f ∗ g}(t) = F (ω) · G(ω)

The Fourier transform simplifies convolution to multiplication in the

frequency domain.
This theorem is widely used in signal processing, where convolutions can be
computationally expensive in the time domain, but easier in the frequency
domain.
Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 46 / 87
 Introduction to Fourier Transform

Multivariate Fourier Transform (2D and Higher

Dimensions)

The Fourier Transform can be extended to functions of multiple variables, making

it essential in areas like image processing.

Definition (2D Fourier Transform)

For a function f (x, y), the 2D Fourier Transform is:
Z ∞Z ∞
F (kx , ky ) = f (x, y)e−i(kx x+ky y) dx dy
−∞ −∞

where kx and ky are the spatial frequencies.

Inverse 2D Fourier Transform

Z ∞ Z ∞
1
f (x, y) = F (kx , ky )ei(kx x+ky y) dkx dky
(2π)2 −∞ −∞

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 47 / 87
 Phase Space and the Fourier Transform

Phase Space Representation

The phase space is a crucial concept in both classical and quantum mechanics,
where the state of a system is described by both position and momentum (or time
and frequency in signal processing).

Phase Space Representation

A signal f (t) can be analyzed in both the time and frequency domains
simultaneously using the **Wigner distribution**:
Z ∞ 
τ ∗ τ  −iωτ
Wf (t, ω) = f t+ f t− e dτ
−∞ 2 2

This provides a joint representation of time and frequency, making it useful in

signal analysis and quantum mechanics.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 48 / 87
 Phase Space and the Fourier Transform

Heisenberg Principle

Heisenberg Uncertainty Principle

The Fourier Transform is inherently linked to the uncertainty principle, which
states that a signal cannot be localized in both time and frequency domains
simultaneously. Mathematically, this is given by:
1
∆t · ∆ω ≥
2
where ∆t is the time uncertainty and ∆ω is the frequency uncertainty.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 49 / 87
 Phase Space and the Fourier Transform

Distributions and Fourier Transform

In many applications, we extend the Fourier transform to operate on generalized
functions, also known as distributions.

The Dirac delta function δ(t) is defined by:

Z ∞
δ(t)f (t) dt = f (0)
−∞

Its Fourier transform is:

F{δ(t)} = 1

Fourier Transform of the Heaviside Function

(
1 if t ≥ 0
H(t) =
0 if t < 0
Its Fourier transform is given by:
1
F{H(t)} = + πδ(ω)
iω
Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 50 / 87
 Phase Space and the Fourier Transform

Applications of Fourier Transform

The Fourier Transform is a powerful tool with numerous applications:

Signal Processing: Decomposing signals into frequency components for
filtering and analysis.
Image Processing: Fourier transforms in 2D are used for image filtering,
compression, and reconstruction.
Quantum Mechanics: The Fourier transform relates position and
momentum space representations of wavefunctions.
Partial Differential Equations: Solutions to PDEs, such as the heat
equation, are often obtained using Fourier methods.
Spectral Analysis: The Fourier transform is used to analyze the spectral
content of time-varying signals.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 51 / 87
 Conclusion

Conclusion

The Fourier Transform is a cornerstone of modern analysis, providing a powerful

method for moving between the time (or spatial) and frequency domains.

Key Points
Fourier transform provides a decomposition of functions into sinusoidal
components.
Theorems such as Parseval’s theorem, the convolution theorem, and the
Fourier inversion theorem offer fundamental insights.
Multivariate Fourier transforms extend these concepts to higher-dimensional
spaces.
Phase space allows us to study signals in both time and frequency
simultaneously.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 52 / 87
 Conclusion

Coding Time
Create a Signal
Let’s create two sine waves with given frequencies and combine them into one
signal. We will use two frequencies: 27Hz and 35Hz.

Figure: 27Hz Sine Wave Figure: 35Hz Sine Wave

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 53 / 87
 Conclusion

Visualization

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 54 / 87
 Conclusion

Time to add some noise

Let’s create some noise and add it to the signal.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 55 / 87
 Conclusion

Doing FFT
Using SciPy’s built in Discrete Fourier transform library to get the signal from
Time to Frequency domain.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 56 / 87
 Conclusion

Remove noise in the fourier graph

From the frequency graph we can see that there is two frequencies that is more
common (higher Y-values). Let’s filter out the other frequencies!

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 57 / 87
 Conclusion

Inverse back to Time-Domain data

We have now combined two sine signals, added noise, used Fourier transform to
understand the frequencies, filtered out the noise and inversed back to a clean
signal!

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 58 / 87
 Least Square Approximation

Inconsistent Systems of Equations

Consider a system of linear equations in matrix form:

Ax = b

where:
A ∈ Rm×n is a matrix of coefficients
x ∈ Rn is the vector of unknowns
b ∈ Rm is the vector of observations
If m > n the system is inconsistent, so there is no exact solution.
An alternative in this situation is to find a vector x that comes the closest to
being a solution.
This special x will be called the least squares solution, which minimizes the
error, defined as the Euclidean distance between the predicted and actual values.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 59 / 87
 Least Square Approximation

Geometric Interpretation

Any m × n system Ax = b can be viewed as a vector equation

x 1 v1 + x 2 v 2 + · · · + x n vn

Vector b, the right-hand side of the

equation, is a linear combination of the columns vectors v1 and v2

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 60 / 87
 Least Square Approximation

Orthogonality

If b lies outside of the plane defined by v1 and v2 , there will be no solution. The
least squares solution x̄ makes the combination vector Ax̄ the one in the plane Ax
that is nearest to b in the sense of Euclidean distance.

Least squares is based on orthogonality.The shortest

distance from a point to a plane is carried by a line segment orthogonal to the
plane.
We want to find the point Ax̄(the projection of b) that lies closest to b in the
space spanned by the columns of matrix A. r = b − Ax̄(residual) is perpendicular
to the plane defined by the column of A.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 61 / 87
 Least Square Approximation

Normal equation

We now
(b − Ax̄)⊥{Ax|x ∈ Rn }
(Ax)T (b − Ax̄) = 0 for all x in Rn
xT AT (b − Ax̄) = 0 for all x in Rn
means that the n-dimensional vector AT (b − Ax̄) is perpendicular to every vector
x in Rn , including itself.
AT (b − Ax̄) = 0
AT Ax̄ = AT b
This system of equation is known as the normal eqautions. So, we should solve
this equation for the least squares solution x̄ that minimizes the Euclidean length
of the residual r = b − Ax

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 62 / 87
 Least Square Approximation

Error Measures

If the residual is the zero vector, then we have solved the original system Ax = b
exactly.
If not, the Euclidean length of the residual vector is a backward error measure of
how far x̄ is from being a solution. There are at least three ways to express the
size of the residual. The Euclidean length of a vector,2-norm
q
∥r∥2 = r12 + · · · + rm2

Squared error
SE = r12 + · · · + rm
2

Root mean squared error

p q
RM SE = SE/m = (r12 + · · · + rm
2 )/m

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 63 / 87
 Least Square Approximation

Fitting data by Least Squares

Given a set of m data points (t1 , y1 ), ..., (tm , ym ).

STEP 1. CHOOSE A MODEL Identify a parameterized model, such as

y = c1 + c2 t, which will be used to fit the data.

STEP 2. FORCE THE MODEL TO FIT THE DATA. Substitute the data
points into the model. Each data point creates an equation whose unknowns are
the parameters, such as c1 and c2 in the line model. This results in a system
Ax = b, where the unknown x represents the unknown parameters.

STEP 3. SOLVE THE NORMAL EQUATION The least squares solution for
the parameters will be found as the solution to the system of normal equations
AT Ax = AT b.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 64 / 87
 Least Square Approximation

Example

Find the best line for the four data points (1, 1), (0, 0), (1, 0), (2, 2)
follow three steps: (1) Choose the model y = c1 + c2 t
(2) Forcing the model to fit the data yields

c1 + c2 (−1) = 1
c1 + c2 (0) = 0
c1 + c2 (1) = 0
c1 + c2 (2) = −2

or, in matrix form,    

1 −1   1
 1 0  c 1
 0 

 1
 = 
1  c2  0 
1 2 −2

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 65 / 87
 Least Square Approximation

(3) The normal equations are

    
4 2 c1 −1
=
2 6 c2 −5

Solving for the coefficients c1 and c2 results in the best line

t y line error
-1 1 1.1 -0.1
y = c1 + c2 t = 0.2 − 0.9t. The residuals are 0 0 0.2 -0.2
1 0 -0.7 0.7
2 -2 -1.6 -0.4
The error statistics are squared error √ √
SE = (−.1)2 + (−.2)2 + (.7)2 + (−.4)2 = 0.7 and RMSE = .7/ 4 = 0.418

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 66 / 87
 Least Square Approximation

Best line y = 0.2 − 0.9t. RMSE is 0.418.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 67 / 87
 Least Square Approximation

Summary what is Least Squares Approxiamation

Least Squares Approximation is a method for finding the best-fitting curve (or
line) to a set of points by minimizing the sum of the squared differences between
the observed values and the values predicted by the curve. The idea is to
minimize the error between the actual data points and the predicted ones. if we

have a set of data (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) and to fit a line, curve, or any
function f (x), you want to find the function that minimizes the sum of squared
residuals (differences between actual and predicted values).
n
X
min (yi − f (xi ))2
i=1

ri = yi − f (xi )

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 68 / 87
 Least Square Approximation

Linear least squares

Definition: The method of linear least squares minimizes the sum of the squared
differences between observed values and the values predicted by a linear function.
Use Case: It’s widely used for fitting a straight line to data points. If your data
points are (xi , yi ), the goal is to find the line y = ax + b that best fits the data.
Equation: The error function (or residual) is
X 2
R(a, b) = (yi − (axi + b))
i

Solution: The optimal values of a and b minimize the sum of squares of these
residuals, and solving this involves matrix operations.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 69 / 87
 Least Square Approximation

Objective: Given a set of data points {(xi , yi )}ni=1 , you want to find the
best-fitting line y = ax + b that minimizes the sum of squared errors. The error is
defined as:
n
X 2
E(a, b) = (yi − (axi + b))
i=1

To minimize E(a, b), we take the partial derivatives of E with respect to a and b
and set them to zero:
n
∂E X
= −2 xi (yi − (axi + b)) = 0
∂a i=1
n
∂E X
= −2 (yi − (axi + b)) = 0
∂b i=1

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 70 / 87
 Least Square Approximation

These give us the normal equations:

n
X n
X n
X
a x2i + b xi = xi yi
i=1 i=1 i=1

n
X n
X
a xi + nb = yi
i=1 i=1

Solving these two equations yields the values of a and b.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 71 / 87
 Least Square Approximation

Constrained least squares

Definition: The constrained least squares (CLS) problem minimizes the sum of
squared errors, similar to the linear least squares, but with additional constraints
on the solution.
Use Case: It is useful when the solution must satisfy certain conditions, such as
equality constraints. Given a system of data points and constraints, CLS finds the
best solution that fits both the data and the constraints.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 72 / 87
 Least Square Approximation

Equation: The objective function for the constrained least squares problem is:

min ∥Ax − b∥22 subject to Cx = d

where:
A ∈ Rm×n is the data matrix,
x ∈ Rn is the solution vector,
b ∈ Rm is the observation vector,
C ∈ Rp×n is the constraint matrix,
d ∈ Rp is the vector of constraints.
Solution: The constrained least squares problem can be solved using the
**Karush-Kuhn-Tucker (KKT) conditions**. The optimality condition is given by:
 T
2A A C T x
   T 
2A b
=
C 0 z d
where z is the vector of Lagrange multipliers.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 73 / 87
 Least Square Approximation

Objective: The goal is to minimize the objective function ∥Ax − b∥22 while
ensuring that the solution satisfies the equality constraint Cx = d.
KKT System: The KKT system combines both the objective function and the
constraints into one system, and solving it yields the optimal values of x and the
Lagrange multipliers z:
   T −1  T 
x 2A A CT 2A b
=
z C 0 d

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 74 / 87
 Least Square Approximation

Nonlinear least squares

Definition: Nonlinear least squares is an optimization problem where the

objective is to minimize the sum of squared differences between observed values
and the values predicted by a nonlinear model.
Use Case: This method is used when the relationship between the independent
and dependent variables is nonlinear. It is commonly applied in fitting curves, such
as exponential, logarithmic, or power-law models.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 75 / 87
 Least Square Approximation

Equation: The objective function for nonlinear least squares is:

n
X 2
min (yi − f (xi , θ))
θ
i=1

where:
xi are the input data points,
yi are the observed data points,
f (xi , θ) is the nonlinear function parameterized by θ,
θ is the vector of parameters to be optimized.
Solution: Nonlinear least squares problems are typically solved using iterative
methods. One popular method is the **Gauss-Newton method**, which
approximates the nonlinear function by a linear one using a first-order Taylor
expansion:
f (xi , θ) ≈ f (xi , θk ) + Ji (θ − θk )
where Ji is the Jacobian matrix of partial derivatives of f (xi , θ) with respect to θ.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 76 / 87
 Least Square Approximation

Objective: The objective is to iteratively update the parameter vector θ by solving

a sequence of linear least squares problems. At each iteration k, the update is:

θk+1 = θk + ∆θ

where ∆θ is the solution to the linear least squares problem:

n
X 2
min (yi − f (xi , θk ) − Ji ∆θ)
∆θ
i=1

Jacobian Matrix: The Jacobian matrix J consists of the partial derivatives of the
model function with respect to the parameters:

∂f (xi , θ)
Jij =
∂θj

At each step, the linearized problem is solved, and the parameters are updated
until convergence.
Convergence: The iterative process continues until the parameter updates ∆θ
are sufficiently small, indicating that the solution has converged.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 77 / 87
 Least Square Approximation

The Gauss-Newton method

The Gauss-Newton method is a popular iterative method used to solve nonlinear

least squares problems. It approximates the Hessian matrix H by neglecting the
second-order terms and linearizing the problem.
Given a set of nonlinear equations r1 (θ), r2 (θ), . . . , rm (θ), the goal is to find θ
that minimizes the sum of squared residuals.
To minimize:
r1 (θ)2 + · · · + rm (θ)2
we follow the iterative process:

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 78 / 87
 Least Square Approximation

1 Initialize the parameter θ0 .

2 For each iteration k = 0, 1, 2, . . .:
1 Compute the Jacobian matrix J(θk ), where:

∂ri (θ)
Jij =
∂θj
2 Solve the normal equation:

AT A∆θ = AT r

where A = J(θk ) is the Jacobian at the current iteration.

3 Update the parameters:
θk+1 = θk + ∆θ
3 Repeat until convergence.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 79 / 87
 Least Square Approximation

Hessian
PmApproximation: In Newton’s method, the Hessian matrix
H = i=1 ∇2 ri involves second-order derivatives. In Gauss-Newton, we
approximate the Hessian by neglecting the second-order term:

H ≈ JT J

where J is the Jacobian matrix.

Update Rule: The update for θ is given by:

θk+1 = θk + v k

where v k is the update direction found by solving the linearized system.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 80 / 87
 Least Square Approximation

Exponential least squares

Definition Exponential least squares is a variant of the least squares method

where the model we are fitting to the data has an exponential form. The goal is
to fit a model of the form:

y = aebx
to a set of data points (xi , yi ), where a and b are parameters to be determined, x
is the independent variable, and y is the dependent variable. Objective In
exponential least squares, the objective is to minimize the sum of squared
residuals between the observed values yi and the values predicted by the
exponential model. The residual for each point is the difference between the
observed value and the predicted value, and the error function to minimize is:
n
X
E(a, b) = (yi − aebxi )2
i=1

where n is the number of data points

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 81 / 87
 Least Square Approximation

Solving To solve this nonlinear problem, we first linearize the exponential model
by taking the natural logarithm of both sides:

ln(y) = ln(aebx ) = ln(a) + bx

This transforms the model into a linear form:

ln(y) = ln(a) + bx
Letting Y = ln(y) and A = ln(a), the equation becomes:

Y = A + bx
This is now a linear model in terms of A and b, which we can solve using the
method of linear least squares. Once we find A and b, we can recover a by
exponentiating A:

a = eA

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 82 / 87
 Least Square Approximation

Steps
1 Transform the Data: Take the natural logarithm of the dependent variable
y to linearize the equation.
2 Apply Linear Least Squares: Solve the linearized equation Y = A + bx
using linear least squares techniques.
3 Recover Parameters: Once the least squares method gives estimates for A
and b, recover a as a = eA .
4 Evaluate the Fit: Use the original model y = aebx with the obtained
parameters to evaluate the fit to the data.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 83 / 87
 Least Square Approximation

Example Suppose we have a dataset {(x1 , y1 ), (x2 , y2 ), . . . , (xn , yn )}, and we

want to fit the model y = aebx . To do this:
1 Compute Yi = ln(yi ) for all i.
2 Solve the linear least squares problem for Yi = A + bxi , where A = ln(a).
3 Extract a = eA and b from the solution.
Use Case Exponential least squares is commonly used for modeling phenomena
that exhibit exponential growth or decay, such as population growth, radioactive
decay, or interest compounding in finance.
Error Function The error function for exponential least squares is:
n
X
E(a, b) = (yi − aebxi )2
i=1

Minimizing this error function requires nonlinear optimization techniques, as the

equation is nonlinear in terms of the parameters a and b. However, after
linearization, it becomes a linear least squares problem, which is easier to solve.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 84 / 87
 Importance of Approximation in Real-World Applications

Approximations are fundamental in many real-world scenarios where exact

solutions are either impossible or unnecessary due to complexity, noise, or
measurement errors. Here’s why approximations are crucial:
1 Imperfect Data:In real-world situations (like weather data, energy
consumption, or financial trends), data is often noisy or incomplete. Exact
modeling is impractical, so we approximate with simpler models that capture
the general trend.
2 Computational Feasibility: In complex systems, exact solutions can be
computationally expensive or impossible (e.g., modeling fluid dynamics,
signal processing). Approximations simplify models to make them solvable
within reasonable time and computational limits.
3 Decision-Making Approximate solutions enable quick decisions in uncertain
conditions. For instance, in machine learning or data science, we approximate
relationships between variables, enabling forecasting, classification, and other
tasks even with noisy data.
4 Optimization:Many optimization problems in operations research or
economics are approximations of real-world systems. Accurate enough
approximations are sufficient to make optimal decisions within given
constraints.
Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 85 / 87
 Importance of Approximation in Real-World Applications

Influence of Accuracy and Errors on Approximation

Methods

1 Accuracy:The quality of an approximation depends on how closely it

replicates the real-world data or underlying function. Higher accuracy
generally means a more complex model (e.g., higher-degree polynomials,
more Fourier components).
2 Overfitting and Underfitting : High accuracy on the training data may
lead to overfitting,a high-degree polynomial might fit every data point but
fail to generalize to unseen data. result in underfitting, where the model fails
to capture the essential features of the data.
3 Error: In approximation methods, residual errors help quantify the model’s
performance. Metrics like Mean Squared Error (MSE) or RMSE are
commonly used to measure error in least squares and other approximation
methods.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 86 / 87
 References

References

Timothy Sauer, Numerical Analysis, 2nd Edition, Pearson, 2011.

Richard L. Burden and J. Douglas Faires, Numerical Analysis, 9th Edition,
Brooks Cole, 2010.
https://www.geeksforgeeks.org.

Ani Okropiridze, Mariami Mamageishvili, Anano Tamarashvili Function and Data Approximation November 28, 2024 87 / 87