From Fast Multiplication to Compressed Image - Fourier Rules
                            CSA Showcase Illumination
Polynomials:       To determine a polynomial f(x) of
                                                          Choosing a Value Representation                                                      Speeding Up DFT:                                      Image Compression :
degree n, suffices to know its value for n+1 distinct
                                                             While we can choose any n+1 values to
values. Thus two points uniquely represents a line,                                                                                             FFT is a technique that reduces the complexity of    An image is a two dimensional signal. The signal
                                                             represent a polynomial in value representation,
three points represents a quadratic equation.                                                                                                  DFT computation from O(n2) to O(n.logn). The idea     varies in space , instead of time. Converting from
                                                             it is useful to compute f(x) at the n roots of unity.
                                                                                                                                               is to formulate DFT recursively                       spatial domain to frequency domain is advantages
                                                             ie f(1), f(ω), f(ω2)..... If f(x) is a0 + a1x + a2
Thus a polynomial         ( 6t- 5t2 + t3 ) can be                                                                                                                                                    because we can analyze frequencies better. DCT
                                                             x2........
represented uniquely in coefficient representation                                                                                                                                                   converts an array of pixel values into an equally
as [ 0 6 -5 1 ] or in value representation eg. values                                                                                          or                                                    sized array of frequency components that also
                                                                   f(1) = a0 + a1 + a2 ........
at t=0,1,2,3 are [ 0 2 0 0 ]                                                                                                                                                                         represents the image
                                                                   f(ω) = a0 + a1 ω + a2 ω2........
                                                                   f(ω2) = a0 + a1 ω2 + a2 ω4........

Why n+1 points uniquely determine the polynomial?         These can be represented in form of matrix.

   suppose f(x) and g(x) (of degree n) both agrees                                                                                             A Note on Fourier Transform
on n+1 points, then f(x) – g(x) is a polynomial with
degree n or less and has n+1 zeros. This happens
only if f(x)- g(x)=0 or f=g.                                                                                                                   A signal f(t) can be expressed as the sum of sum of
                                                                                                                                               cosines and sinusoids. In the Euclidean space. Any
This can be used to find the polynomial given its                                                                                              point can be given in terms of weighted sum of
values at n+1 points say f(0)=1, f(1)=4 , f(2)=13,        then [ f(1), f(ω), f(ω2).. ] has same information as                                 basis. Likewise any function f(t) can be considered
f(3)=34, f(4)=73                                          [ a0 , a1 , a2 ]. This representation is called the                                  as a point in the vector space of functions whose
                                                          Discrete Fourier Transform                                                           basis are the cosines and sinusoids.
                                                                                                                                                                                                     The human eye, cannot detect high frequency
using finite differencing we can find                                                                                                                                                                components. Using the process of quantization,
the polynomial. Babbage used it in                                                                                                                                                                   we retain all low frequency components. (The high
his Difference Engine                                     Polynomial Multiplication using DFT                                                                                                        frequency components are retained only if they
                                                          Suppose we have to multiply f(x) = x2 + x + 2 and                                                                                          have large coefficients).      On the resulting
                                                          g(x)= x+3 . then DFT of g(x) ( whose coefficient                                                                                           coefficients run length encoding and huffman
Polynomial Multiplication                                 representation is [3 1 0 0]) is given by                                                                                                   compression is applied
                                                                                                                                               The advantages of converting it from time domain
If f(x) and g(x) has to be multiplied using coefficient                                                                                        representation to frequency domain representation
representation, the kth coefficient          would be                                                                                          is that many manipulations (like filtering out high
calculated as Σ ai*bk-i where i varies from 0 to k.                                                                                            frequencies) can be easily done in frequency
Thus we get (a0b0, a0b1+a1b0, a0b2+a1b1+a2b0....) .                                                                                            domain. FT handles continuous signals. DFT can
This is also know as vector convolution of ai and bi
                                                                                                                                                                                                     Applications
                                                                                                                                               be viewed as forming a continuous function of ω
                                                          DFT of f(x) is [4, i+3, 2 , -i+3] (calculation not shown)                            from discrete values
A better method is to use value representation.           DFT of g(x) is [4, i+1, 2 , 1-i ] (calculated above)                                                                                       SEQUENCE RETRIEVAL: We have a database of
Suppose f(x) is of degree 10 and g(x) is of degree        p(x).q(x)= [16, 4i+2,4, 2-4i] (multiplying pointwise)                                                                                      sequences and a query sequence. Each sequence
20, then we can compute f(x) and g(x) at 31 points                                                                                                                                                   is represented by its first few coefficients from
and multiply them. since f(a)*g(a) = (f*g) (a) , we                                                                                            The Real Matrix:                                      DFT. The query sequence is also transformed
would get the value representation of f*g from which                                                                                                                                                 using DFT and its first few coefficients are
                                                                                                                                               since ? is a complex root of unity                    compared.
we can get the coefficient representation.
                                                                                                                                                                                                     OTHER        APPLICATIONS         include   filtering,
                                                                                                                                                                                                     convolution, audio processing, medical imaging,
                                                                                                                                               If the function is even, the the sin component is     pattern recognition, partial differential equations,
                                                                                                                                               zero and then the entries of the DFT matrix are       multiplication of large numbers, analysis of time
                                                          computing set in 2, c3 similarly we get, equivalent, italic, 18 to 24 points,
                                                           Captions to be c1,c Times or Times New Roman or                                     real. Such a transformation is called Discrete
                                                           to the length of the column in case a figure takes more than 2/3 of column width.                                                         series, solving system of linear equations
                                                                                                                                               Cosine Transformation.

                                                                                                                                                                                                                                 Presented by D.Arvind

Fourier Transforms

  • 1.
    From Fast Multiplicationto Compressed Image - Fourier Rules CSA Showcase Illumination Polynomials: To determine a polynomial f(x) of Choosing a Value Representation Speeding Up DFT: Image Compression : degree n, suffices to know its value for n+1 distinct While we can choose any n+1 values to values. Thus two points uniquely represents a line, FFT is a technique that reduces the complexity of An image is a two dimensional signal. The signal represent a polynomial in value representation, three points represents a quadratic equation. DFT computation from O(n2) to O(n.logn). The idea varies in space , instead of time. Converting from it is useful to compute f(x) at the n roots of unity. is to formulate DFT recursively spatial domain to frequency domain is advantages ie f(1), f(ω), f(ω2)..... If f(x) is a0 + a1x + a2 Thus a polynomial ( 6t- 5t2 + t3 ) can be because we can analyze frequencies better. DCT x2........ represented uniquely in coefficient representation converts an array of pixel values into an equally as [ 0 6 -5 1 ] or in value representation eg. values or sized array of frequency components that also f(1) = a0 + a1 + a2 ........ at t=0,1,2,3 are [ 0 2 0 0 ] represents the image f(ω) = a0 + a1 ω + a2 ω2........ f(ω2) = a0 + a1 ω2 + a2 ω4........ Why n+1 points uniquely determine the polynomial? These can be represented in form of matrix. suppose f(x) and g(x) (of degree n) both agrees A Note on Fourier Transform on n+1 points, then f(x) – g(x) is a polynomial with degree n or less and has n+1 zeros. This happens only if f(x)- g(x)=0 or f=g. A signal f(t) can be expressed as the sum of sum of cosines and sinusoids. In the Euclidean space. Any This can be used to find the polynomial given its point can be given in terms of weighted sum of values at n+1 points say f(0)=1, f(1)=4 , f(2)=13, then [ f(1), f(ω), f(ω2).. ] has same information as basis. Likewise any function f(t) can be considered f(3)=34, f(4)=73 [ a0 , a1 , a2 ]. This representation is called the as a point in the vector space of functions whose Discrete Fourier Transform basis are the cosines and sinusoids. The human eye, cannot detect high frequency using finite differencing we can find components. Using the process of quantization, the polynomial. Babbage used it in we retain all low frequency components. (The high his Difference Engine Polynomial Multiplication using DFT frequency components are retained only if they Suppose we have to multiply f(x) = x2 + x + 2 and have large coefficients). On the resulting g(x)= x+3 . then DFT of g(x) ( whose coefficient coefficients run length encoding and huffman Polynomial Multiplication representation is [3 1 0 0]) is given by compression is applied The advantages of converting it from time domain If f(x) and g(x) has to be multiplied using coefficient representation to frequency domain representation representation, the kth coefficient would be is that many manipulations (like filtering out high calculated as Σ ai*bk-i where i varies from 0 to k. frequencies) can be easily done in frequency Thus we get (a0b0, a0b1+a1b0, a0b2+a1b1+a2b0....) . domain. FT handles continuous signals. DFT can This is also know as vector convolution of ai and bi Applications be viewed as forming a continuous function of ω DFT of f(x) is [4, i+3, 2 , -i+3] (calculation not shown) from discrete values A better method is to use value representation. DFT of g(x) is [4, i+1, 2 , 1-i ] (calculated above) SEQUENCE RETRIEVAL: We have a database of Suppose f(x) is of degree 10 and g(x) is of degree p(x).q(x)= [16, 4i+2,4, 2-4i] (multiplying pointwise) sequences and a query sequence. Each sequence 20, then we can compute f(x) and g(x) at 31 points is represented by its first few coefficients from and multiply them. since f(a)*g(a) = (f*g) (a) , we The Real Matrix: DFT. The query sequence is also transformed would get the value representation of f*g from which using DFT and its first few coefficients are since ? is a complex root of unity compared. we can get the coefficient representation. OTHER APPLICATIONS include filtering, convolution, audio processing, medical imaging, If the function is even, the the sin component is pattern recognition, partial differential equations, zero and then the entries of the DFT matrix are multiplication of large numbers, analysis of time computing set in 2, c3 similarly we get, equivalent, italic, 18 to 24 points, Captions to be c1,c Times or Times New Roman or real. Such a transformation is called Discrete to the length of the column in case a figure takes more than 2/3 of column width. series, solving system of linear equations Cosine Transformation. Presented by D.Arvind