Fourier Analysis Using A Spreadsheet

Robert A. Dory, and Jeffrey H. Harris

Citation: Computers in Physics 2, 83 (1988); doi: 10.1063/1.4822799

View online: https://doi.org/10.1063/1.4822799
View Table of Contents: https://aip.scitation.org/toc/cip/2/6
Published by the American Institute of Physics

ARTICLES YOU MAY BE INTERESTED IN

Spreadsheets for physics

Computers in Physics 2, 70 (1988); https://doi.org/10.1063/1.168300

Performing Fourier transforms on extremely long data streams

Computers in Physics 3, 59 (1989); https://doi.org/10.1063/1.168338

Solving coupled, nonlinear differential equations with commercial spreadsheets

Computers in Physics 3, 75 (1989); https://doi.org/10.1063/1.168340

Spreadsheets Tackle Physics ProblemsUsing Lotus 1-2-3 and software from the AAPT bulletin board,
physicists can quickly solve and graph a multitude of problems
Computers in Physics 2, 37 (1988); https://doi.org/10.1063/1.4822723

Feynman’s proof of the Maxwell equations

American Journal of Physics 58, 209 (1990); https://doi.org/10.1119/1.16188

On Feynman’s proof of the Maxwell equations

American Journal of Physics 60, 301 (1992); https://doi.org/10.1119/1.16921
SPREADSHEETS
Fourier Analysis Using ASpreadsheet
Robert A. Dory and Jeffrey H. Harris
or many tasks involving Fourier analysis, a spread-
F sheet is an excellent tool. Using three examples, this
article shows that for data sets involving perhaps a
hundred points and a few tens of coefficients, a
spreadsheet method is practical and quick on personal
computers such as the Macintosh SE or the IBM PC AT
and compatibles. With enhanced hardware such as
floating-point co-processor chips, the numbers may be
extended by about an order of magnitude. This article
presents three spreadsheet templates that should permit
rather easy generalization to other orthogonal basis
functions, provided those can be defined in terms of
standard functions or recurrence formulas.
Fourier analysis is used here to illustrate provision of
modularity to spreadsheet element definitions, so that
the available replication processes allow simple exten-
sion of the ranges of the calculation; in this case, the
number of harmonics and integration nodes.
Fourier analysis of a real function also illustrates
some techniques in spreadsheet use. These include
selective application of array Ivector arithmetic and the
important trick of formulating formula blocks that can be
replicated (blocks copied with relative references adjust-
ed appropriately and absolute references unchanged) to
extend the range or resolution of the calculation.
For a moderate problem, programming and debug-
ging of functional Fourier analysis templates like the
Fourier templates discussed below might take a spread-
sheet user less time than finding the user's manual and
the object code for an existing Fourier analysis program.
The reader is assumed to have some experience with
Excel or a similar spreadsheet program so the general
operations given here can be carried out without detailed
instructions. The examples use the RIC2 option for
addressing the elements of the template; this is a matter
of taste, but has the property that variable names can take
the form aO and bt, which otherwise conflict with the
slightly more compact normal addressing scheme AI,
AB137, 00..
Two templates for the Excel spreadsheet program
illustrate the trade-off between extensive use of memory
for possible reduction of calculation time and, converse-
ly, the deliberate recalculation of data to reduce memory
requirements.
Fast Fourier transform (FFT) techniques are not used
here because the logic required is not straightforwardly
well suited to the simpler tools available in the spread-
sheet. By development of function macros for retrograde
indexing, formula entry and logic, the FFT or the related
Robert A. Dory and Jeffrey. H. Harris are Research Physicists at Oak
Ridge National Laboratory, Oak Ridge, Tenn.
fast Hartley transform (FHT) should be practical.
The latter could give an interesting challenge to
readers; an article on the FHT in the April, 1988 issue of
Byte magazine, p. 93, by Mark O'Neill would be a fine
starting place for development of this process, which
uses only real (non-complex) arithmetic.
The procedure used here is straightforward integra-
tion with the formulas:
.fi,x) = "i.fn cos nx + "i.gn sin nx, with n = 0,1,2'00'
fn = J .fi,x)cos nx dxlf cos" nx dx, from 0 to XM = 21T,
gn = J J\x)sin nx dx/ f sin" nx dx, from 0 to XM = 21T,
where XM = 21T is the periodicity interval of .fi,x). With
this approach, the basis functions of the expansion could
be readily changed for application of another basis, such
as Legendre or Bessel functions, especially if the higher
index elements can be calculated with sufficient accuracy
using two-term recurrence formulas as in the Fourier
case. Accuracy of this is discussed further under
Example 2.
The integrations use the trapezoidal rule for non-
uniform (arbitrary) grid spacing:
fg(X)dx = 1/2 "i. (xj + 1 - Xj _ 1 ) g(xj )
where the nodes are Xj' for j = 0, jM and for the special
terms j = 0 and jM' the corresponding X_ 1 and XjM + 1
are to be taken as Xo and XjM' If a new set of basis func-
tions called for a weight function, it could be grouped
conveniently with the factor (Xj + 1 - Xj _ 1 ).
A test function is provided. It is given by y" with pa-
rameters a" b o' b, and "noise," where
y, = a , sin,x + b, + b 2 cos 2x + noise' (R - 1/2)
and R is the pseudo-random function with uniform
distribution in [0,1).
Fourier Template 1
Sine and cosine values are tabulated at the integra-
tion nodes and recurrence formulas are used to obtain
the higher harmonic functions sin nx j and cos nx j. The
accuracy should be adequate for reasonable values of nx j
since Excel arithmetic is done to roughly 15 decimal
places. For small values of nand iM , for example, ten and
50, Excel recognizes that the basis data do not need to be
recalculated when changes are made only to test data
f(xi ). The process is then reasonably speedy. However,
when the number of nodes increases by a factor of three
or four, tables of dependency become large and Excel
may then not be able to avoid some recalculation.
Figure 1 shows one arrangement. We assume that a
limited number of harmonics are needed, but that
COMPUTERS IN PHYSICS. NOV/DEC 1988 83
SPREADSHEOS

Fourier [Hample 1.

21 ooints 0 2 dxO 0 .31 4 bO -0 .25 a1 12 .34 b2 0 .123

1 x dx u 51 co 5i 2 co 2 513 co 3 514 co 4
0 0 .0 00 0 .3 14 -0 .127 0 .000 1.000 0 .000 1.000 0 .000 1.000 0 .000 1.000
1 0 .314 0 .628 3 .663 0 .309 0 .95 1 0 .588 0 .809 0 .809 0 .588 0 .951 0 .309
2 0 .628 0.628 7.041 0 .588 0 .809 0 .951 0 .309 0 .951 -0 .309 0 .588 -0 .809
3 0 .942 0 .628 9 .695 0 .809 0 .588 0 .951 -0 .309 0 .309 -0 .951 -0 .588 - 0 .809
19 5 .969 0 .628 -3 .964 -0 .309 0 .95 1 -0 .588 0 .809 -0 .809 0 .588 -0 .951 0.309
20 6 .283 0 .314 - 0 . 12 7 0 .000 1.000 0 .000 1.000 0 .000 1.000 0 .000 1.000
(011) Non uni for m int~rvals OK. us es Q'Jadratic forms to do int~Q r ations
Co~ffic 1~nt5 0 I 1 1 I 2 2 3 3 I 4 4 I
1. This Is the Fourier analysis template 1. C051n~ fn -0 .250 I 0 .000 1 0 .123 0.0001 0 .000 1
A two-column block from "SI n Co n " Si n~ en 0 .000 11 2.3 4 0 10 .0 0 0 0.000 10.000
nois e 0 .0 I Test function : y = bO + al sin x + b2 cos 2x + noise * (randO- .5)
through " Coe ffici e n ts Cosine and Sine"
2011 011 011 011 011 011 011 011 011
can be copied to blank columns on the
factor 1.00 I norm 12 .566 6.283 6 .28 3 6 .283 6 .283 6 .283 6.283 6 .283 6 .283
right to add calculation for another n
value.

acceptable accuracy may require a substantial number of The "norm" elements are entered in the form
integration nodes ,x-j ' Integration nodes are arranged "I = surntsi' si' dxl j"using the command key and a
downward, with harmonic numbers across, because that carriage return, to make them array function formulas,
fits the screen format best. If the data to be analyzed are which can act, for example, to give the inner product of
gen erate d by another program, Fortran for example, they the elements of the vectors "sin " and "dx" [that is,
can b e stored one number per row, so that they can ~ (sin ) j . (sin), . tdxl] . Function formulas are set off by
con ven ie n tly be pasted into Excel using the Macintosh braces: I···J.
Clipboard. The rows labeled " Cosin e Coefficients" and "Sin e
At the bottom of the column blocks labeled "si n " Coefficients" are filled with similar array function
and "co n " are formulas for the integrations. The formulas having the forms:
numbers labeled "n orm " are the values of j'sin" nx dx and "! = SUM(y' R[ - 23]C:R[ - 3]C' dxl/B] + 3]Cj " and
jcos'' nx dX" multiplied by an overall normalizing number
" facto r." For the present sine and cosine basis and ,, ' = SUM(y' R[ - 24]C:R[ - 4]C' dX"l/R[ + 2]Cj"
assumed integration interval [O,p1T], these values are all where the divisor picks up the value of "n o rm" from the
p1T exce p t those for n = 0, which are 0 and 2p1T. To save same colu m n and the symbol "R[ - 25]C:R[ - 5]C " calls
time, these could be changed to con stan t values 1, 0, and 2 out the basis function vector from the same column.
and " fa cto r " to p1T; they are as indicated to provide for If more integration nodes are needed, one can in sert
use with another set of orthogonal functions. the appropriate number of blank rows somewhere after

Fourier [Hamp le I. Formu las:l

2 3 4 5 6 17
2 =COUNT(n) iooi nt s Ip i2 dx O 1 ~ p *P IO /(po int s - 1)
3 1 x dx III 51 !C O

2. A partial listing of the 'program' for

template 1. As shown, columns 5 and 6, 25 j(p11) 1 jNon- uni for m intervals 10K
are too narrow to give complete formulas
In all rows. The notation "RC:R[2]C(5]"
denotes a 3 X 6 range of cells, from the
cell containing the reference to the cell
five columns to the right and two rows
~~ no;s~ :O ! .
I}~~ t.~.~~.!.j~.~ : ~ ..:::..~Q.lp;~.! ~!.~i~; ~~ ~~.~ ~~ ~ ~~.i . . . ..
down. 31 f act or !1 Inor m :=SUM( dx) 1=SUM( R[ =SUM( R[- 27 ]C:R[-

Fourier [Hample I. Formula s:2

7 8 9
2 =p*p 10I(points-l ) bO 1-0 .25
3 co 2 ,2
.....L =COS(x) =si*RC[-I]+co*RC[-2)
;;;s·i"*·RcF·i · ·i+~·~·*·R·C"F2f· ....··..........·......·....·....·....··..·=co*RC[-2 )-si*RC[-3]
..·......···....·..·....···..·..·
.z., ;;;COS(~)"·"""""·""·"·" ;;·s;·*R d :·l i"~~~ *R cF2T ......·........··..·....·..........·....·..·
;;;~·~·*·RC"F2Fsi*RC ·[:3·j"
......................·......·..
6 '~6)s'(~')""" """""""'''' ·;;·~~· * RC "[:·2· j":S·;·*·R·Ci": 3 f

24 =COS(x ) =5i*RC[-l )+co*RC[-2] =co *RC [-2]-si*RC[-3)

3 . A s e c o n d partial listing of the 25 i uses Quadrati c forms to do int~Qrations do int~Qrations
26 1 I=R [-23 ]C I=R[ -23]C
'program' for template 1. For a nicely
27 = SUH ( u *R[-23)1 , = SUH ( Il.R [-2 3 )C :R [ - 3 )C. dx ) 1
labeled display, row 3, columns 8 and 9,
28 I= SUM ( '1 *R[-24 )C :R [ - 4 I C. dx ) / nor-m'
which contain the values of n , have been
x + b2 cos 2x + noi !se * (randO- .5)
assigned number formats " Si "0 and " Co ~
30
1011·· ··.. ..·....·..·..·..·....·....'(011..·..·..·....·..........·....·........·......................................·....·....·....:·011·....·....·....·........·....·....·....·........................................
I

" 0 which preface the n value with the text =SUM(R[ -27)C :RH - SUM(R[- 27 )C:R[- 7 )C" 2 * dx) ' -SUM (R(-27)C :R[- 7 )C"2 * dx)
31
in quotes.

84 CO MPlJTE IIS IN PHYSICS. NOVIO EC 198 H

 Fo uri er EHamp le 2.

21 ooints D 2 (dx ) 0 .314 I ~ 1 1 I bO -0 .25 b2 0 .1231

i x dx II Test function : \I ~ bO + al sin x + b2 cos 2x + noise (randO- .5) I'

0 0 .000 0 .31 4 - 0 . 12 7 I noise 0 .0

1 0.31 4 0 .628 0 . 159 Copfficipnts n I 0 1 2 3 4
2 0 .628 0 .628 0 .376 Co sinp fn -0 .250 0 .000 0 .123 0 .000 0 .000
3 0 .942 0 .628 0 .521 Sinp on 0 .000 1 .000 0 .000 0 .000 0 .000
4 1.25 7 0 .628 0 .602 norm 6 .283 I
5 1.57 1 0 .628 0 .627 4. Fourier analysis template 2 is a space-
19 5 .969 0 .628 -0 .460 saving program that uses half the
20 6 .283 0 .3 14 -0 .127 Usu quadratic forms to do int egr ations Non- equidist ant pts OK memory of the previous example; it also
(p'l1) has half the speed, but the added time
may be unimportant.

the second row of nodes and before the last one. Then the Macintosh II and use of a floating point co-processor chip
relative duplication command ("Fill Down") can be used the calculation time for about a thousand elements and a
to populate the template from the second node down half dozen coefficients is roughly a minute.
through the last. This should provide the required higher The approach here is generally the same as in the
integration accuracy when new data Wi) are added. first example, except that the coefficients are entered as
If another harmonic value is required, the user can the array function formulas
copy the two-column rectangle from "si n co n" down
through the "n orm " line into the two blank columns at
the right of the presently filled zone. To make this
"! = SUMly' SIN (n*X))/norm I" and
possible, the sine and cosine formulas were written in "I =SUMly 'COS(n*X))/normj " where the
mixed form .. = si ' RC[ - 1] + co RC[ - 2]," which
means multiply the element on this row of the vector row array n is defined as labeled, and norm is set to p1T for
named "si" by the element of this row and one column the Fourier transform to eliminate some repetition of the
back, and add the corresponding element of vector "co " sine and cosine calculations . An easily misleading
times the element two columns back. These formulas are complication arose in this example. An initial assumption
not array function formulas because the intent is to proved incorrect: that the proper element of a row array
eva lu ate a column vector row by row using data from the norm, as in Example 1, would be used by the array
same row of other column vectors. The vectors "si" and formula definingfn and Sn- Instead, the first element of
" co " were defined to be columns C6 and C7, elements R4 the norm array was used even for succeeding columns.
through R24 in the present case, and should not be This feature could be called a bug in Excel (version 1.04),
confused with functions SIN(argument) and COS(argu- but might be specified somewhere in fine print in the
ment). manual. The work-around chosen here was to reduce
norm to a scalar and then insert the necessary factor 2 in
Fourier Template 2
the formula for fa.
To reduce memory usage (by about half) in the case Integration nodes could be added to this template in
of large data sets f (;JCj), the basic functions in this example the same manner as in the first example. Providing more
are not determined once and stored. Instead, they are coefficients is also about the same: insert new columns
recalculated for each coefficient and case. Calculation after the n = 1 calculation and replicate the n,fn, gn and
time is about doubled, but still not large; with a norm columns to the right, as far as the last one.

Fourier EH ample 2. Formula s:!

2 I 3 4 I S 6 I 7
2 ~ COUNT( ;) ' Do;nt s !2 (d x) , ~ n * P IO /(no ints - l )
3 x dx u
.!.!~Lf:-!·n.~·!·! ~·n.":·I..··\I..=··~Q··:':..~·~..· ~·!~..~..·
~ R[1 )C[-I I-RC[-I] =bO+al *SI N(x) +
~ Q................. . . ;;;RFi
0
. . i·c·~·R2C'7· ;;;R'i'j"jc(:'i"i=RF,']Cr:Tj' ;;;b'o~'~'l';';'sir~(~):;:'
~ .!...................... .. £~!..f.f!~~!~t~.! .............................................
';;'R'r:Tjc:;:R2C7' ';;bo"~"~'l";s'iN(';;-j'~' ~.~.~.j.~.~...............

t.~ffl~l-tf!~:~;:~·rl~lfff
~ :?....__........ ..
l..._......._....._....._................
tz, .~_.._............... ';;;;RFi·iE~· R·2'c·i
;'Ri': 'j" jC-';:R2C7'
;;;b·ii·~·~·i ;,;·s·i~i(~f.;:·. Si neo
·;;·bo:;::;·i"*s·ii·j·(·~·j·~·
j
,
8 4
5. A partial listing of the 'program' for
1
2 4 20
2S
=R[-1 lC+R2C7 =RC[- 1)- R[- 1lC[-1
!( 0'l1) , !
=bO+al *S IN( x ) +
,.l!.~~.~..9~~.~~.C~.!.t~·i·f.~r.·T.~..!.~..~~..!.n..\~.gr..~. template 2. Columns 5 and 7 are not
shown fully.

Fourie r EHam ple 2. Formu l as:2

8 I 9 I 10 I 11 I 12 T
2 ~I il bO ,- 0 .25 b2 ,0. 12<;
,
~ ··:!:"~:?·i""·""·"""·"""··""·"·"·""""·""·""·"":·"··""·"""..·..··..·..··.._··..·..........·....·..·..·..·..··..·....··..·..··..........··..1....··..··..·....·......··tnoi.e
4 ,0
S n !O 01 ;2 ,3 ,4 6. A second partial listing of the
6 fn ,= SUH ( u + d x ) / n or m /2 ' -SUH ( u *C OS( R[ - I )C +x) +dx)/norm ' - SUH(Il + ' = SUH(u ~~S U H 'program' for example 2. Columns 8, and
7 Qn 10 I~ S U H ( u +S IH( R[ - 2 )C+ x )* dx) / nor m ,- SUH ( u + j =S UH ( u +I-SUH 11 through 13 are not shown fully. See
8
9
norm ,~o *PIO
, ~ .. ........ .. ......... . .... .. . . . .... .... . . . . .. . . . . . . .... . \.... . . . ... . . . . j.. . . . . .........t·.. . ....... the note in the text concerning the
change in use of norm from Example 1.

C.O MPIlT ERS IN PHYSICS, NOV/I)EC 19M 85

SPREADSHEOS

Fouri er [ Hampl e 3.

Test f unction : y = bO + al sin x +

+ b2 cos 2x + noise*(randO- .5) Act on's Comoaction
21 ooints p 2 dj n 0 1 2 3 I
.11 1 bO - 0 .2 5 0.05 Iphin 0 0 .31 42 0 .6283 0 .94 25
b2 0 .12 3 no i se' 0 .0 dphi c phin2 2 .00 1.9 0 1.62 1.18
(il.x) 0 .3 142 km ax 20 0 .3 1 cK p2 0 0 0 0
i x u uf1 k cK p l 0 0 0 0
0 0 .0 0 0 -0 .12 7 - 0. 127 20 c20 - 0 .12 7 -0 .127 -0 .127 - 0 .127
1 0 .3 14 0 .1585 - 0 .46 19 c19 - 0 .7 14 - 0. 70 1 - 0 .66 5 - 0 .609
? n "'7l'l n :>;7Sl'l _n" '0 ~I " _7 1 _ ? nn.:. - I 7 4 <1 _ I :>;l'l l'l

19 5 .969 - 0 .46 0 .158 5 1 cl - 114 .4 32 .361 l E-1 5 - 2E- 16

20 6 .283 - 0 .127 -0 .127 0 cO - 119 .5 30 .65 1.103 -0 .127
( pn) y (f1ipped) csum -5 .12 7 - 0 .127 1.10 3 - 0 .127
Nota bene : Equal int er v al s r equired. dsum 0 10 6E-16 -2E-16
7. Fourier template 3 is a compact Ref : Numerical Met hods that lI'o r k I
fn - 0 .2 5 2 E-15 0 . 123 0 I
analysis as described by F. S. Acton in the F. S. Acto n; Har per , Row ; 19 70 . err 9E- 16 2E- 17
book Numerical Methods That Work. The gnI 0 1 ('E - 17 0 I
method can accept an arbitrary number err -7E-16
of equally space data points.

Fourier [ Hamp le 3 Form ulas : !

4 I s 8 9 I 10 I 11
4 ! Acton's Cornoact ion
:; 10 i2 .~ L n '0 i - RC[- 1)+1
s bO i- 0.25 -I Ikma x loh i n ,-R[- l )C* dohi i - R[- l ]C* dphi
7 no i se1 0 .~ p..h.i c phi n2 -2*COS(phin) i - 2 * COS(phin)
8 kma x! - MAX(") =2 * oi * di cKn2 '0 10

8. A partial listing of the 'program' for

template 3. Notice that a single sine and
cosine pair is evaluated for each
harmonic coefficient. In the second
column shown, the data vector y is
'flipped' to simplify the downward
recurrence for the harmonic
coefficients.

Discussion cos nt/J = (2 cosel-cosrn + Ilt/J - cos(n + 2)t/J

In spreadsheet programming, higher efficiency is and
possible with more careful planning and improved
mathematical technique, The Fourier analysis process
sin nt/J = (2 c0stPl·sin(n + 1)t/J - sin(n + 2)t/J
has been the focus of a great deal of optimization in the with the observation that:
last couple of decades, The FFT process earlier permits K

vel)' rapid analysis, at the expense of substantially more '2.Jikt/Jn ) cos kt/Jn = [co - CI costPn ]
k~o
involved logic, restriction to uniform data spacing and to
a number of data points that can be decomposed into a and
K
product of powers of small prime numbers [z'" ·3" ...].
"Lfikt/Jn) sinkt/Jn = C 1 sint/Jn,
A compact and elegant process arising from a k=o
suggestion in F. S. Acton's book Numerical Methods That
Work, Harper and Row, 1970, provides an intermediate
when ck = (2 cost/Jn ),ck + 1 - ck + 2 + fl.kt/Jn), for
k=K, ...,I,O, with CK+l =CK+2 =0.
level of speed and complexity. It applies to uniformly
spaced data with arbitral)' number of points, and gives
an attractively simple template requiring only one sine
FOR FURTHER READING
and cosine per harmonic. The scaling of calculation time Numerical Methods That Work, mentioned above,
is asjM times nmex rather thanjM InjM for the FFT gives a fine discussion of the merits of using recurrence
process; it can be competitive then when only a few tens formulas to extend tables of standard functions. A more
of data points are needed and the required highest recent book is Numerical Recipes by Press, Flannel)',
harmonic, nmex, is not large. The speed is about twice Teolkowsky and Vetterling, Cambridge Press, 1986.
that of Example I, and four times that of Example 2.; Readers' suggestions for useful or especially effective
storage requirements are slightly (10%) under those for texts on numerical/mathematical methods would be
Example 1, twice that of Example 2. appreciated. They can be sent in care of Computers In
The process is based on recurrence formulas for Physics, American Institute of Physics, 335 East 45th
cosines and sines: Street, New York, N.Y., 10017. •

86 COMPUTERS IN PHYSICS, NOV/DEC 1988