TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5420 ~ 54
2
9
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.527
0
5420
Re
cei
v
ed
De
cem
ber 1, 20
13; Re
vised
F
ebruary 23,
2014; Accept
ed March 1
5
, 2014
Study of Additive Dither on Restraining Signal
Truncation Error
Tao Liu*, Shulin Tian, Zhigang Wa
ng, Lianping Gu
o
Schoo
l of Auto
mation En
gi
ne
erin
g, Unives
ity of
Electro
n
ic
Scienc
e an
d T
e
chn
o
lo
g
y
of C
h
in
a
Xi
yu
a
n
Ave, W
e
st Hi-T
e
ch Z
one, Ch
e
n
g
du, P.R.Chin
a, 86-
28-6
183
13
18
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: liutao
p
rivate
@gmai
l
.com
A
b
st
r
a
ct
Ow
ing to th
e l
i
m
ite
d
c
a
lcu
l
ati
on
precis
io
n d
u
r
ing
di
gital
si
gn
al
process
i
n
g
, the
inter
m
e
d
i
a
te stag
es
’
sign
al-b
it-w
idth truncati
on s
h
o
u
ld
be
exec
ute
d
to re
ali
z
e
th
e
conv
ersio
n
fro
m
hig
h
pr
ecisi
o
n to l
o
w
on
e. A
s
meth
od
of d
i
re
ct truncatio
n w
ill
deg
en
erate t
he S
puri
ous Fr
eeDy
n
a
m
ic
Ra
nge (SFD
R) p
e
rformanc
e of
th
e
output si
gn
al, this pa
per pr
o
pose
d
that a
d
d
itio
nal
dig
i
tal
dither sh
ou
ld
be a
dde
d b
e
f
ore op
eratio
n
of
truncatio
n, w
h
i
c
h cou
l
d
decl
i
ne the
har
mon
i
c distorti
on
efficiently
an
d ex
tend th
e dyn
a
m
ic r
a
n
ge of t
h
e
truncated
sig
n
a
l si
gn
ificantly
. T
w
o simu
lat
i
ons f
o
r tr
unc
ation
op
erati
o
n tow
a
rds s
i
g
nal w
i
th
ad
diti
ve
Gaussia
n
d
i
th
er a
nd
un
ifor
m
dith
er ar
e
carried
o
u
t to
prov
e the
va
lidity
of the
p
r
opos
ed
meth
od.
Co
mp
arative st
udi
es de
mo
nstrate t
hat the propos
ed al
gor
ithm a
p
p
lie
d in
Gaussia
n
dithe
r
ing a
nd un
ifor
m
dither
ing co
ul
d improve th
e ou
tput SF
DR perfo
rmanc
e by ab
out 16dB
and
1
5dB resp
ective
ly.
Ke
y
w
ords
: dig
i
tal dith
er, truncation err
o
r, precisio
n conv
er
sion, SF
DR
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
With the
intro
ductio
n
of
ad
vance
d
Fiel
d
Pr
ogramma
bl
e Gate
Array
(FPGA) archit
ectures
whi
c
h provid
e built-in Di
gi
tal Signal Proce
s
sing (DS
P
) sup
port
such a
s
em
be
dded multipli
ers,
block RAM
s
DSP blocks a
nd so o
n
, a n
e
w ha
rd
wa
re alternative
is available
for DSP
desi
gne
rs
who
ca
n get
even hig
her l
e
vel of
pe
rformance tha
n
those a
c
hiev
a
b
le on
gen
eral pu
rpo
s
e
DSP
proc
es
sors
[1]. In order to
redu
ce
the
cost of
ha
rd
wa
re
while
in
cr
e
a
sin
g
t
h
r
oug
h
put
r
a
t
e
s,
m
o
st
digital FPGA
implem
entat
ions of
sign
a
l
algo
rith
m
s
rely solely o
n
fixed-p
o
int
app
roximati
ons,
su
ch
as fre
q
uen
cy mixing
, sign
al d
e
ci
mation,
a
nd f
iltering
and
so on.
The
in
evitable p
r
obl
em
cau
s
e
d
by fix-point calcula
t
ion is the growth in
bit wi
dth. Especi
a
ll
y after multistage calculati
on,
the increme
n
t of bit width of final output
data, co
mp
a
r
ed to that of
origin
al input
data, turns
o
u
t
to be very consi
derable.
Whe
n
the lat
t
er stag
es’
al
lowa
ble p
r
o
c
essing bit
wi
dth tend
s to be
sho
r
t, we
ne
ed to tra
n
sfo
r
m the p
r
evio
us
stage
s’
ou
tput bit width
into some
e
x
tent in ord
e
r to
accomm
odat
e the width re
quire
m
ent of the latter stag
e.
The strategie
s
for bit-wi
dth
conversio
n
c
an be ro
ughl
y categori
z
e
d
into two gro
ups [2].
The first on
e
is basi
c
ally
an analytical approa
ch
co
ming from th
ose al
gorith
m
desig
ne
rs
who
analyze the finite word len
g
th effects du
e to m
antissa
processin
g
a
r
ithmetic. The
other app
ro
a
c
h
is b
a
sed
on
b
i
t-true
simul
a
tion o
r
igin
atio
n from
t
he
ha
rdware d
e
si
g
ners. Th
e a
n
a
lytical a
ppro
a
ch
started
from attempts
to model
q
uanti
z
ation erro
r statistically; then it wa
s
exp
ande
d to spe
c
ific
linear time in
variant
(LTI)
system
s
su
ch a
s
digital
filters, FF
T, et
c. In
the
pa
st three
de
ca
d
e
s,
nume
r
ou
s
pa
pers
have
be
en d
e
voted to
this
app
ro
ach [2-5]. The
b
i
t-true
sim
u
lat
i
on m
e
thod
h
a
s
been
extensi
v
ely used
re
cently [6-8]. Its pote
n
ti
al b
enefits lie
in
its ability to
handl
e no
n-L
T
I
sy
st
em
s
a
s
well
a
s
LTI
sy
st
em
s.
W
h
at
ev
er
, the a
f
oreme
n
tione
d two
app
ro
ach
e
s
have
the
simila
rity of handli
ng a
n
d
analy
z
ing
d
a
ta influen
ce
at the o
u
tpu
t
port of
som
e
certain
sta
ge.
While thi
s
pa
per p
r
op
oses a new
way,
which tend
s to firstly introdu
ce some
differen
c
e (i.
e
.
digital dithe
r
the belo
w
m
entione
d, co
mpared to
an
alog dith
er i
n
ADC-optimi
z
ing field [9]) into
the input port of some ce
rtain pro
c
e
s
si
ng uni
t and then execute mantis
sa p
r
o
c
e
ssi
ng at the
output p
o
rt.
There a
r
e
m
a
inly two
typ
e
s
of ma
ntissa
processin
g
: trun
catio
n
and
roun
d
to
nearest. Rou
nd to nea
re
st
employs a
n
extra add
er
f
o
r the roun
di
ng ope
ratio
n
, while tru
n
cation
dire
ctly chop
s the bits lowe
r than req
u
ire
d
lest
signifi
cant bit, which
is the main type we co
ncern
in this pap
er.
Mantissa pro
c
e
ssi
ng towa
rds m
u
lti-
bit d
a
ta, is a tran
sform
a
tion p
r
oce
s
s from hi
gh
quanti
z
ed p
r
eci
s
ion to lo
w one.
Due t
o
the red
u
cti
on in qu
antization width, t
r
un
cation
error is
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Study of Addi
tive Dithe
r
on
Rest
rainin
g Signal Tru
n
cation Erro
r (T
ao Liu)
5421
arou
se
d, which le
ad
s to t
he noi
se
pea
k in th
e outp
u
t spe
c
trum
cau
s
e
d
by ha
rmoni
c di
sto
r
tion,
and re
du
ce
s the sp
urio
us fr
ee dynami
c
range (SF
D
R).
Sampling o
p
e
ration
of ADC can al
so b
e
co
ns
i
dered
as a tran
sfo
r
mation
pro
c
ess from
high
qua
ntized p
r
e
c
i
s
ion
to lo
w
pre
c
isi
on.
Dithe
r
tech
nolo
g
y
[10-14]
aim
s
to redu
ce
t
h
e
quanti
z
ation
effect of anal
og-to
-digital
conver
to
r and i
m
prove the S
F
DR
perfo
rm
ance drasti
ca
lly,
mainly by mean
s of addi
ng ran
dom
noise into
ADC in
put sig
nal. Based
on the dithe
r
ing
prin
ciple, thi
s
pape
r mai
n
ly dra
w
s digital
dither
i
n
to a
nalysi
s
of tru
n
catio
n
erro
r
and
con
d
u
c
t the
study of re
straining tru
n
cation error. In
orde
r to
improve the harm
onic di
stortio
n
cau
s
e
d
by bit-
width trun
cati
on, we nee
d to add app
ro
priate rando
m noise into
the pre
-
trun
cated sig
nal with
long bit-width
before t
r
un
cation ope
rati
on is ex
e
c
ut
ed. The p
r
op
ose
d
metho
d
works in p
u
r
e
digital domai
n, which will
be easily a
nd efficient
ly
applied to most of the digital pro
c
e
s
sing
sy
st
em
s.
2. Trunca
t
ion Error Anal
y
s
is
Gene
rally sp
eaki
ng, erro
r
sou
r
ces in di
gital
sy
st
e
m
s
mainly
co
me
f
r
om t
he f
o
llo
wing t
w
o
asp
e
ct
s: 1
)
Quanti
z
ation
error expo
rte
d
by A
DC;
2) Tru
n
cation
e
rro
r int
r
od
uce
d
by finite
word
length effe
ct. For o
p
tion
s in whi
c
h the
circuit
stru
ct
ure is fixed, it is imp
r
a
c
tical
to redu
ce A
D
C’s
quanti
z
ation
error. As a re
sult, trun
catio
n
error be
co
mes the maj
o
r target ne
edi
ng re
solved.
Digital sig
nal
processing
is
b
a
sed on
a
se
rie
s
of
algo
rithm,
who
s
e
comp
utational
accuracy d
e
termin
es the
a
c
cura
cy
of the final output
result. In orde
r to obtain hig
her p
r
e
c
isi
on
in
field prog
ram
m
able gate
a
rray (FP
G
A), the bit-wi
dth allocated in i
n
terme
d
iate
stage
s tend
s to be
longe
r tha
n
t
hat in
the fin
a
l result. For
instan
ce,
mo
st of
digital
si
gnal
cal
c
ul
ation in
FPGA
are
develop
ed in
fixed-poi
nt
arithmeti
c
, which
mea
n
s
the op
era
n
d
s
are
all i
n
te
gers. Th
e
whole
cal
c
ulatio
n p
r
oce
s
s al
way
s
co
ntain
s
a
seri
es of
st
eps, such as multiplicat
io
n, filtering,
sig
nal
comp
re
ssing
and
so o
n
, whi
c
h
will ab
solutely in
cre
a
se th
e effici
ent bit-wi
dth
of interme
d
ia
te
results. T
a
ke
a loo
k
at the
logi
c resource of
FFT
IP core
within
AL
TERA seri
es’
FPGA, wh
ose
longe
st a
c
cepta
b
le i
nput bit-wi
dth is
2
4
-bits. After op
eration such
as
quanti
z
ation
、、
mixing
filtering an
d so o
n
, signal, wh
ose bit-
width
is mostly long
er than 24
-bit
s, is
sent to
the i
nput p
o
rt of t
he IP core.
Then,
red
und
ant bits shou
ld be
cut
so
as to
satisfy
the
requi
rem
ent.
A
ssu
me
()
x
n
is the final outp
u
t signal
with
effective bit
width of A after a se
rie
s
of
cal
c
ulatio
n, and
()
yn
is
the trunc
a
ted s
i
gnal with effec
t
iv
e bit width of
B. Executing fix-pointed
cal
c
ulatio
n in
FPGA, the
most
simpl
e
and
also the
easi
e
st
meth
od fo
r trun
cat
i
on i
s
di
sca
r
d
i
ng
the lowest
ce
rtain bit
s
dire
ctly. Let
()
zn
re
prese
n
ts the
di
scard
ed
error sign
al with
b
i
t width of
A-B, then, we
get the following e
quatio
n:
()
()
2
(
)
AB
x
ny
n
z
n
(1)
Let
()
j
X
e
、
()
j
Ye
and
()
j
Z
e
represent the Di
screte
Fouri
e
r T
r
an
sform
a
tion
(DF
T
) of
()
x
n
、
()
yn
and
()
zn
resp
ect
i
vely. Accord
ing to the linear
pro
pert
y
of Fourier
Tran
sfo
r
m, we can g
e
t:
00
()
2
(
)
(
)
jA
B
j
n
j
n
nn
X
ey
n
e
z
n
e
(2)
Whi
c
h mea
n
s
,
()
2
(
)
(
)
jA
B
j
j
X
eY
e
Z
e
(3)
Equation
(3) i
ndicates th
at, in the case o
f
relevant
spe
c
trum
pa
ram
e
ters (e.g. SF
DR) of
()
x
n
unde
r
ce
rtai
n, the spe
c
trum qu
ality of trun
cated
si
gnal
()
yn
is affe
cted
by trun
cation
error
()
zn
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5420 – 54
29
5422
As mentio
ne
d above,
()
yn
is achi
eved
by disca
r
din
g
the lo
we
st A-B bits of
()
x
n
.
Signal co
nstructed by the
discard
ed A-B bits is call
ed
()
zn
, then,
()
x
n
and
()
zn
b
e
c
ome
establi
s
h
ed a
s
:
()
()
(
m
o
d
2
)
A
B
zn
x
n
(4)
As
sho
w
n
in
Equatio
n (4), symbol
‘mo
d
’
represents the
co
ngrue
nce
op
erat
or,
whi
c
h
mean
s the
remaind
e
r
achieved by
()
x
n
d
i
viding
2
AB
. As
sume
1
()
[
0
,
2
]
A
xn
m
, then
Equation (4) i
s
equival
ent with the follo
wing formula:
()
(
m
o
d
2
)
A
B
zm
m
(5)
Equation (5
) means retai
n
ing the lowest
A
B
bits of
m
. Let
21
AB
, then
Equation (5)
can b
e
de
scri
bed a
s
Figu
re
1.
m
1
()
zm
1
Figure 1. Fun
c
tion of Tru
n
cation Erro
r
Having
a
clo
s
e-up
view o
f
Figure
1, it is not
difficu
lt to find th
at, the tru
n
cation e
r
ro
r
function
()
zm
has
the si
milar p
r
operty
with A
DC’
s
qua
ntiz
ation e
r
ror,
e
s
pe
cially th
e
perio
dicity
and linea
rity.
The differen
c
e lies in that,
()
zm
is a periodi
c discrete lin
ea
r function, wit
h
a perio
d
of
12
AB
N
. Such p
e
ri
odicity of the
truncation e
r
ror
()
zm
is refle
c
te
d in the h
a
rm
oni
c
on the output
spe
c
tru
m
, wh
ich will d
egen
erate the
spe
c
trum q
uality after trun
catio
n
.
3. Dither Pri
n
ciple
Dither i
s
a ki
nd of rando
m jitter signa
l, wh
ich is completely ind
epen
dent wit
h
ADC’
s
analo
g
inp
u
t sign
al, [15]. Harmoni
c ma
y come
from ADC’
s cohe
rent
sa
mpling,
quanti
z
ed n
o
ise
and p
e
rio
d
icit
y of differential nonlin
ea
r
error, w
ho a
r
e created
by the existed
certain
releva
nce
among
sam
p
ling, quanti
z
a
t
ion and in
pu
t waveform, [
16]. Dither i
s
just empl
oyed in orde
r to
damag
e the
relative fixed relation
ship. A
dding
dit
her i
n
to ADC’s i
n
put sig
nal a
n
d
wipi
ng off t
h
e
noise at the
output port
with digital
methods, th
e SF
DR
will be im
proved.
The operation princi
ple
is as
sho
w
n i
n
Figure 2. From the view of amp
litude
, the additive dither can b
e
cla
ssifie
d
into
dither with hi
gh amplitud
e
and low am
plitude. Wh
il
e from the view of freque
ncy, the addi
tive
dither
can b
e
classified int
o
wide
ban
d d
i
ther
and
narrowb
and dith
e
r
. Applicatio
n
occasi
on
s var
y
with different
kind
s of dithe
r
.
Figure 2. Prin
ciple of Dith
er
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Study of Addi
tive Dithe
r
on
Rest
rainin
g Signal Tru
n
cation Erro
r (T
ao Liu)
5423
The p
r
in
ciple
of dithe
r
sho
w
n i
n
Fig
u
re
2 involve
s
A
D
C an
d its
relevant a
nalo
g
ci
rcuit
desi
gn, which
is not that e
a
s
y to implem
ent in a
fixed digital
sy
stem
.
That’s why we
n
eed a
ki
nd
of pure digital
method to re
strain trun
cati
on er
ror, i.e. this pa
per
refe
rre
d digital dit
her.
4. Dither
-ba
s
e
d Res
t
rain
ts of Trunc
ati
on Error
Similar
with
ADC’
s q
uanti
z
ation
pri
n
ci
p
l
e, tr
un
cation
towa
rd
s lo
n
g
data
bit-wi
dth is a
transfo
rmatio
n pro
c
e
s
s fro
m
high q
uant
ization
pre
c
i
s
ion to lo
w p
r
eci
s
ion.
Duri
ng the p
r
o
c
e
ss,
detailed info
rmation bet
ween two
adj
ace
n
t sam
p
l
e
point de
crease with th
e red
u
ctio
n of
quanti
z
ation
step
s, whi
c
h
leads to the
fact t
hat co
ntinuou
sly variable d
e
tails among
seve
ral
sampl
e
poi
nt
of the high
preci
s
ion
sig
nal
turn in
to
a flat step
witho
u
t any variati
on. Mea
n
whil
e,
harm
oni
c dist
ortion is int
r
o
duced an
d SFDR i
s
de
cre
a
se
d, [17, 18].
Among
th
e rang
e of
()
[
0
,
2
1
]
A
xn
m
, t
r
unc
a
tion error
()
zm
is a di
screte
sa
wtooth wa
ve
functio
n
with
peri
od of N.
When
[0
,
1
]
mN
,
()
zm
m
. Thus the
N-p
o
in
t
Fouri
e
r Seri
e
s
of
()
zm
is listed
as:
0
11
(2
/
)
00
()
NN
jk
m
j
kN
m
kk
kk
zm
a
e
a
e
(6)
Coeffici
ent
k
a
in the Fourie
r Serie
s
expression Equation
(6) i
s
:
0
1
0
1
(2
/
)
0
/
1
()
1
c
o
s(
2
/
)
1
sin(
2
/
)
2[
1
c
o
s
(
2
/
)
]
2s
i
n
(
/
)
N
jk
m
k
m
N
jk
m
N
m
jk
N
am
ze
N
m
e
N
kN
j
k
N
kN
je
kN
(7)
Fouri
e
r transf
o
rmatio
n exp
r
es
sio
n
of pe
riodi
cal
sign
a
l
()
zm
within on
e perio
d can b
e
obtaine
d from
Equation (7
).
/
1
0
2
()
(
)
si
n
(
/
)
jk
N
N
j
k
ek
Ze
j
kN
N
(8)
Whe
n
k
value is small,
the amplitud
e-fre
que
ncy
cha
r
a
c
teri
stic function ca
n be
descri
bed a
s
:
2
()
(
)
j
k
Nk
Ze
kN
(9)
What
can
be
dedu
ce
d fro
m
Equation
(9) is th
at, variation of ha
rmonic
amplit
ude h
a
s a
inverse ratio t
o
the valu
e of
k
, and atten
u
a
tes
slo
w
ly. In order to rap
i
dly attenuate
the ha
rmo
n
ic
amplitude
in the
tru
n
catio
n
e
rro
r, we p
l
an
to add
ra
ndom
noi
se
with width of
A
B
bits into
sign
al
()
x
n
before trun
catio
n
operation i
s
execute
d
.
Th
e ad
ded
ra
nd
om noi
se
is the
so-call
e
d
dither di
scussed in thi
s
pa
per. Wi
dth of dither
i
s
ba
se
d on the follo
wing
con
s
id
e
r
ation: 1)
Over-
high amplitude will int
r
o
duce floor-noi
se-ri
sing
of
()
yn
; 2) Und
e
rsi
z
e a
m
plitude i
s
in
adeq
uate
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5420 – 54
29
5424
for ch
angin
g
the step
-characteri
stic of
()
yn
. The additive
dither with
A
B
bits sh
ould n
o
t
affect the floor-noise, but will well
change the step-cha
racteri
s
tic of
()
yn
.
As dithe
r
i
s
a
statisti
c
sign
al, we
ne
ed t
o
re
present t
r
un
cation
e
r
ror
with m
a
th
ematical
expectatio
n
. Let
si
gnal
d
indicates the
additive dith
e
r
, an
d
[0
,
2
1
]
AB
dZ
, wh
ere
Z
rep
r
e
s
ent
s the set of integers, which mean
s
d
is a discrete
ran
dom
variable. Assume the ne
w
truncation e
r
ror is
()
zm
, then:
1
0
()
[
(
)
]
()
(
)
N
l
zm
E
z
m
d
zm
l
p
l
(10)
In Equatio
n (10),
2
AB
N
, and
()
pl
is
the di
strib
u
tion fun
c
tion
of
rand
om va
ria
b
le
d
,
and:
()
{
}
pl
P
d
l
(11)
Equation (10
)
ca
n be reckon
ed a
s
the
cro
s
s-co
rrel
a
tion functio
n
of signal
()
zm
and
distrib
u
tion fu
nction
()
pl
. The freque
ncy-dom
ain expre
s
sio
n
can b
e
writt
en as:
()
(
)
()
jj
j
Z
eZ
e
P
e
(12)
In the followi
ng analy
s
is,
dither
with G
aussi
a
n
dist
ri
bution an
d u
n
iform di
strib
u
tion are
introdu
ce
d to demon
strate the re
strai
n
t on truncation e
rro
r.
4.1. Gaussia
n
Dither
Res
t
rains Tru
n
c
a
tion Error
Whe
n
dithe
r
i
s
a di
screte signal
with Ga
ussian
distri
b
u
tion [19], we
need to
wo
rk out the
distrib
u
tion fu
nction of Equ
a
tion (11
)
, i.e.
()
pl
.
Assu
me the
continu
o
u
s
rando
m variable
()
dt
obeys Gaussia
n
d
i
stributio
n, with
averag
e valu
e of
and variance of
2
, i.e.
2
()
(
,
)
dt
N
. Suppose
()
dt
an
d
d
abide by
the relation
sh
ip as:
()
dd
t
(13)
()
dt
mean
s th
e b
i
gge
st intege
r not g
r
eate
r
t
han
()
dt
, then
()
pl
can b
e
a
c
hiev
ed
from the follo
wing e
quatio
n:
2
1
2
1(
)
(
)
e
xp[
]
2
2
l
l
t
pl
d
t
(14)
The Fou
r
ie
r transfo
rmatio
n of
()
pl
is
()
j
Pe
, and:
1
0
2
1
1
2
0
()
(
)
1(
)
ex
p
[
]
2
2
N
jj
l
l
N
l
j
l
l
l
Pe
p
l
e
t
dt
e
(15)
The amplitu
d
e
-fre
que
ncy chara
c
te
risti
c
is:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Study of Addi
tive Dithe
r
on
Rest
rainin
g Signal Tru
n
cation Erro
r (T
ao Liu)
5425
2
1
1
2
0
2
1
1
2
0
2
1
1
2
0
2
2
0
1(
)
()
e
x
p
[
]
2
2
1(
)
e
xp[
]
2
2
1(
)
e
xp[
]
2
2
1(
)
ex
p
[
]
2
2
N
l
j
jl
l
l
N
l
j
l
l
l
N
l
l
l
N
t
Pe
d
t
e
t
dt
e
t
dt
t
dt
(16)
Let
2
2
0
1(
)
e
xp[
]
1
2
2
N
t
Qd
t
, whe
r
e v
a
lue
of
Q
indicates th
e a
r
e
a
amo
n
g
the ran
ge of
[0
,
]
N
below the probability density plot of
()
dt
. Value of
Q
will diminish, as
var
i
anc
e
2
increa
se
s.
B
e
ca
use
()
1
j
Pe
Q
,
()
j
Pe
will diminish
with the increase
of
2
. Thus, wh
at can b
e
inferre
d
from Equati
on (12
)
is:
()
(
)
()
()
()
(
)
jj
j
j
jj
Ze
Z
e
P
e
Ze
Q
Z
eZ
e
(17)
Equation
(17
)
tells u
s
that, becau
se of t
he influe
nce
of Gau
ssi
an
dither, the h
a
rmoni
c
amplitude
in t
r
un
cation
e
r
ror
spe
c
trum i
s
smalle
r th
a
n
that
without
Gau
s
sian
dit
her. M
ean
whi
l
e,
the harm
oni
c amplitude will decrea
s
e
si
gnifi
cantly with the increase of
2
, whic
h effic
i
ently
illustrates dither’s
restraint on
truncation error.
4.2. Uniform
Dither
Res
t
r
a
ins Trunc
ation Error
Whe
n
the
ad
ditive dithe
r
i
s
a
u
n
iform
distrib
u
ted
si
gnal
and
[0
,
1
]
dN
Z
, the
distrib
u
tion fu
nction
()
pl
ca
n b
e
achieved
a
c
cordi
ng to t
he
definitio
n of
uniform di
stributio
n
,
[20].
1
()
{
}
pl
P
d
l
N
(18)
Execute Fou
r
ier tran
sform
a
tion towa
rd
s Equation (1
8
)
:
1
0
1
0
()
(
)
1
1
(1
)
N
j
jl
l
N
j
l
l
jN
j
Pe
p
l
e
e
N
e
Ne
(19)
Then, the am
plitudefre
que
ncy ch
ara
c
te
ri
stic of the tru
n
catio
n
error i
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5420 – 54
29
5426
2/
2/
2
2/
()
(
)
()
21
()
(1
)
12
()
(1
)
12
()
(1
)
12
()
(1
)
0
jj
j
jN
j
k
jN
j
k
jk
N
N
jk
N
k
jk
jk
N
k
Ze
Z
e
P
e
Nk
e
kN
N
e
ek
ke
N
ek
ke
N
ek
ke
N
(20)
In actu
al
ca
ses, b
e
cau
s
e
of the finite l
ength
effect
of uniform di
stribute
d
dith
er, the
occurre
n
ce p
r
oba
bility of e
a
ch
point i
n
Equation
(18
)
turn
s n
o
t out
to be
always the same
an
d
exists
a
cert
ain differen
c
e from
the i
deal valu
e o
f
1/
N
. As a
re
sult, the a
c
tu
al value
of
()
j
Z
e
will not equal to zero, but
only be close to
zero. Any
how, what can be inferred from
the above an
alysis is that
uniform distributed dith
e
r
also ha
s the ability to restrain trun
catio
n
error effe
ctively.
5. Simulation and Verific
a
tion
A
ssu
me
()
x
n
is a sine sig
nal with bit-width o
f
14
A
bits. The sample
-rate i
s
1MSPS
and f
r
eq
uen
cy stand
s at
170
kHz.
Widt
h of
sig
nal
()
yn
obtaine
d afte
r truncation
i
s
10
B
.
Then the wi
dth of truncation error
()
zm
is
4
A
B
.
Firstly, add a
Gau
ssi
an dit
her
with wi
dth of 4 bits int
o
sig
nal
()
x
n
. C
o
mp
a
r
is
on
c
har
t in
time-dom
ain
and fre
quen
cy-domai
n of the trun
catio
n
erro
r
()
zm
of signal witho
u
t dither an
d
()
zm
of signal with
dither ca
n be
sho
w
n in Fig
u
re 3 an
d Fig
u
re 4.
Figure 3. The
Time-do
m
ain
Compa
r
i
s
on
of
Trun
catio
n
Error
without an
d with Gau
s
si
an
Dither
Figure 4. The
Frequ
en
cy-d
omain Comp
arison
of Trun
cation
Erro
r witho
u
t
and with G
a
u
ssi
an Dith
er
Figure 3 tell
s
the fact that
waveform of t
r
un
cation
error without Ga
ussian dithe
r
in
time-
domain appe
ars obviou
s
perio
dicity,
which
le
ad
s
to
the harm
oni
c disto
r
tion in
the spe
c
tru
m
in
Figure 4. The influence
caused by
harm
onic
distortion of truncation
error is that, it will introduce
a redu
ndant
pea
k
sign
al i
n
the
spe
c
tru
m
of trun
cate
d si
gnal
()
yn
. Performan
c
e
of
SFDR go
es
0
100
200
300
400
500
600
0
5
10
15
T
r
unc
at
i
o
n
er
r
o
r
w
i
t
h
out
G
aus
s
i
an di
t
her
0
100
200
300
400
500
600
0
5
10
15
T
r
unc
at
i
on er
r
o
r
w
i
t
h
G
aus
s
i
an di
t
her
0
50
100
150
200
250
300
350
400
45
0
500
-
150
-
100
-5
0
0
F
r
equenc
y
/
k
H
z
A
m
p
l
i
t
ude
/
d
B
Er
r
o
r
s
pec
t
r
um
w
i
t
hout
G
aus
s
i
an di
t
her
0
50
100
150
200
250
300
350
400
45
0
500
-8
0
-6
0
-4
0
-2
0
0
F
r
equenc
y
/
k
H
z
A
m
pl
i
t
u
de/
d
B
Er
r
o
r
s
pec
t
r
um
w
i
t
h
G
aus
s
i
an di
t
her
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Study of Addi
tive Dithe
r
on
Rest
rainin
g Signal Tru
n
cation Erro
r (T
ao Liu)
5427
worse
owi
ng
to the mentio
ned influe
nce
.
With addi
tive dither, the
perio
dicity is
damag
ed, wh
ich
help
s
de
cre
a
s
e the ha
rm
onic di
stortio
n
and incre
a
s
e SFDR of
signal
()
yn
, as sho
w
n in
Figure 5.
Figure 5. The
Frequ
en
cy-d
omain Comp
arison of
Tru
n
cate
d Signal
without and
with Gau
s
sia
n
Dither
It can be infe
rre
d from Fi
g
u
re 5 th
at, SFDR
of trun
cated si
gnal
with additive G
aussia
n
dither in
crea
ses ab
out 16d
B.
Put another
two si
gnal
wi
th freque
ncy
of 110kHz
and 27
0kHz
in
()
x
n
, and ad
d
a
uniform dithe
r
with 4-bit
s
, then, the time
-do
m
ain an
d freque
ncy
-
domain
comp
arison
s of
()
zm
and
'
()
zm
are described in Fi
gure 6 and Figu
re 7 respe
c
tively.
Figure 6. The
time-domai
n Comp
ari
s
o
n
of
Trun
catio
n
Error
without an
d with Unifo
r
m
Dither
Figure 7. The
Frequ
en
cy-d
omain Comp
arison
of Trun
cation
Erro
r witho
u
t and with
Unif
orm
Dither
Same as wha
t
demonstrate
s
in Figure 3, Figur
e 6 sho
w
s that the a
dditive uniform dither
damag
es
sig
nal’s p
e
ri
odi
ci
ty, which hel
p
s
in
cre
a
se SF
DR
of trun
cat
ed si
gnal
()
yn
, as sho
w
n in
Figure 8.
0
50
100
15
0
20
0
25
0
30
0
35
0
40
0
45
0
50
0
-
150
-
100
-50
0
X
:
230
Y
:
-
76.56
F
r
eq
uen
c
y
/
k
H
z
A
m
pl
i
t
ude/
d
B
Sp
ec
t
r
u
m
w
i
t
h
ou
t
G
a
u
s
s
i
an
dit
h
e
r
0
50
100
15
0
20
0
25
0
30
0
35
0
40
0
45
0
50
0
-
150
-
100
-50
0
F
r
eq
uen
c
y
/
k
H
z
A
m
pl
i
t
u
d
e/
dB
S
p
e
c
tr
u
m
w
i
th
G
a
u
ssi
a
n
d
i
th
e
r
0
100
200
30
0
400
500
600
0
5
10
15
T
r
un
c
a
t
i
on er
r
o
r
w
i
t
hout
uni
f
o
r
m
di
t
h
er
0
100
200
30
0
400
500
600
0
5
10
15
T
r
unc
at
i
o
n
e
r
r
o
r
w
i
t
h
uni
f
o
r
m
di
t
her
0
50
10
0
15
0
20
0
25
0
30
0
35
0
40
0
45
0
50
0
-1
50
-1
00
-5
0
0
F
r
equ
en
c
y
/
k
H
z
A
m
pl
i
t
ud
e/
dB
E
r
r
o
r
sp
e
c
tr
u
m
w
i
th
o
u
t u
n
i
f
o
r
m
d
i
th
e
r
0
50
10
0
15
0
20
0
25
0
30
0
35
0
40
0
45
0
50
0
-1
00
-5
0
0
F
r
equ
en
c
y
/
k
H
z
A
m
p
l
i
t
u
de/
dB
E
r
r
o
r
sp
e
c
tr
u
m
w
i
th
u
n
i
f
o
r
m
d
i
th
e
r
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5420 – 54
29
5428
Figure 8. The
Frequ
en
cy-d
omain Comp
arison of
Tru
n
cate
d Signal
without and
with Unifo
r
m
Dither
It can
be i
n
ferred from Fi
gure
8 th
at, SFDR of trun
cated
si
gnal
with a
dditive
uniform
Dither in
crea
se
s abo
ut 15
dB.
6. Conclusio
n
In this paper,
we have pre
s
ente
d
a ne
w way to an
alyze the tru
n
catio
n
error in digital
system
s, e
s
p
e
cially in
FPGA t
hat refe
rred
above.
Compa
r
ing to t
he an
alog
dither te
ch
nolog
y in
ADC-optimi
z
i
ng field, digit
a
l dither is i
m
porte
d
into
the analy
z
in
g pro
c
e
s
s of
trun
cation
e
rro
r,
whi
c
h hel
ps restrai
n
it and improve the
system perfo
rmance.
The
pro
p
o
s
e
d
ap
proa
ch t
heoretically
a
nalyz
e
s
th
e p
eak noi
se
introdu
ced
by truncation
harm
oni
c di
stortion, a
nd
d
edu
ce
s the
mathemati
c
expre
ssi
on
of trun
cation
error with
addit
i
ve
dither ba
sed
on statistical approa
ch.
In addition,
MA
TLAB
sim
u
lat
i
on
of situatio
ns with additi
ve
Gau
ssi
an dit
her
and
unif
o
rm
dither is pre
s
e
n
ted.
The
simulati
on result rev
eals t
hat, sig
nals’
SFDR
pe
rformance will be
in
crea
sed
sig
n
ificantly
if we
add
anothe
r dith
er
sign
al b
e
fore
operation of
signal
s’ preci
s
i
on c
onversio
n
. The theo
re
tically ded
uc
e
d
expre
s
sion
kee
p
s i
denti
c
al
with s
i
mulation result.
Referen
ces
[1]
Sang
hamitra
Ro
y
,
Prith Ba
nerj
ee. An Alg
o
rithm
for
T
r
adin
g
off Quantizatio
n Error w
i
t
h
Hard
w
a
re
Resources for
MAT
L
AB-Based FPGA Design.
IEEE Transactions
on Com
p
uters.
20
05;
54(7): 8
86-
895.
[2]
C Shi. Statistical Method for F
l
oati
ng-Po
int to F
i
xed Po
int C
onvers
i
on. MS T
hesis, Electrical Eng. An
d
Computer Science Dept., Univ. of
California, Berkeley
. 2002.
[3]
KH Ch
an
g, W
G
Bliss. F
i
nite
W
o
rd-Le
ngth
E
ffects of Pipe
line
d
R
e
curs
iv
e Di
gital
Filters
.
IEEE Trans.
Sign
al Process
i
ng.
19
94; 42(
8
)
: 1983-1
9
9
5
.
[4]
LBJackson, K
H
Chang, WG Bliss.
Co
mmen
t
s on
‘
F
inite
W
o
rd-L
engt
h Effects of Pi
pel
in
ed
Recurs
i
v
e
Digita
l
Filters
.
IEEE Trans. Signal Processing
. 1995; 4
3
(12)
: 3030-3
0
3
2
.
[5]
RM Gray
, DL
Neuhoff.
Quanti
z
at
io
n
.
IEEE Tans. Informatio
n
Theory
. 19
98
; 44(6): 232
5-2
383.
[6]
W
Sung, KI K
u
m. Simul
a
tio
n
-
Based W
o
r
d
-L
ength
Optimiza
tion Meth
od f
o
r F
i
xe
d-Po
int
Digita
l
Si
gn
al
Processing S
ystems.
IEEE
Tr
ans. Signal Processing.
199
5; 43(1
2
): 220
9-2
212.
[7]
S Kim, WSung.
F
i
xed-Po
int
Error An
alysis
and
W
o
rdl
e
n
g
th Opti
mi
z
a
t
i
on
of a
Distri
but
ed Arit
h
m
etic
Based 8×
8
2D-
I
DCT
Architecture
. Proc. W
o
rkshop VLSI Si
gna
l Processi
n
g
. 1996: 3
98-4
07.
[8]
H Ked
i
n
g
, M
Willems, M C
o
ors, H Me
yr. F
R
IDGE:
A F
i
xed-Poi
n
t Desi
gn
and
Si
mu
lati
o
n
Envir
o
n
m
ent
.
Proc. Desig
n
, Automatio
n
, an
d T
e
st in Europ
e
. 1998: 4
29-4
35.
[9]
Dias P, Silva
G, Cruz S.
Ditheri
ng p
e
rfor
mance of ov
er sampl
ed
ADC s
ystem
s affected by hysteres
is
.
Journ
a
l of the Internati
o
n
a
l Measur
em
ent C
onfed
erati
on. 2
002; 32(
1): 51-
59.
[10]
Wagd
y Z, Fa
w
z
y
M. Effect of
add
itive dit
her
on the res
o
luti
on of ADC’s
w
i
th singl
e-bit or
mulib
it errors.
IEEE Transactions on Instrum
entation and
Measur
em
ent.
1
996; 45(
2): 610
-615.
[11]
Suresh
B, W
o
ll
man H
B
. T
e
sting
an
A
DC li
n
eariz
ed
w
i
th ps
eud
oran
dom
di
ther.
IEEE Transactions on
Instrume
ntatio
n and Me
asur
e
m
e
n
t.
1998; 4
7
(
4): 839-8
48.
[12]
Z
hang Yu
n, Li Guang
jun.
A
pipel
in
ed AD
C structure adapta
b
l
e
to dither introd
ucti
on
. Moder
n
Electron
ics T
e
chni
que. 2
011;
34(10): 16
0-1
62.
0
50
10
0
15
0
20
0
250
300
35
0
400
45
0
500
-150
-100
-50
0
X
:
89.97
Y
:
-
73.8
F
r
equ
enc
y
/
k
H
z
A
m
pl
i
t
ude
/
d
B
Sp
ec
t
r
um
w
i
t
hout
un
if
o
r
m
d
i
t
her
0
50
10
0
15
0
20
0
250
300
35
0
400
45
0
500
-150
-100
-50
0
F
r
equ
enc
y
/
k
H
z
A
m
pl
i
t
ude
/
d
B
S
p
e
c
tr
u
m
w
i
th
u
n
i
f
o
r
m
d
i
th
e
r
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Study of Addi
tive Dithe
r
on
Rest
rainin
g Signal Tru
n
cation Erro
r (T
ao Liu)
5429
[13]
Blesser B, Lo
cantii B.
T
he app
licati
on
of narrow
b
a
nd
D
i
t
her
oper
atin
g
at the Nyquis
tfrequcney i
n
digit
a
l system
s
to provide im
proved signal to
nois
e
ratio over
conventional Dithering
. Au
di
o Eng. 1
987
;
35(6): 44
6-4
5
4
.
[14]
Anna
D. A-D c
onvers
i
on
w
i
th
Dither si
gn
al-p
ossibi
liti
e
s an
d
limitatio
ns.
Me
asure
m
ent Sci
ence
Revi
ew
.
200
1; 1(1): 75-
78.
[15]
Shu YS, So
n
g
BS. A 15
bit li
near
20M
Sampl
e
/spip
e
li
ned A
DC
dig
i
tall
y ca
libr
a
ted
w
i
t
h
si
gna
l-
dep
en
dentD
i
th
erin
g.
IEEE Journal Soli
d-State Circuits.
200
8; 43(2): 34
2-3
50.
[16]
Yu CH
H, LI J
L
. A W
h
ite N
o
i
s
e F
ilteri
ng Me
t
hod
for
DOA Estimation of Con
here
n
t
Sig
nals un
der
L
o
w
SNR. Sign
al
Pr
ocessi
ng.
20
12
; 28(7): 957-9
6
2
.
[17]
W
ang LB,
Cui
CH, Sh
a Z
H
H. Sparse
De
compos
ition
Method
of Smo
o
th Sig
n
a
l
U
n
der T
r
uncatio
n
Effect.
Signal Processi
ng.
20
11; 27(6): 9
56-
960.
[18]
Che
n
T
Q
, Xu J, Z
hu K. Error Anal
ys
is and
S
y
stem
Desig
n
of Hi
g
h–Accur
a
c
y
P
i
peli
n
e
d
A/D
Conv
erters.
Microel
ectronics.
200
8; 38(1): 12
6-12
8.
[19]
Che
ng M Z
H
, Jing W
P
. Design a
nd Ana
l
ysis of a Nov
e
l Pipe
li
ned A
DC.
Journa
l of University of
Electron
ic Scie
nce an
d T
e
chn
o
lo
gy of Chi
n
a
.
2008; 3
7
(6): 9
30-9
33.
[20]
W
agd
y MF
, Ng
W
.
Validit
y
of uniform q
u
a
n
ti
zation
error m
ode
l of sin
u
soi
dal si
gn
als
w
i
t
hout a
nd
w
i
t
h
Dither.
IEEE Transacti
ons o
n
Instrumentati
o
n
and Meas
ure
m
e
n
t
. 1989; 3
8
(
3): 718-7
22.
Evaluation Warning : The document was created with Spire.PDF for Python.