<s>
Noise-Predictive	B-Device
Maximum-Likelihood	I-Device
(	O
NPML	B-Device
)	O
is	O
a	O
class	O
of	O
digital	O
signal-processing	O
methods	O
suitable	O
for	O
magnetic	B-Architecture
data	I-Architecture
storage	I-Architecture
systems	I-Architecture
that	O
operate	O
at	O
high	O
linear	O
recording	O
densities	O
.	O
</s>
<s>
It	O
is	O
used	O
for	O
retrieval	O
of	O
data	O
recorded	O
on	O
magnetic	B-Architecture
media	I-Architecture
.	O
</s>
<s>
NPML	B-Device
aims	O
at	O
minimizing	O
the	O
influence	O
of	O
noise	O
in	O
the	O
detection	O
process	O
.	O
</s>
<s>
Alternatives	O
include	O
peak	O
detection	O
,	O
partial-response	B-Device
maximum-likelihood	I-Device
(	O
PRML	B-Device
)	O
,	O
and	O
extended	O
partial-response	O
maximum	O
likelihood	O
(	O
EPRML	O
)	O
detection	O
.	O
</s>
<s>
Although	O
advances	O
in	O
head	O
and	O
media	O
technologies	O
historically	O
have	O
been	O
the	O
driving	O
forces	O
behind	O
the	O
increases	O
in	O
the	O
areal	O
recording	O
density	O
,	O
digital	B-General_Concept
signal	I-General_Concept
processing	I-General_Concept
and	O
coding	B-General_Concept
established	O
themselves	O
as	O
cost-efficient	O
techniques	O
for	O
enabling	O
additional	O
increases	O
in	O
areal	O
density	O
while	O
preserving	O
reliability	O
.	O
</s>
<s>
The	O
NPML	B-Device
family	O
of	O
sequence-estimation	O
data	O
detectors	O
arise	O
by	O
embedding	O
a	O
noise	O
prediction/whitening	O
process	O
into	O
the	O
branch	O
metric	O
computation	O
of	O
the	O
Viterbi	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
<s>
Reliable	O
operation	O
of	O
the	O
process	O
is	O
achieved	O
by	O
using	O
hypothesized	O
decisions	O
associated	O
with	O
the	O
branches	O
of	O
the	O
trellis	O
on	O
which	O
the	O
Viterbi	B-Algorithm
algorithm	I-Algorithm
operates	O
as	O
well	O
as	O
tentative	O
decisions	O
corresponding	O
to	O
the	O
path	O
memory	O
associated	O
with	O
each	O
trellis	O
state	O
.	O
</s>
<s>
NPML	B-Device
detectors	O
can	O
thus	O
be	O
viewed	O
as	O
reduced-state	O
sequence-estimation	O
detectors	O
offering	O
a	O
range	O
of	O
implementation	O
complexities	O
.	O
</s>
<s>
By	O
judiciously	O
choosing	O
,	O
practical	O
NPML	B-Device
detectors	O
can	O
be	O
devised	O
that	O
improve	O
performance	O
over	O
PRML	B-Device
and	O
EPRML	O
detectors	O
in	O
terms	O
of	O
error	O
rate	O
and/or	O
linear	O
recording	O
density	O
.	O
</s>
<s>
In	O
the	O
absence	O
of	O
noise	O
enhancement	O
or	O
noise	O
correlation	O
,	O
the	O
PRML	B-Device
sequence	O
detector	O
performs	O
maximum-likelihood	O
sequence	O
estimation	O
.	O
</s>
<s>
requiring	O
a	O
-state	O
NPML	B-Device
detector	O
if	O
no	O
reduced-state	O
detection	O
is	O
employed	O
.	O
</s>
<s>
The	O
NPML	B-Device
detector	O
is	O
efficiently	O
implemented	O
via	O
the	O
Viterbi	B-Algorithm
algorithm	I-Algorithm
,	O
which	O
recursively	O
computes	O
the	O
estimated	O
data	O
sequence	O
.	O
</s>
<s>
In	O
other	O
words	O
,	O
the	O
optimum	O
MLSE	O
can	O
be	O
implemented	O
efficiently	O
by	O
using	O
the	O
Viterbi	B-Algorithm
algorithm	I-Algorithm
,	O
in	O
which	O
the	O
branch-metric	O
computation	O
involves	O
data-dependent	O
noise	O
prediction	O
.	O
</s>
<s>
Because	O
the	O
predictor	O
coefficients	O
and	O
prediction	O
error	O
both	O
depend	O
on	O
the	O
local	O
data	O
pattern	O
,	O
the	O
resulting	O
structure	O
has	O
been	O
called	O
a	O
data-dependent	O
NPML	B-Device
detector	O
.	O
</s>
<s>
Reduced-state	O
sequence	O
detection	O
schemes	O
can	O
be	O
applied	O
to	O
data-dependent	O
NPML	B-Device
,	O
reducing	O
implementation	O
complexity	O
.	O
</s>
<s>
NPML	B-Device
and	O
its	O
various	O
forms	O
represent	O
the	O
core	O
read-channel	O
and	O
detection	O
technology	O
used	O
in	O
recording	O
systems	O
employing	O
advanced	O
error-correcting	O
codes	O
that	O
lend	O
themselves	O
to	O
soft	O
decoding	O
,	O
such	O
as	O
low-density	B-Error_Name
parity	I-Error_Name
check	I-Error_Name
(	O
LDPC	B-Error_Name
)	O
codes	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
noise-predictive	O
detection	O
is	O
performed	O
in	O
conjunction	O
with	O
a	O
maximum	B-General_Concept
a	I-General_Concept
posteriori	I-General_Concept
(	O
MAP	O
)	O
detection	O
algorithm	O
such	O
as	O
the	O
BCJR	B-Error_Name
algorithm	I-Error_Name
then	O
NPML	B-Device
and	O
NPML-like	O
detection	O
allow	O
the	O
computation	O
of	O
soft	O
reliability	O
information	O
on	O
individual	O
code	O
symbols	O
,	O
while	O
retaining	O
all	O
the	O
performance	O
advantages	O
associated	O
with	O
noise-predictive	O
techniques	O
.	O
</s>
<s>
Beginning	O
in	O
the	O
1980s	O
several	O
digital	O
signal-processing	O
and	O
coding	B-General_Concept
techniques	O
were	O
introduced	O
into	O
disk	O
drives	O
to	O
improve	O
the	O
drive	O
error-rate	O
performance	O
for	O
operation	O
at	O
higher	O
areal	O
densities	O
and	O
for	O
reducing	O
manufacturing	O
and	O
servicing	O
costs	O
.	O
</s>
<s>
In	O
the	O
early	O
1990s	O
,	O
partial-response	O
class-4	O
(	O
PR4	O
)	O
signal	O
shaping	O
in	O
conjunction	O
with	O
maximum-likelihood	O
sequence	O
detection	O
,	O
eventually	O
known	O
as	O
PRML	B-Device
technique	O
replaced	O
the	O
peak	O
detection	O
systems	O
that	O
used	O
run-length-limited	O
(	O
RLL	O
)	O
(	O
d	O
,	O
k	O
)	O
-constrained	O
coding	B-General_Concept
.	O
</s>
<s>
This	O
development	O
paved	O
the	O
way	O
for	O
future	O
applications	O
of	O
advanced	O
coding	B-General_Concept
and	O
signal-processing	O
techniques	O
in	O
magnetic	B-Architecture
data	I-Architecture
storage	I-Architecture
.	O
</s>
<s>
NPML	B-Device
detection	O
was	O
first	O
described	O
in	O
1996	O
and	O
eventually	O
found	O
wide	O
application	O
in	O
HDD	O
read	O
channel	O
design	O
.	O
</s>
<s>
The	O
“	O
noise	O
predictive	O
”	O
concept	O
was	O
later	O
extended	O
to	O
handle	O
autoregressive	B-Algorithm
(	O
AR	O
)	O
noise	O
processes	O
and	O
autoregressive	B-Algorithm
moving-average	O
(	O
ARMA	O
)	O
stationary	O
noise	O
processes	O
The	O
concept	O
was	O
extended	O
to	O
include	O
a	O
variety	O
of	O
non-stationary	O
noise	O
sources	O
,	O
such	O
as	O
head	O
,	O
transition	O
jitter	O
and	O
media	O
noise	O
;	O
it	O
was	O
applied	O
to	O
various	O
post-processing	O
schemes	O
.	O
</s>
<s>
The	O
pioneering	O
research	O
work	O
on	O
partial-response	B-Device
maximum-likelihood	I-Device
(	O
PRML	B-Device
)	O
and	O
noise-predictive	B-Device
maximum-likelihood	I-Device
(	O
NPML	B-Device
)	O
detection	O
and	O
its	O
impact	O
on	O
the	O
industry	O
were	O
recognized	O
in	O
2005	O
by	O
the	O
European	O
Eduard	O
Rhein	O
Foundation	O
Technology	O
Award	O
.	O
</s>
<s>
NPML	B-Device
technology	O
were	O
first	O
introduced	O
into	O
IBM	O
's	O
line	O
of	O
HDD	O
products	O
in	O
the	O
late	O
1990s	O
.	O
</s>
<s>
In	O
2010	O
,	O
NPML	B-Device
was	O
introduced	O
into	O
IBM	O
's	O
Linear	B-Operating_System
Tape	I-Operating_System
Open	I-Operating_System
(	O
LTO	O
)	O
tape	O
drive	O
products	O
and	O
in	O
2011	O
in	O
IBM	O
's	O
enterprise-class	O
tape	O
drives	O
.	O
</s>
