Title: Coherent Detection for Optical Communications using Digital Signal Processing
1Coherent Detection for Optical Communications
using Digital Signal Processing
Atlantic Sciences
- Michael G. Taylor
- Optical Networks Group, University College London
- and
- Atlantic Sciences
- e-mail mtaylor_at_atlanticsciences.com
2Outline
- Why use coherent detection?
- Why is coherent detection feasible now? arrival
of real-time DSP - Burst mode coherent detection experiments
- Constraints imposed by parallel digital
processing phase estimation - Future developments summary
3Benefits of coherent detection
- Optical gain
- Only light in close neigbourhood of local
oscillator wavelength is seen by coherent
detection - acts like an ultra-narrow WDM filter
- behaves as a tunable filter if tunable LO is used
- Phase encoded modulation formats can be detected
- e.g. binary (BPSK) quadrature (QPSK) modulation
formats have 3dB better sensitivity than on-off
formats - QPSK carries 2 bits/symbol
- Equalisation of propagation impairments in
electrical domain is equivalent to electric field
equalisation - compensate for chromatic dispersion in IF using
microstrip line
4Difficulties with coherent detection
- More complex receiver
- To get best sensitivity and detect high bit rate
signals homodyne detection must be used - LO phase locked to incoming signal
- Polarisation management needed to match SOP of LO
to incoming signal - active polarisation control or polarisation
diversity or polarisation switching - To achieve best sensitivity synchronous detection
needed - electronics to lock to wandering phase
5Sampled coherent detection
- Apply real time digital signal processing
technology to coherent detection - already used in receivers for impairment
compensation after direct detection - "Hard" part of coherent detection will be done by
a digital processor - polarisation management
- phase estimation
- equalisation of propagation impairments
- Very flexible solution, since DSP can be
reconfigured under software control - inadequacies of transmitter/receiver hardware can
be compensated in DSP - All benefits of coherent detection available
simultaneously - detects phase- polarisation-encoded formats
- allows many bits/symbol
- best possible receiver sensitivity
- ultra-narrow WDM
- etc.
- Transceiver can re-use transmit laser as local
oscillator in receiver
6Quadrature sampling
90 hybrid - passive unit
incoming signal
sin(wLOt)
DSP
cos(wLOt)
local oscillator
photo- detector
A/D converter
extra phase shift by 90
- Phase diverse apparatus used to combine signal
LO - DSP unit processes a digitised representation of
detected signals in two arms - Polarisation tracking done by two 90 hybrids in
polarisation diverse topology - Local oscillator can be nominally same frequency
as signal but not phase locked to it
7Coherent detection experiments
8Proof of principle experiment
noise loading apparatus
pattern generator 10.7 Gb/s
real time sampling scope 20 GSa/s
arrangement of fiber pigtailed passive splitters
photo-detector
variable attenuator
EDFA
phase modulator
tunable laser
1nm filter
tunable laser
- Continuous sample 4ms long recorded on scope,
then processed later offline by PC - BPSK modulation format
- Polarisation of LO matched only approximately to
signal by manipulation of fiber coils - 90 phase shift achieved by coincidental
difference in length between arms - multiple samples recorded and then best result
chosen
9How data from experiment is processed
Two waveforms downloaded from oscilloscope.
Two channels combined to give complex electric
field.
Equalisation filter applied to each channel -
reverses non-flat frequency response of
electronics. Same filter applied to all data
sets.
For measurements over fiber, CD equalisation is
applied.
Q factor calculated using decision threshold
method (based on samples at bit centres).
Clock frequency (10.66GHz) beat envelope (about
100MHz) recovered.
Smooth waveform by interpolating points in
between half bit times, and hence generate eye
diagram.
Channels retimed to sample rate of 2 x clock
frequency (alternate samples at bit centre).
10Experiment results
- Example of measured data OSNR 31dB data point
waveforms at two outputs of 90 hybrid
eye diagram
11Experiment results
- Each measured data point comes from a 4ms sample
- Q calculated by decision threshold method
- Typical IM-DD result from Taylor et al., ECOC
2002 - Theoretical sensitivity from Yamamoto, J. Quantum
Electron., QE-16, p. 1251, 1980
theoretical limit
typical 10G IM-DD transmitter-receiver
2.5dB
4.5dB
12Experiment results
- Equalisation done by convolution with 9 element
vector (FIR filter fractional spacing) - vector determined by simple adaptive process to
give best Q
without equalisation
with equalisation
Q 8.3
Q 12.7
13Why 2.5dB penalty?
- Early experiment showed sensitivity 2.5dB from
theoretical minimum - Penalty contributions are
- single element phase modulator was used instead
of MZ modulator driven through 2Vp some wasted
energy in quadrature component - SOP of LO did not exactly match signal
- shape of transmit pulses not adjusted for zero
intersymbol interference - Receiver noise did not contribute to 2.5dB
penalty - By fixing contributors above it should be
possible to demonstrate near theoretical
sensitivity using sampled coherent detection - with or without propagation impairments
- Combined with appropriate FEC, Shannon limits
should be achievable
14Experiment results CD equalisation
- Chromatic dispersion compensation applied by
simple convolution with vector (FIR filter) - vector is impulse response of CD transfer
function for 89km NDSF, truncated to 7 elements - Penalty from chromatic dispersion is reduced to
zero
FIR filter
89km NDSF with CD equalisation
back-to-back
89km NDSF
real
imaginary
15Experiment results CD equalisation
89km NDSF without CD equalisation OSNR 27dB Q
5.3
89km NDSF with CD equalisation OSNR 27dB Q
12.3
162500km PM-QPSK transmission experiment
- 4x10.7Gb/s polarisation multiplexed QPSK signal
transmitted over 2480km NDSF - Polarisation diverse (and phase diverse) coherent
receiver - Polarisation demultiplexing performed in digital
domain, as well as phase estimation impairment
compensation
Launch power-5dBm
1554.94 nm ??2MHz
DSP (applied offline)
x
y
1554.94 nm ??100kHz
172500km PM-QPSK transmission experiment
Ch1
Carrier recovery
Frequency offset
Decision circuitry
CD comp
x
Ch2
Ch3
CD comp
Carrier recovery
Frequency offset
Decision circuitry
y
Ch4
128 tap FIR
13 tap FIR
- PMD compensation performed using four adaptive
FIR filters - cross terms interact between polarisations
- tap coefficients updated using stochastic
gradient constant modulus algorithm no training
sequence - Bit error rate after 2480km 9.5x10-4 (average),
1.6x10-3 (worst quadrature) - 1.5dB penalty compared to back-to-back
18Other published results using sampled coherent
detection
- 2.5b/s/Hz spectral density demonstrated bu U.
Tokyo group (Tsukamoto et al., paper PD29, OFC
2005) - 10Gbaud QPSK, two polarisations muxed, 16GHz
spaced - record information spectral density
19Other published results using sampled coherent
detection
- Real time (not burst mode) coherent receiver
demonstrated by U. Paderborn (Pfau et al., paper
CThC5, COTA 2006) - 400Mbaud QPSK
- 1MHz wide DFB lasers for transmitter LO
- 2.2Gbaud QPSK real time receiver built by Alcatel
Lucent using Atmel A/D converters, Xilinx FPGA
(Leven et al., paper OThK4, OFC 2007)
20Other published results using sampled coherent
detection
- CoreOptics/Siemens/Eindhoven U. systems
experiment (Fludger et al., OFC 2007, paper
PDP22) - 10 WDM channels x 111Gb/s (28Gbaud), 50GHz
spaced, over 2375km NDSF - Alcatel-Lucent systems experiment (Charlet et
al., OFC 2007, paper PDP17) - 40 WDM channels x 40Gb/s (10Gbaud) PM-QPSK,
100GHz spaced, over 4080km - post-detection compensation for 100ps mean PMD
- Nortel systems experiment (Laperle et al., OFC
2007, paper PDP16) - 40 WDM channels x 40Gb/s (10Gbaud) PM-QPSK, 50GHz
spaced, over 3200km NDSF without inline DCF - post-detection compensation of chromatic
dispersion 33ps mean PMD
21How parallel computation architecture impacts DSP
phase estimation
22Parallel DSP architectures
ts
Lts
long delay
- The DSP must operate in parallel because maximum
clock speed lt symbol rate - parallel operation is eqiuvalent to a delay in
computing a result - result n-1 is not available to compute result n
- algorithms employing feedback are compromised
- Phase estimation algorithms typically use
feedback - resolution is to reduce phase noise by employing
low linewidth lasers - DFB lasers and miniature external cavity lasers
may have too large linewidth to use for sampled
coherent detection
23Open loop Wiener phase estimate
arg( )
Wiener filter
exp( )
complex signal
phase estimate
2
( )2
d eif p
ei2f p
q
- Neglecting high order noise terms, applying small
angle approximation - q 2f (additive
noise component) - Estimation theory says that best linear estimate
of f is Wiener filter applied to q
Gaussian noise
quantity we want Gaussian random walk
quantity observed
24Wiener filter responses
Zero lag Wiener filter
Finite lag Wiener filter
- Finite lag Wiener filter is best, because it sees
D symbols ahead in time as well as the past
25Look-ahead computation
- But the Wiener filters involve feedback to
immediately preceding result not allowed! - Either Wiener filter can be written as
- To resolve, apply look-ahead computation so these
relationships refer to a distant past result, L
symbols ago - multiply numerator and denominator by polynomial
- now uses feedback to L symbols ago, at expense of
more feedforward taps
feedback from previous result
L symbols past
26Experiment
pattern generator 1.5 Gb/s
polarisation controllers
real time sampling scope
polarisation beamsplitter
variable attenuator
photo-detectors
1.2nm filter
DFB
MZ modulator (biased at null)
LO DFB
OSA
phase diverse hybrid
EDFAs
var. atten.
ASE source
- DFB lasers used for signal and LO laser
- combined linewidth Dn 48MHz
- Low symbol rate 1.5Gbaud
- tsDn 0.032
- Long measurement burst 1ms duration, contains
1.5x106 symbols - statistically significant number of bit errors
cycle slips seen - Optical SNR -5dB, in 0.5nm resolution bandwidth
27Results of experiment
- Look-ahead computation tested by comparing L 1
case with L 32 case - found to give identical results
- Q-factor of 8.6dB obtained
- Example of estimated phase vs. time
- uses Wiener filter with D 10
28Future possibilities for coherent detection
29Coherent optical add/drop
Ein(t)
Eout(t)
LO in
Eout(t)- Ein(t)
modulation subsystem
input monitor
from DSP
to DSP
laser
- Inserted signal interferes with input signal to
produce desired output signal - Modulation on inserted signal must take into
account optical phase and SOP of input signal - Can be applied as optical add/drop function,
regenerator function - enables add/drop to be implemented with minimal
channel spacing
30Downconversion by analog multiplication
multipliers
sums
photodetectors
I data out
optical signal
Q data out
local oscillator
phase diverse hybrid
sin(Dwt)
cos(Dwt)
phase estimate
DSP
- Symbol rate for digital downconversion operation
limited by availability of wideband A/D
converters, DSP fabric - Analog multiply can use similar technology to tap
weight in tapped delay line - Weight input of multiplier must have bandwidth
maximum offset frequency, e.g. 1GHz - Symbol rate of e.g. 40Gbaud possible using
todays technology - good solution for 100 GigE
31Conclusions
- Coherent detection is the best mode of detection
of optical signals - offers best receiver sensitivity
- ultra-narrow WDM
- compensation of propagation impairments without
residual penalty - Introduction of real time DSP can overcome cost
issues - Sensitivity 2.5dB from theoretical limit
demonstrated at 10Gb/s - Compensation of chromatic dispersion, PMD over
2500km NDSF demonstrated - Phase estimate can be made in a parallel digital
processor with wide linewidth lasers - synchronous phase estimation has been performed
for an optical signal having tsDn 0.032
32Additional slides
33Phase estimation
?
no noise
phase noise only
phase amplitude noise
- Phase is estimated and applied to signal before
making 1/0 decision - Smoothing function is needed to reduce effect of
additive noise and pass actual phase change - Errors in the phase estimate lead to
- increase in number of bit errors
- cycle slip errors, i.e. data inversion in case of
BPSK
34Optimal phase estimate
- Approach to phase estimation problem
- try to calculate optimal phase estimate
- try to implement optimal estimate on a parallel
digital processor - Best possible estimate of phase is maximum a
posteriori (MAP) estimate - joint estimate of phase f(n) and data d(n) that
maximises - r(n) complex signal
- sp2 normalised variance of amplitude noise
- MAP estimate was calculated by applying a per
survivor method to a group of symbols, and
calculating phase by successive Newton's
approximation for each symbol group instance
35Phase unwrapping
- Wiener filter must operate on unwrapped phase, so
argument function must include phase unwrapping - q(n) arg(s(n)) g(n)
- g(n) g(n-1) 2p f( arg(s(n)) - arg(s(n-1)) )
- where f(x) 1 if x lt p
- f(x) 0 if x lt p
- f(x) -1 if x gt -p
- g(n) keeps count of phase cycles
- However g(n) depends on g(n-1) in expression
above not allowed! - Phase unwrapping function can also be recast
using look-ahead computation to depend on result
L symbols ago - more computations needed than original version
- sum function can be calculated in log2(L) steps,
so can always be calculated by processor of
sufficient parallelism
36Phase estimation methods comparison
BPSK
QPSK
- 1dB penalty point at tsDn 0.014
- 1dB penalty point at tsDn 0.0016