Multi-Rate Kalman Filter for Carrier Phase

Recovery in 200 Gbps PDM Coherent Optical

Receivers
Wrivu Sanyal Srishti Sharma Pradeep Kumar Krishnamurthy
Department of Electrical Engineering Department of Electrical Engineering Department of Electrical Engineering
Indian Institute of Technology Kanpur Indian Institute of Technology Delhi Indian Institute of Technology Kanpur
Kanpur, 208016, India New Delhi, 110016, India Kanpur, 208016, India
Email: [email protected] Email: [email protected] Email: [email protected]

Abstract—The Kalman filter is often used for tracking and and computational capabilities. We verify the performance of
estimation of effects such as LPN and NLPN in long haul coher- the proposed filter by carrying out simulations of long haul
ent optical communication systems. However, real-time symbol- transmission with various laser linewidths and launch powers.
by-symbol estimation of these parameters is computationally
challenging. We use a multi-rate Kalman filtering scheme that II. S YSTEM D ESCRIPTION AND M ULTI - RATE K ALMAN
allows for different sampling and state update rates in the
system. This scheme achieves high Q-factor by making use of F ILTER
maximum available samples while reducing computational load. A. System Model
Simulations are performed for 200 Gbps PDM-16-QAM system
by transmitting 20000 symbols over 800 km optical channel. We use the system setup shown in Fig. 1 for a 200
The filter has Q-factor of 17.25 dB with state estimates being Gbps PDM-16-QAM coherent optical communication link. We
updated after every 20 samples. The filter shows more than 1 generate two sequences of random symbols and map them
dB improvement in Q-factor when compared to a KF where the onto square 16-QAM constellations. We pass them to two
intermediate samples are not utilised for phase estimation.
I/Q modulators, which modulate them onto optical carriers of
I. I NTRODUCTION wavelength 1550 nm. We take the modulating laser linewidth
to be δν. We combine the x̂ and ŷ polarization optical signals
A linear Kalman filter (KF) can be used as a linear minimum
using a polarization beam combiner (PBC) and transmit it
mean squared error (LMMSE) estimator for carrier phase
over an optical channel. The channel consists of Ns spans of
in high data rate coherent optical communication systems
standard single mode fiber (SSMF). Inline optical amplifiers
[1] [2]. However, performing a real-time symbol-by-symbol
mitigate the attenuation caused by each fiber span, but they
estimation of the phase at such data rates is very difficult
add amplified spontaneous emission (ASE) noise to the signal.
computationally [3]. In [4], a KF was proposed that performed
phase estimations using a reduced sampling rate for a polariza-
tion division multiplexed (PDM)-16-QAM system. This led to
a reduction in the computational complexity while maintaining
satisfactory Q-factor performance. However, the filter in [4]
did not use the intermediate samples in phase estimation.
A review of available literature provides promising schemes
for a multi-rate Kalman filter to be used as a phase estimator
in coherent receivers. In [5] such a scheme was proposed
by the authors that accounted for different rates present in
the system. This is practically true as high-speed analog-to-
digital converters (ADCs) available for optical transceivers are
capable of sampling at rates much greater than what the current
computational hardware can cope up with. [6].
In this paper, we modify and use the scheme proposed in
[5] for tracking laser phase noise (LPN) and nonlinear phase
Fig. 1. System model for 200 Gbps PDM-16-QAM coherent optical com-
noise (NLPN) in a 200 Gbps PDM-16-QAM long haul optical munication system. LD: Laser diode; PBS: Polarization beam splitter; PBC:
communication link. The novelty of this scheme is the fact Polarization beam combiner; SSMF: Standard single mode fiber; OA: Optical
that it can be adjusted according to the available sampling amplifier; PMD: Polarization mode dispersion; CD: Chromatic dispersion;
ADC: Analog to digital converter; DSP: Digital signal processing; MrKF:
978-1-6654-4177- 3/21/$31.00 ©2021 IEEE Multi-rate Kalman filter.
 We receive the optical signal by a coherent receiver and pass The system matrices are given by:
it to the digital signal processing (DSP) blocks. We assume  
the receiver local oscillator (LO) to have zero frequency offset A= 1 (3)
 
with the modulating laser and perfect timing synchronisation. B = 1 1 . . . n terms (4)
The k th received symbol on either polarization, after chromatic 
C = 1 1 . . . n1 terms
T
(5)
dispersion (CD) and polarization mode dispersion (PMD)  
0 0 ... 0
compensation, is given by:
Da 0 . . . 0
rk = sk ejΨk + nk
 
(1)  a
D Da .. 
D= . 0 
(6)
√  ..

where sk = Pin dk is the k th transmitted symbol. Ψk is .. 
 . n1 rows .  
the uncompensated phase due to the laser phase noise (LPN) 
..
and nonlinear phase noise (NLPN). NLPN results due to the Da . Da 0
interaction of fiber Kerr nonlinearity with ASE noise generated  
by optical amplifiers [13]. Here nk is the ASE noise added where Da = 1 1 . . . m terms
to the signal, modelled as zero mean additive white Gaussian The Kalman filter gives an LMMSE estimate of (2). Equa-
noise (AWGN) [7]. tions for the filter are as given in [5]. They are stated here for
The LPN is modelled as a discrete-time Wiener process with completeness.
ΨPk
N
= ΨP N PN
k−1 +δΨk , where δΨk
PN
is a zero mean Gaussian Ψ̂pk+1 = P Ψ̂ck + BDT (R∗ )−1 Zk (7a)
random variable with variance 2π δν/R (R is the symbol rate)
Σpk+1 = P Σck (P )T + B(Q − DT (R∗ )−1 D)BT (7b)
[8]. We employ the multi-rate Kalman filter described below
to estimate and mitigate this uncompensated phase Ψk . Kk+1 = Σpk+1 CT (CΣpk+1 CT + R∗ )−1 (7c)
Ψ̂ck+1 = (I − Kk+1 C)Ψ̂pk+1 + Kk+1 Zk+1 (7d)
B. Multi-rate Kalman Filter
Σck+1 = (I − Kk+1 C)Σpk+1 (7e)
The multi-rate Kalman filter proposed in [5] allows for
different rates to be present in the system being estimated. where P = (A − BDT (R∗ )−1 C). Σ is the error covariance
The rates considered in our system are as follows: matrix and Kk is the Kalman gain matrix of necessary
• h0 = 1 is the original state update rate of the system. dimensions. We tune Q and R∗ manually to give optimal
• h1 = m × h0 = m is the rate of sampling of measure- filter performance. ()p represents apriori estimates and ()c
ments by the filter. represents corrected estimates.
• h2 = n × h0 = n is the rate of update of state estimates To better understand the filter functionality, we state explic-
by the filter. itly the equations for KF for an example case of measurement
We consider a case where the rate of update (n) is slower sampling rate (m) = 1 and estimate update rate (n) = 5.
than the measurement sampling rate (m). This allows to These values should be selected based on the available hard-
reduce the computational load, but utilise more samples and ware and computational resources. Suitable values of Q and
hence give better estimates. The ratio m n
= n1 must be an R∗ are obtained by manually tuning the filter. The values are
integer value. put into (7a). The prediction stage of the filter becomes:
For estimation of the phase Ψk in (1), we use the following
 
zk
state space model [5]:
 zk+1 
 
Ψ̂pk+5 = 0.8726 Ψ̂ck + 0 0.0125 0.025 0.038 0.052 

Ψk+1 = AΨk + BWk z
 k+2 

(2) zk+3 
Zk = CΨk + V∗ k
zk+4
where
The Kalman gain matrix settles to:
Ψk = Ψ(kn),
Kk+5 = [0.1722 0.1633 0.1567 0.1524 0.1503]
Zk = col {z(kn + j), j = 0, m, 2m . . . (n − m)},
Wk = col {w(kn + i), i = 0, 1, 2 . . . (n − 1)}, These weights are used in the correction step of KF given
by (7d), which becomes:
V∗ k = Vk + D Wk ,
zk+5 − Ψ̂pk+5
 
Vk = col {v(kn + j), i = 0, m, 2m . . . (n − m)}. p
 k+6 − Ψ̂k+5 
z 
Here zk = ∠rk s∗k is the measurement variable. We remove c p p
Ψ̂k+5 = Ψ̂k+5 + Kk+5 zk+7 − Ψ̂k+5 
 
the modulation by multiplying with the conjugate, allowing zk+8 − Ψ̂pk+5 
 
us to use linear KF equations. wk and vk are the zero mean
zk+9 − Ψ̂pk+5
process and measurement noise samples respectively. Multi-
rate properties of the system lead to correlation between the Thus, the filter uses a weighted sum of intermediate samples
noises, via matrix D. in prediction step and smoothing of state estimate using
the next available samples in correction step. Since more
measurement samples are being used by the filter to generate
each state estimate, we expect superior Q-factor performance
by this scheme.
III. S IMULATION S ETUP AND R ESULTS
We verify the performance of the filter by considering
carrier phase recovery of 200 Gbps PDM-16-QAM coherent
optical link shown in Fig. 1. We use MATLAB to perform
the simulations. To simulate fiber propagation, we use the
Optilux toolbox [9]. We neglect carrier frequency offset in (a) (b)
these simulations. We generate two data streams of 20000 Fig. 2. Constellation diagrams of received signal for 20000 symbols, Pin =
random symbols and propagate them through optical channel. 2 dBm, Ns =10 spans, δν = 100 kHz. (a) Before MrKF, (b) After MrKF with
The channel consists of Ns =10 spans of SSMF links (total m (sampling rate) = 1 and n (estimate update rate) = 10.
length = 800 km). Each optical amplifier provides a gain
of 16 dB and noise figure of 5 dB. For fiber propagation,
we numerically solve vector form of nonlinear Schrödinger
equation (NLSE) using split-step Fourier method (SSFM). We
also include the effects of polarization mode dispersion during
propagation of the signal through the fiber. The parameters for
fiber propagation are given in Table I.

TABLE I
PARAMETERS FOR SSMF

Attenuation coefficient α = 0.2 dB/km

Dispersion coefficient Ds = 17 ps/nm-km
Kerr nonlinear coefficient γ = 1.7 /W/km
√ Fig. 3. Trajectory of tracked phase noise for 20000 symbols, Pin = 2 dBm,
PMD coefficient Dp = 0.1 ps/ km
Ns =10 spans, δν = 100 kHz, using MrKF with m (sampling rate) = 1 and
Length of span L = 80 km
n (estimate update rate) = 10.

After propagation through fiber, we receive the signal by

a coherent receiver and pass it to the DSP block. In the comparison. We see that the multi-rate KF scheme gives better
DSP chain, we use vector form of digital back-propagation Q-factor performances as a result of using the intermediate
(DBP) to jointly equalise CD and PMD [10]. The dispersion samples available to it. At 1 dBm launch power, the MrKF
compensated signal is in the form as shown in (1). We with n = 10 outperforms the downsampled KF by a factor
pass this signal to the multi-rate Kalman filter for estimation of 1.1 dB while having similar computational complexity.
and tracking of phase impairments. We use the error vector This improvement in Q-factor is because the filter utilises 10
magnitude (EVM) method to compute Q-factor and evaluate times more samples than the downsampled KF. The MrKF’s
the performance of the system [11], [12]. We calculate the utilisation of as many available samples as possible makes
Q-factor by the following equations:
q√
3
BER = erf c 2/10(EV M )2 (8)
8 √
2 erf c−1 (2BER)


Q(dB) = 20 log10 (9)
Fig. 2a shows the constellation of the received signal after
PMD and CD compensation. We use a multi-rate Kalman filter,
with m = 1 and n = 10 to track and mitigate the phase
impairments. Fig. 2b shows the constellation after processing
by the filter. Fig. 3 shows trajectory of the phase noise tracked
by the filter. As seen from Fig. 2b, the filter has estimated and
compensated LPN and NLPN so that the phase error in the
received constellations are corrected.
Fig. 4 shows the variation of Q-factor over launch powers
for different estimate update rates of the filter. The Q-factor
performance of KF with reduced sampling rates (downsampled
by a factor of 10), which does not use any intermediate Fig. 4. Q-factor vs. Launch Power for different MrKF schemes, for δν = 100
samples (as proposed in [4]), is used as a benchmark for kHz and Ns = 10 spans.
it a suitable choice for phase estimation in optical receivers. [2] L. Barletta, M. Magarini and A. Spalvieri, “Bridging the Gap Between
Interestingly, the MrKF with n = 20 also outperforms the Kalman Filter and Wiener Filter in Carrier Phase Tracking,” in IEEE
Photonics Technology Letters, vol. 25, no. 11, pp. 1035-1038, June1,
downsampled KF by a factor of 0.92 dB, even though it 2013, doi: 10.1109/LPT.2013.2259476.
calculates half as many state estimates. Also noticeable is the [3] X. Zhou, “Efficient Clock and Carrier Recovery Algorithms for Single-
fact that at lower launch powers, the MrKF scheme performs Carrier Coherent Optical Systems: A systematic review on challenges
and recent progress,” in IEEE Signal Processing Magazine, vol. 31, no.
almost equal to a full rate Kalman filter. The degradation in 2, pp. 35-45, March 2014, doi: 10.1109/MSP.2013.2281071.
Q-factor is less than 0.2 dB at Pin = −5 dBm for n = 20. [4] S. Sharma and P. K. Krishnamurthy, “Reduced sampling rate Kalman
filters for carrier phase and frequency offset tracking in 200 Gbps 16
QAM coherent communication system,” Scientific Reports, vol. 11, no.
1, pp. 1-12, 2021.
[5] Y. Liang, T. Chen, and Q. Pan, “Multi-rate optimal state estimation,”
International Journal of Control, vol. 82, no. 11, pp. 2059–2076, 2009.
[6] C. Laperle and M. Osullivan, “Advances in High-Speed DACs, ADCs,
and DSP for Optical Coherent Transceivers,” Journal of Lightwave
Technology, vol. 32, no. 4, pp. 629–643, 2014.
[7] A. Jain and P. K. Krishnamurthy, “Phase Noise Tracking and Com-
pensation in Coherent Optical Systems Using Kalman Filter,” in IEEE
Communications Letters, vol. 20, no. 6, pp. 1072-1075, June 2016, doi:
10.1109/LCOMM.2016.2550429.
[8] M. Seimetz, High-order modulation for optical fiber transmission,
Berlin: Springer, 2009.
[9] P. Serena, M. Bertolini and A. Vannucci, Optilux toolbox. Available at
optilux.sourceforge.net/Documentation/optilux-doc.pdf, 2009.
[10] E. Ip and J. M. Kahn, “Compensation of Dispersion and Nonlinear
Impairments Using Digital Backpropagation,” Journal of Lightwave
Technology, vol. 26, no. 20, pp. 3416–3425, 2008.
Fig. 5. Q-factor vs. linewidth for different MrKF schemes, with Pin = 1 [11] F. Zhang et al., “Experimental Comparison of Different BER Estimation
dBm and Ns =10 spans. Methods for Coherent Optical QPSK Transmission Systems,” in IEEE
Photonics Technology Letters, vol. 23, no. 18, pp. 1343-1345, Sept.15,
2011, doi: 10.1109/LPT.2011.2160718
Finally, we test the filter for various laser linewidths and [12] R. Schmogrow et al., “Error Vector Magnitude as a Performance
transmission spans. Fig. 5 shows variation of Q-factor with Measure for Advanced Modulation Formats,” in IEEE Photonics
δνTs . The filter is tested with δν from 50 kHz up to 3 MHz. Technology Letters, vol. 24, no. 1, pp. 61-63, Jan.1, 2012, doi:
10.1109/LPT.2011.2172405.
Simulations show a degradation of ≈ 0.5 dB between δν [13] J. P. Gordon and L. F. Mollenauer, “Phase noise in photonic communi-
= 50 kHz and δν = 1 MHz for MrKF with n = 10. This cations systems using linear amplifiers,” Optics Letters, vol. 15, no. 23,
result demonstrates a high tolerance to laser linewidth for the p. 1351, 1990.
proposed scheme. For an increase in the number of spans to
14 (total length of channel 1120 km), the MrKF with n = 20
showed a Q-factor of about 15.9 dB. Thus, the filter can be
used for long haul communication links of over 1000 km.
IV. C ONCLUSION
We show that the multi-rate Kalman filter scheme suc-
cessfully estimates the LPN and NLPN present in a PDM-
16-QAM system with laser linewidth of up to 1 MHz. By
using the intermediate samples available, and by virtue of the
smoothing of estimates based on future samples, the multi-
rate Kalman filter gives 1.1 dB Q-factor improvement over a
downsampled Kalman filter, where the intermediate samples
were discarded. We see the filter performs especially well
at low launch powers. At launch powers less than -3 dBm,
we obtain a Q-factor within 0.2 dB of a full rate Kalman
filter, even when estimating the state after every 20 samples.
The flexibility in terms of the sampling rates and estimate
update rate makes this scheme useful as ADCs and ASICs
used in optical receivers usually differ in the speeds they can
achieve. We intend to extend the work further to include CFO
estimation by modifying the linear KF scheme to a vector KF.
R EFERENCES
[1] T. Inoue and S. Namiki, “Carrier recovery for M-QAM signals based
on a block estimation process with Kalman filter,” Optics Express, vol.
22, no. 13, p. 15376, 2014.