Complex-Valued Kernel-based Phase and Amplitude Distortion

Compensation in Parametrically Amplified Optical Links

Long H. Nguyen(1) , Sonia Boscolo(1) , Stylianos Sygletos(1)

(1)
Aston Institute of Photonic Technologies, Aston University, Birmingham B4 7ET, United Kingdom
[email protected]

Abstract We develop a complex-valued kernel-adaptive-filtering based method for phase and amplitude
distortion compensation in cascaded fibre-optical parametric amplifier (FOPA) links. Our algorithm
predicts and cancels both distortions induced by pump-phase modulation across all amplification stages,
achieving more than an order of magnitude improvement in BER. ©2024 The Author(s)

Introduction version of the real-valued kernel recursive least-

arXiv:2406.16239v1 [eess.SP] 23 Jun 2024

squares (KRLS) adaptive algorithm introduced in[9]

Fibre-optical parametric amplifiers (FOPAs) offer for mitigating the dithering-induced phase distor-
unique properties that make them attractive for fu- tion in OPC systems following conventional carrier
ture optical communications, including their ability phase recovery (CPR)[10] .
to operate at nearly any frequency with significant In this paper, we extend the operation of the
gain bandwidth, ultra-fast response, and capabili- SWKRLS algorithm to the complex domain to ad-
ties for wavelength conversion and optical phase dress also the related amplitude distortions in cas-
conjugation (OPC)[1] . However, stimulated Bril- caded FOPA links. Our complex-valued (CV) algo-
louin scattering (SBS) poses a major challenge to rithm functions as a single-stage online equaliser
integrating FOPAs into optical systems[2] as it limits handling both amplitude and phase distortions
the pump power deliverable to the highly nonlinear from pump dithering and random phase noise in-
fibre (HNLF), thereby restricting the achievable duced from laser sources, without needing prior
signal gain. To mitigate SBS, a common strat- knowledge of the dithering frequencies unlike the
egy is to broaden the pump source’s linewidth method demonstrated in[7] . The performance of
via external phase modulation, e.g. using a set the equaliser is verified numerically in 28-Gbaud
of radio-frequency (RF) tones[3] , which effectively single-polarisation 16 quadrature-amplitude mod-
suppresses SBS and enables higher gains[4] . How- ulation (QAM) signal transmission over a cascade
ever, this modulation can introduce temporal fluc- of ten FOPAs, achieving substantial bit-error-rate
tuations in the complex parametric gain, impacting (BER) improvement over conventional CPR.
coherently-detected complex amplitude signals
Recent research has addressed SBS mitigation Method
at the subsystem and device level, showing promis- Our proposed scheme is illustrated in Fig. 1(a)
ing results but often resulting in more complex de- while Fig. 1(b-c) displays the complex gain re-
vice architectures which could hinder commercial- sponse of each FOPA unit and the corresponding
ization due to increased costs[5] . Conversely, inte- root-mean-square (RMS) fluctuations induced by
grating machine learning (ML) into DSP has shown dithering as a function of the signal detuning from
potential for enhancing equalization functionality in the pump wavelength. At the point of maximum
commercial systems. Specifically, ML-enhanced gain, the signal primarily experiences phase distor-
DSP tailored for FOPAs could offer a viable solu- tion, while smaller wavelength detunings introduce
tion for their broader adoption in communication also amplitude distortion.
systems. Our recent efforts have developed DSP The accumulated impact of the aforementioned
techniques that effectively mitigate phase distor- distortions appears as an unknown complex fac-
tion from pump dithering in FOPA devices[6],[7] . tor on the received signal rn , where n represents
Kernel-based online learning algorithms, often the symbol index. At the core of our compen-
referred to as kernel adaptive filters, leverage sation scheme is the complex valued SWKRLS
the theory of reproducing kernel Hilbert spaces (CV-SWKRLS) algorithm that creates an estimate
(RKHSs) for nonlinear prediction in real-time appli- of this distortion, denoted as dˆn , subsequently
cations. These methods transform inputs to a high- used for the compensation purpose. The corrected
dimensional feature space through a positive def- symbol zn = rn /dˆn is then processed through
inite kernel function, allowing nonlinear optimiza- a decision circuit, which outputs the decision-
tion problems in the input space to be treated as directed symbol [zn ]D . The complex ratio of the
convex optimization problems in RKHS[8] . In[6] , we received symbol to the decision-directed symbol
have successfully applied the sliding-window (SW) d˜n = rn /[zn ]D , represents a decision-driven dis-
 Fig. 1: (a) Block diagram of the proposed CV SWKRLS-based equaliser for nth symbol recovery. (b-c) Amplitude and phase
spectral responses of the FOPA’s gain versus wavelength detuning from the pump, and respective dithering-induced RMS
fluctuations. The green vertical line indicates the operating point for the performance evaluation of the proposed algorithm.

tortion, which has critical role in updating our al- equation requires recalculation of the inverse ker-
gorithm and maintaining its online operation. A nel matrix at every symbol update and it is very
set of reference symbols is utilised for the initial computationally expensive and memory-intensive.
training of the model until its convergence before Therefore, instead of explicitly computing Kn and
the algorithm enters a decision-directed mode. then K−1 n every time, these matrices can be up-
A way to operate kernel algorithms in the com- dated recursively by leveraging the matrices from
plex domain is through "complexification" of real the previous cycle, Kn−1 and K−1 [9]
n−1 . Specifi-
RKHS[11] . In this case, the input complex data cally, the new regularised kernel matrix Kn is con-
x = xr + jxi , where xr , xi ∈ RL is mapped structed from Kn−1 , by removing its first row and
to a "complexified" RKHS according to the rule column, now referred to as K̄n−1 , and then adding
Φ̂(x) = Φ(xr , xi ) + jΦ(xr , xi ), where Φ(xr , xi ) = kernels calculated with the new data as the last
κ(·, (xr , xi )) is the feature map of a chosen row and column:
real kernel, whichPin our case is Gaussian, i.e.  
2L
κ(p, q) := exp[− k=1 (pk − qk )2 /(2σ 2 )], where K̄n−1 bn
Kn = , (1)
2L
p, q ∈ R and the hyperparameter σ defines the bTn cn
kernel width. In the regression task between the in-
where bn = [κ(x̂n−M +1 , x̂n ), ..., κ(x̂n−1 , x̂n )]T ,
put X̂ ∈ RM ×2L (consisting of real and imaginary
and cn = κ(x̂n , x̂n ) + λ. The inverse kernel matrix
parts of the complex data points) and the output
K−1
n can then be updated accordingly as
y ∈ CM where M is the number of observations,
the algorithm aims to achieve minα∈CM ||y−Kα||2 ,  −1
K̄n−1 − K̄−1 bn eT en

−1 n−1 n
where K is the kernel matrix calculated on the in- Kn = , (2)
eT
n fn
put observations, i.e., K(p, q) = κ(X̂p , X̂q ), with
X̂p and X̂q are pth and qth rows of X̂ respectively. −1 −1
where fn = (cn − bT n K̄n−1 bn ) and en =
The coefficient vector α can be solved directly but −1
−K̄n−1 bn fn . After calculating the updated solu-
in the online scenario, a recursive update through tion αn , our estimation algorithm, in the next cycle,
a SW approach will reduce the complexity. can predict distortion for the (n + 1)th symbol as
The online SWKRLS algorithm will make predic- dˆn+1 = hT n+1 αn by applying the learned αn vec-
tion of the one-step ahead complex distortion dˆn tor on the kernel vector evaluated on the points
based on a vector of past L decision-driven distor- within the dictionary X̂n and the new point x̂n+1 ,
tions xn = [d˜n−1 , ..., d˜n−L ], which at the input of i.e., hn+1 = [κ(x̂n−M +1 , x̂n+1 ), . . . , κ(x̂n , x̂n+1 )]T .
the algorithm is converted at to a real valued com-
posite representation x̂n = [Re(xn ), Im(xn )] ∈ Results and discussion
R2L , to facilitate calculation with real Gaussian We considered a system comprising ten identi-
kernels, and the learning coefficient vector αn−1 cal cascaded stages, each incorporating linear
carried from the previous update. After obtaining loss followed by a FOPA. We modelled the FOPA
the new decision-driven distortion d˜n from the re- using the complex signal gain[1] with the phase
cently predicted dˆn , in the CV SWKRLS algorithm, mismatch comprised an additional instantaneous
we only consider the last M input-output pairs term induced by the phase modulation of the
as observations and form the observation input pump[12] . Our pump-phase modulation employed
matrix X̂n = [x̂n , ..., x̂n−M +1 ]T and the observa- a four-tone scheme with a base frequency of
tion output vector yn = [d˜n , ..., d˜n−M +1 ]T . The 100 MHz and a multiple of three spacing between
regularised real kernel matrix can be defined as successive tones, i.e., the set of RF tones was
Kn = X̂n X̂T n + λI (I is the identity matrix and λ is [0.1, 0.3, 0.9, 2.7] GHz. The tone’s amplitudes and
a regularisation constant). The updated solution phases were optimised to ensure nearly uniform
is then obtained as αn = K−1 n yn . However, this power distribution among the peaks across the
 Fig. 3: Amplitude and phase RMS variations over number of
cascaded FOPA stages between two methods.

We compared the performance of our proposed

kernel-based compensation scheme across the
FOPA cascade with conventional CPR accom-
plished by the one-tap least mean-squares (LMS)
algorithm[10] , as shown in Fig. 2(b). Being de-
signed to operate on the slower laser phase noise,
the LMS-CPR scheme is unable to handle the
high-frequency, dithering-induced phase distor-
tions effectively, leading to phase cycle slips and in-
Fig. 2: (a) Hyyper-parameter optimisation with input block creased phase detection errors. The random time
length L = 20. The blue circled point represents the selection
for the subsequent study. (b) BER performance comparison
shift at each stage leads to the different phase and
over number of cascaded FOPA stages. amplitude RMS fluctuations of the received signal
in different runs, and when the LMS method fails
broadened pump spectrum. This optimisation was to predict it, its BER values are unstable as in Fig.
performed using stochastic gradient descent in 2(b) despite of high number of symbol batches.
TensorFlow[7] . The optimised pump-phase modu- In contrast, our kernel-based scheme significantly
lation significantly raised the SBS power threshold, outperforms the LMS method, achieving a BER
allowing for SBS-free operation of the FOPA at improvement of at least one order of magnitude
its peak power gain of 25 dB[7] . To illustrate the with a consistent performance. The capability of
equalisation capability of our kernel method, the the CV SWKRLS algorithm to mitigate both phase
FOPA was operated at a signal detuning from the and amplitude distortions accrued along the FOPA
pump of |λs − λp | = 20 nm (see Fig. 1(b-c)). cascade is illustrated by Fig. 3. The figure shows
We performed numerical simulations of the the evolution over the transmission length of the
transmission of a single-polarisation 28-Gbaud RMS deviations of the phases and amplitudes, re-
16-QAM Nyquist shaped signal with a roll-off fac- spectively, of the received symbols compared to
tor of 0.1. The laser line-widths were 50 kHz and the transmitted ones, after post-processing by the
30 kHz for the transmitter and receiver units and one-tap LMS algorithm and our method. The RMS
the FOPA pumps, respectively. To avoid the signal values displayed are averaged over the alphabet
symbols experiencing exactly the same phase dis- size. While the suppression of phase distortion
tortion along the FOPA link, we included a random was notably effective, amplitude distortion suppres-
time shift in the pump-phase modulation sinusoidal sion was also achieved. These results highlight
waveform at each FOPA stage. The amplifier’s the superior distortion correction capabilities of our
noise figure was 4.5 dB. We employed the directly- method compared to conventional CPR.
counted bit-error-rate (BER) as a system’s perfor-
mance metric measured over 100 × 216 symbols. Conclusions
The kernel and the training algorithm were opti- We developed an kernel-based algorithm for
mised by carrying out a grid search over the hyper- phase and amplitude compensation in systems
parameter space of dictionary size M , kernel width with multiple cascaded FOPAs. Our scheme can
σ and regularisation parameter λ with 50 × 216 outperform conventional CPR schemes, which typ-
symbols generated after 10 FOPA stages. The ically react slowly to temporal phase variations.
block length was L = 20. We can see from Fig. The efficacy of our method has been demonstrated
2(a) that the BER is minimised by the selection of in 16-QAM signal transmission, achieving a BER
hyper-parameters: (M = 50, σ = 100.5 , λ = 10−1 ) improvement of at least an order of magnitude
(marked by a blue circle). We then used these over the one-tap LMS phase recovery algorithm
values for the performance evaluation. across a cascade of ten FOPAs.
Acknowledgements
This work was supported by the H2020
MSCA ETN project POST-DIGITAL (EC 263 GA
860360), and the UK EPSRC grants TRANSNET
(EP/R035342/1) and CREATE (EP/X019241/1).
References
[1] M. E. Marhic, Fiber Optical Parametric Amplifiers, Os-
cillators and Related Devices. Cambridge, United King-
dom: Cambridge University Press, 2007. DOI: 10.1017/
CBO9780511600265.
[2] V. Gordienko, Á. D. Szabó, M. F. C. Stephens, V. Vas-
siliev, C. B. Gaur, and N. J. Doran, “Limits of broadband
fiber optic parametric devices due to stimulated Brillouin
scattering”, Optical Fiber Technology, vol. 66, p. 102 646,
2021. DOI: 10.1016/j.yofte.2021.102646.
[3] J. B. Coles, B. P.-P. Kuo, N. Alic, et al., “Bandwidth-
efficient phase modulation techniques for stimulated
Brillouin scattering suppression in fiber optic parametric
amplifiers”, Optics Express, vol. 18, no. 17, pp. 18 138–
18 150, 2010. DOI: 10.1364/OE.18.018138.
[4] T. Torounidis, P. A. Andrekson, and B.-E. Olsson, “Fiber-
optical parametric amplifier with 70-dB gain”, IEEE Pho-
tonics Technology Letters, vol. 18, no. 10, pp. 1194–
1196, 2006. DOI: 10.1109/LPT.2006.874714.
[5] M. Bastamova, V. Gordienko, N. J. Doran, and A. D. El-
lis, “Impact of pump phase modulation on QAM signals
in polarization-insensitive fiber optical parametric am-
plifiers”, Optical Fiber Technology, vol. 84, p. 103 758,
2024. DOI: 10.1016/j.yofte.2024.103758.
[6] S. Boscolo, T. T. Nguyen, A. A. I. Ali, S. Sygletos, and
A. D. Ellis, “Kernel adaptive filtering-based phase noise
compensation for pilot-free optical phase conjugated
coherent systems”, Optics Express, vol. 30, no. 11,
pp. 19 479–19 493, 2022. DOI: 10.1364/OE.456963.
[7] L. H. Nguyen, S. Boscolo, and S. Sygletos, “Online
digital compensation of pump dithering induced phase
and amplitude distortions in transmission links with cas-
caded fibre-optical parametric amplifiers”, Optics Ex-
press, vol. 32, no. 8, pp. 13 467–13 477, 2024. DOI:
10.1364/OE.516840.
[8] J. L. Rojo-Álvarez, M. Martínez-Ramón, J. Muñoz-Marí,
and G. Camps-Valls, Digital Signal Processing with Ker-
nel Methods. Hoboken, NJ, USA: Wiley-IEEE Press,
2018. DOI: 10.1002/9781118705810.
[9] S. V. Vaerenbergh, J. Via, and I. Santamaria, “A sliding-
window kernel RLS algorithm and its application to non-
linear channel identification”, in 2006 IEEE International
Conference on Acoustics Speech and Signal Process-
ing Proceedings, Toulouse, France, 2006, pp. V–V. DOI:
10.1109/ICASSP.2006.1661394.
[10] I. Fatadin, D. Ives, and S. J. Savory, “Blind equalization
and carrier phase recovery in a 16-QAM optical coher-
ent system”, Journal of Lightwave Technology, vol. 27,
no. 15, pp. 3042–3049, 2009. DOI: 10.1109/JLT.2009.
2021961.
[11] P. Bouboulis and S. Theodoridis, “Extension of
Wirtinger’s calculus to reproducing kernel Hilbert spaces
and the complex kernel LMS”, IEEE Transactions on
Signal Processing, vol. 59, no. 3, pp. 964–978, 2011.
DOI : 10.1109/TSP.2010.2096420.

[12] A. Mussot, A. Durecu-Legrand, E. Lantz, et al., “Im-

pact of pump phase modulation on the gain of fiber
optical parametric amplifier”, IEEE Photonics Technol-
ogy Letters, vol. 16, no. 5, pp. 1289–1291, 2004. DOI:
10.1109/LPT.2004.826123.