Source-linked AI summary
Nonlinear Interference Mitigation via Deep Neural Networks
Christian Häger, Henry D. Pfister
TL;DR
High-speed fiber systems face nonlinear interference, while conventional digital backpropagation is computationally burdensome. The paper unrolls the split-step Fourier method into a trainable complex-valued deep neural network, yielding learned DBP that significantly reduces complexity compared with conventional implementations.
Problem
Nonlinear interference challenges high-speed fiber-optic communication, and digital backpropagation has a large computational burden.
Method
The paper interprets the SSFM as a blueprint for a complex-valued deep neural network and optimizes its parameters using machine learning.
Results
For a 32-span, 100-km-per-span fiber link, learned DBP significantly reduces complexity compared with conventional DBP implementations.
Takeaways & Limitations
Unrolling the SSFM provides clear network-design choices and a useful initialization for gradient-based optimization while retaining a DBP-based structure.
Takeaways & Limitations
The reported approach is based on simulated data, with future work proposed to optimize parameters using experimental data.
Abstract
from arXiv · showhide
A neural-network-based approach is presented to efficiently implement digital backpropagation (DBP). For a 32x100 km fiber-optic link, the resulting "learned" DBP significantly reduces the complexity compared to conventional DBP implementations.
1. Introduction
The paper frames nonlinear interference as a major challenge in high-speed fiber-optic communication and addresses DBP complexity using deep neural networks designed from the SSFM computation graph.
- Digital backpropagation mitigates nonlinear interference by solving the nonlinear Schrödinger equation with negated fiber parameters during receiver processing.The paper identifies DBP as an established approach whose computational burden motivates complexity reduction.
- The central computational challenge is approximating a partial differential equation solution with as few resources as possible.
- Deep neural networks are applied to nonlinear-interference mitigation by interpreting the unrolled SSFM as a network blueprint.The approach uses the SSFM’s functional structure rather than a standard black-box network design.
2. System model
The system models pulse-shaped symbols propagating through a nonlinear fiber and recovers them by numerically solving the inverse propagation problem; the SSFM trades accuracy against computational complexity.
- 2. System model: The transmitted waveform is formed from constellation symbols shaped by p(t), launched into the fiber, low-pass filtered, and sampled to produce y.The signal model defines x(t) from symbols, pulse shape, and symbol rate before propagation.
- 2.1. Digital backpropagation: Digital backpropagation recovers the transmitted symbol vector by solving the NLSE with negated fiber parameters and applying a digital matched filter.The receiver begins from the sampled observation vector and implements the inverse propagation mapping.
- 2.2. Split-step Fourier method: The SSFM estimates the backpropagated signal after M fiber segments, with smaller step size improving accuracy but increasing the number of computational steps.The unrolled iterations form the computation used to approximate DBP.
- 2.2. Split-step Fourier method: Fig. 1 contrasts conventional DBP through the SSFM with learned DBP obtained by unrolling the SSFM.Before deep-learning parameter optimization, both have the same performance when using the same number of steps or layers per span.
3. Deep neural networks
Deep feed-forward neural networks alternate linear and element-wise nonlinear transformations across layers, and their hierarchical structure parallels the iterative decomposition used by the SSFM.
- A deep neural network maps an input vector through successive affine transformations and nonlinearities across ℓ layers.The layer-specific weights, biases, and nonlinear functions define the network computation.
- Deep neural networks and the SSFM share a functional form in which linear operators alternate with simple element-wise nonlinear operators.This structural correspondence motivates interpreting the SSFM as a neural-network blueprint.
- The SSFM provides a hierarchical sequence of elementary steps that matches the kind of structure often associated with successful deep neural networks.The paper presents this relationship as a reason the SSFM–NN similarity may be functionally meaningful rather than coincidental.
4. Learned digital backpropagation
Learned DBP interprets the SSFM as a complex-valued neural-network blueprint and optimizes its parameters with machine learning. Its architecture incorporates span processing, differentiable nonlinearities, constrained convolutional weights, and training against transmitted symbols.
- 4. Learned digital backpropagation: LDBP interprets the SSFM as a blueprint for a complex-valued deep neural network and optimizes its parameters using machine learning tools.The resulting method is called learned DBP.
- 4.1. Neural network parameters: Each LDBP layer uses two weight matrices and an element-wise differentiable nonlinearity, followed by a linear matched-filter layer.The nonlinear function uses a trainable coefficient, while the matched filter is represented by an additional linear layer.
- 4.1. Neural network parameters: Circular-convolution constraints reduce each weight matrix from n2 free parameters to K + 1 and support time-domain filter implementation.The restriction uses symmetric filters of length 2K + 1 with K much smaller than n.
- 4.1. Neural network parameters: Before optimization, LDBP is initialized to match conventional DBP with the same number of steps per span, assuming sufficiently large K.The initialization uses SSFM linear operators, nonlinear coefficients, and the matched filter.
- 4.2. Objective function and optimization procedure: LDBP is trained from received vectors y and transmitted symbol vectors x using mean squared error and Adam-based stochastic gradient descent.Optimization uses mini-batches of 30 input–output pairs.
5. Numerical results
The numerical-results section specifies the simulated 16-QAM, 20-Gbaud optical-link setup and receiver-processing parameters used for evaluation.
- 5. Numerical results: The evaluation assumes 16-QAM transmission at 20 Gbaud over an optical link with Nsp = 32 and Lsp = 100 km.Root-raised-cosine pulses use a 0.1 roll-off factor.
- 5. Numerical results: Receiver processing applies 35 GHz brick-wall low-pass filtering before 40 GHz sampling, yielding 2 samples per symbol.
- 5. Numerical results: Forward propagation is simulated with 6 samples per symbol and 50 SSFM steps per span, with higher values reported as having no effect on results.
6. Conclusion and future work
The paper concludes that an SSFM-unrolled deep-neural-network approach can implement learned DBP with substantially lower complexity than conventional DBP. Future work targets experimental-data optimization and modified nonlinear designs to improve the performance–complexity trade-off.
- 6. Conclusion and future work: The proposed SSFM-unrolled network provides clear hyperparameter choices and a useful initialization for gradient-based optimization compared with black-box deep neural networks.
- 6. Conclusion and future work: The resulting learned DBP significantly reduces complexity compared with conventional DBP implementations.
- 6. Conclusion and future work: Future work could optimize parameters using experimental data and explore modified nonlinear functions with additional filtering to improve the performance–complexity trade-off.