Recurrent Optical Spectrum Slicers

• Recurrent optical spectrum slicers are optical front-end processors that employ delayed feedback to realize an infinite impulse response, sharpening spectral features for dispersion compensation.
• They are implemented using programmable photonics or integrated silicon photonics to enable multi-wavelength processing and enhance performance in IM/DD and self-coherent QAM systems.
• Recent experiments show significant BER improvements and power savings, positioning ROSS as an optical accelerator that offloads complex equalization tasks from digital DSP.

Searching arXiv for papers on recurrent optical spectrum slicers and closely related work. I’m checking arXiv records for the cited ROSS papers and adjacent literature to ensure the article reflects the current research landscape. Recurrent optical spectrum slicers (ROSS) are optical front-end processors in which spectrum slicing is made recurrent by delayed feedback, so that optical filtering is no longer a static band-selection operation but an optical-domain infinite impulse response mechanism that reshapes the received spectrum before direct detection. Across recent work, ROSS has been developed as a programmable photonic, multi-wavelength optical pre-processor for dispersion-impaired intensity modulation/direct detection (IM/DD) links, as an integrated silicon-photonic recurrent spectral filter for long-reach PAM-4 reception, and as a self-coherent direct-detection aid for QAM reception with modest linear DSP (Sozos et al., 15 Jul 2025).

1. Emergence and research trajectory

The ROSS line of work first appeared as a neuromorphic photonic architecture in which optical filters placed in a feedback loop operate as recurrent spectral processors rather than as conventional feedforward wavelength slicers. In that formulation, the filter bank and loop define a recurrent neural or reservoir-computing substrate, with photodetection providing the nonlinear readout and linear regression or FFE supplying the task-specific output stage. Later work moved from numerical studies to experimental IM/DD receivers implemented first with a Waveshaper and programmable photonics, then with purpose-built silicon photonics, and finally to multi-wavelength WDM equalization and self-coherent M-QAM reception (Sozos et al., 2022).

Paper	Focus	Reported result
"High Speed Photonic Neuromorphic Computing Using Recurrent Optical Spectrum Slicing Neural Networks" (Sozos et al., 2022)	ROSS-NN concept and numerical evaluation	100+ Gbaud equalization and NARMA-10 benchmarking
"Experimental Investigation of a Recurrent Optical Spectrum Slicing Receiver for Intensity Modulation/Direct Detection systems using Programmable Photonics" (Sozos et al., 2024)	Experimental IM/DD validation	Equalization of 80 km of 64 Gb/s PAM-4 in C-band
"Experimental Analysis of a Self-Coherent M-QAM Receiver by Means of Recurrent Optical Spectrum Slicing and Direct Detection" (Sozos et al., 12 Mar 2025)	Self-coherent QAM reception	32 Gbaud QAM-4/16 over 25 km, 50 km, and 75 km
"Integrated recurrent optical spectral slicer for equalization of 100-km C-band IM/DD transmission" (Teofilovic et al., 15 Jul 2025)	Ad hoc silicon-photonic ROSS	Below-FEC operation over 100 km
"Recurrent Optical Spectrum Slicers as multi-λ processors for WDM optical equalization of IM/DD channels" (Sozos et al., 15 Jul 2025)	Multi-wavelength WDM equalization	Simultaneous equalization of three wavelengths after 75 km

This progression is significant because it shifts ROSS from a general recurrent-photonic computing proposal toward a communications-specific receiver architecture. A plausible implication is that the topic now spans three linked domains: dispersion compensation in IM/DD, programmable or integrated photonic equalization, and low-complexity alternatives to coherent DSP.

2. Recurrent filtering and spectral slicing mechanism

A conventional optical spectrum slicer splits an input spectrum into different frequency bands using a set of optical filters. The recurrent version introduces feedback, so the optical field recirculates after a delay and the current output depends on both the present input and delayed feedback. In the integrated 2025 formulation, this is described explicitly as an optical-domain IIR filter; in the programmable-photonics experiments, the receiver consists of recurrent optical filter nodes followed by separate photodetectors and, optionally, a simple FFE (Teofilovic et al., 15 Jul 2025).

For the experimentally validated IM/DD receiver, the recurrent node transfer function is given by

$$H(f)=\frac{\sqrt{a}\,\sqrt{b}\,H_0(f)}{1-\sqrt{1-ab}\,H_0(f)\,e^{-j(2\pi f \Delta T+\phi)}} ,$$

where $H_0(f)$ is the nonrecurrent base filter response, $a$ and $b$ are the input/output coupler strengths, $\Delta T$ is the loop delay, and $\phi$ is the loop phase shift. The denominator encodes the recurrence: a portion of the output is fed back to the input after delay and phase rotation, producing sharper spectral features than a simple pass-through filter. In the same work, $H_0(f)$ is instantiated as either a recurrent 1st-order Butterworth filter in a Waveshaper implementation or as a recurrent Mach–Zehnder delayed interferometer (MZDI) in the programmable photonic implementation (Sozos et al., 2024).

The physical relevance of this mechanism follows from IM/DD power fading. After fiber dispersion, optical sidebands acquire frequency-dependent phase shifts; under square-law detection, carrier-sideband beating yields periodic spectral nulls in the detected electrical response. ROSS counteracts this by introducing frequency-selective optical preprocessing before detection, so faded spectral components are reweighted before the photodiode imposes irreversible information loss. The 2024 experimental paper further states that multiple recurrent nodes can expose multiple decorrelated spectral “views” of the same signal to the downstream FFE, provided the filters are sufficiently detuned from one another (Sozos et al., 2024).

In the broader ROSS-NN formulation, recurrence is expressed through a filter-in-a-loop node with coupler ratios, loop loss, phase shift, and delay as tunable parameters, while the center frequencies and bandwidths of the filters act as the physical analog of synaptic weights. This interpretation makes ROSS simultaneously a communications equalizer and a recurrent-neuromorphic spectral processor (Sozos et al., 2022).

3. Devices, programmability, and implementation platforms

A central implementation route uses the iPRONICS Smartlight programmable photonic platform. In the multi-wavelength WDM study, each node contains an MZDI and an external feedback loop controlled with a phase shifter. The MZDI is realized using 4 programmable unit cells (PUCs), and the feedback loop uses 2 PUCs. The main path difference is $\Delta T = 45\ \text{ps}$, yielding a free spectral range of $22.2\ \text{GHz}$ and a bandwidth of $11.1\ \text{GHz}$; the feedback-loop length is 1.622 mm and the feedback-loop delay is 22.5 ps. The phase difference between the MZDI arms, $H_0(f)$0, sets the central frequency and the shape of the transfer function, while the external phase shift further modifies the filter response (Sozos et al., 15 Jul 2025).

The 2024 experimental study used the same platform to compare a small loop and a large loop. The small loop uses 2 PUCs in the feedback path, corresponding to about 1.622 mm and 22.5 ps delay, while the large loop uses 14 PUCs, corresponding to about 11.36 mm and 157.5 ps delay. The smaller feedback loop consistently performed better because the larger loop had higher round-trip loss and therefore weaker effective feedback. That work explicitly notes that the Waveshaper and the large-loop Smartlight filters behave like weak-feedback nodes, approximately akin to $H_0(f)$1, whereas stronger recurrence is needed to reproduce the sharper peaks predicted by the recurrent-filter model (Sozos et al., 2024).

A second route is an ad hoc silicon-photonic implementation on CEA-Leti’s 300 mm silicon photonics platform using immersion lithography. The recurrent node is a feedback-coupled asymmetric MZI with thermo-optic tuning. Two feedback delays were fabricated: a short-$H_0(f)$2 design of about 23 ps and a long-$H_0(f)$3 design of about 44 ps; the design description also states a general feedback-loop delay of approximately 19 ps. The measured transfer functions exhibit sharp spectral features with extinction ratio up to 15 dB, and heater tuning changes not only the spectral position but also the spectral shape, allowing different ROSS nodes to be realized with the same device design. Measured losses were about $H_0(f)$4 per coupler, $H_0(f)$5 waveguide propagation loss, and overall insertion loss of 9.8 dB for the long-$H_0(f)$6 device and 10.4 dB for the short-$H_0(f)$7 device; the feedback-path loss was estimated below 0.1 dB, contrasted with greater than 1 dB in the earlier programmable PIC implementation (Teofilovic et al., 15 Jul 2025).

These implementation results clarify a persistent design tradeoff. Programmable photonics offers broad reconfigurability across wavelength grid and link condition, whereas purpose-built integrated devices offer lower loop loss and stronger recurrence. This suggests that future ROSS receivers may be judged less by whether they are programmable or custom and more by the attainable combination of loop loss, insertion loss, and tuning granularity.

4. IM/DD equalization in dispersive links

The most established application of ROSS is optical preprocessing for dispersive IM/DD PAM-4 transmission. In the 2024 experimental validation, pseudorandom 32 Gbaud PAM-4 was transmitted over 50 km and 80 km of single-mode fiber with $H_0(f)$8. After the ROSS stage, the waveform was captured by an 80 GSa/s oscilloscope and processed with a low-complexity symbol-spaced FFE using 25 taps, intentionally chosen to highlight the optical recurrent stage rather than heavy DSP. For the Waveshaper implementation, a single recurrent filter gave limited benefit, while two and three recurrent filters substantially improved BER. At 50 km, a two-node ROSS receiver pushed BER below the soft-decision FEC region; at 80 km it achieved a BER around $H_0(f)$9, and the three-node case improved further to roughly $a$0. In the Smartlight experiments, the small-loop configuration performed best: with two recurrent nodes, BER could be reduced from about 0.11 to below $a$1, and with three nodes performance improved by nearly two orders of magnitude relative to the baseline (Sozos et al., 2024).

The integrated silicon-photonic ROSS of 2025 extended this operating regime. It was tested in a 32-GBd PAM-4 IM/DD C-band link with an AWG at 64 GSa/s, an external cavity laser at 1550.1 nm, launch power of 0.5 dBm, and transmission over 50 km to 100 km of SMF. Two ROSS nodes were emulated sequentially by changing heater voltage and recording different traces, and the synchronized electrical outputs were combined by a 62-tap FFE. At 50 km, both long- and short-$a$2 designs reached roughly $a$3, below the KP4 HD-FEC threshold of $a$4. The best short-$a$5 design kept BER below the 400ZR C-FEC threshold of $a$6 even at 100 km of SMF transmission (Teofilovic et al., 15 Jul 2025).

The significance of these results is not that ROSS eliminates DSP, but that it changes where the equalization burden is carried. The reported architecture is repeatedly described as optical pre-processing before direct detection followed by a comparatively simple FFE. The papers therefore position ROSS as an optical accelerator that reduces reliance on DFE, MLSE, Volterra methods, or other complex post-detection compensation stages rather than as a complete replacement for digital equalization (Sozos et al., 2024).

5. Multi-wavelength WDM processing and self-coherent reception

A defining 2025 extension is the use of ROSS as a multi-$a$7 optical equalizer for IM/DD WDM links. In that work, the same pair of recurrent nodes simultaneously equalized three 32-GBd PAM-4 wavelengths in a 3-channel WDM system after 25 km, 50 km, and 75 km of SMF in the C-band. The transmitter used two Mach–Zehnder modulators driven by a 64-GSa/s AWG; the central channel was at $a$8, the total launched power was 0.8 dBm, and the programmable photonic circuit introduced more than 20 dB insertion loss, so EDFAs were required before and after the circuit. After optical processing, a 200-GHz AWG demultiplexer, three 50-GHz photodetectors, and an 80-GSa/s DSO were used, followed by a 31-tap T-spaced FFE per channel. Over the full tested range, BER remained below $a$9 for all three channels simultaneously up to 75 km; for a 200-GHz grid, all channels achieved BER below $b$0, while direct detection without ROSS was around $b$1 (Sozos et al., 15 Jul 2025).

An important feature of that WDM receiver is wavelength-grid tolerance. The WDM spacing was varied from 193 GHz to 215 GHz, and the same pair of filters still provided good equalization. The paper reports robust performance around two phase settings, roughly 0.5 rad and 3 rad, corresponding to a frequency detuning of about 8.75 GHz between the two filters, and explicitly states that both integer and non-integer multiples of the filter FSR can be handled. This is the basis for describing ROSS as a multi-wavelength optical processor rather than as a single-channel tuned filter (Sozos et al., 15 Jul 2025).

ROSS has also been extended to self-coherent M-QAM reception. In the 2025 QAM study, recurrent optical spectrum slicing is used as a photonic accelerator ahead of direct detection so that the optical front-end assists complex-signal recovery with minimal DSP. The receiver used the iPronics Smartlight programmable photonic processor, 32 Gbaud QAM-4 and QAM-16, a 25 GHz IQ modulator with $b$2 V, and drive swings below $b$3, with the modulator biased at quadrature so the transmitted constellation stayed in the first IQ quadrant. For QAM-4, two nodes and a 2-output FFE were used; for QAM-16, four nodes and a real-valued $b$4 MIMO FFE were used. The experimental demonstrations covered 25 km, 50 km, and 75 km in the C-band, with QAM-4 remaining around $b$5 to $b$6 over the tested distances and QAM-16 near $b$7 in back-to-back and acceptable up to 50 km (Sozos et al., 12 Mar 2025).

That study adds several distinctive claims. It states that each radian of $b$8 corresponds to about 3.6 GHz in the experimental configuration, and elsewhere that the MZDI phase difference corresponds to a frequency shift of 11 GHz per $b$9. It also introduces geometric constellation shaping using DBSCAN, with the update rule

$\Delta T$0

and reports tolerance to laser drift such that BER remains almost constant for drifts below 1 GHz and degrades noticeably beyond 3 GHz. Under the stated power-accounting assumptions for 1.6T-class modules, the paper reports total power of 6.22 W for the ROSS architecture, or 5.22 W without TEC, and states this corresponds to more than 40% power improvement relative to the lightweight coherent comparison (Sozos et al., 12 Mar 2025).

6. Neuromorphic interpretation, limitations, and open questions

ROSS retains a strong neuromorphic interpretation. In the ROSS-NN paper, the architecture consists of $\Delta T$1 filter banks each containing $\Delta T$2 filters, with each node implemented as a filter embedded in a delayed feedback loop. The central frequencies, bandwidths, frequency offsets, feedback phase, feedback strength, and delay are treated as the physical analog of trainable or tunable recurrent weights. The paper emphasizes that the architecture is passive, compatible with devices such as micro-ring resonators and MZDIs, and can be read out either digitally or optically. On NARMA-10, $\Delta T$3 and $\Delta T$4 gave 12 physical nodes and achieved $\Delta T$5; with $\Delta T$6, $\Delta T$7, and 200 random fabrication-like instances, the reported result was $\Delta T$8, versus an ideal optimized value of 0.079 (Sozos et al., 2022).

The same paper positioned ROSS-NN as a high-baud-rate equalizer. For 112 Gbaud PAM-4, 16-QAM, and 32-QAM, it reports numerical performance superior to simple spectral slicing, FFE, Volterra nonlinear equalization, and MLSE; in the tested setting, it is described as the only method that keeps BER below the HD-FEC threshold at high chromatic dispersion or long reach. The paper also claims about 60 km reach in the studied O-band IM/DD scenario, about 2 dB better tolerance to Kerr nonlinearity than a linear algorithm, and a complexity reduction corresponding to about a factor of 10 relative to state-of-the-art solutions, with ROSS-NN requiring about 40–100 multiplications where Volterra could require over 2400 for the reported cases (Sozos et al., 2022).

At the same time, the literature repeatedly identifies practical limitations. Programmable-photonics demonstrations incur high insertion loss: more than 20 dB in the WDM IM/DD receiver and about 20.5–22.5 dB minimal overall loss in the self-coherent QAM prototype, necessitating EDFAs and exposing the receiver to amplified noise. Several experiments require manual polarization alignment using polarization controllers. The 100 km integrated demonstration emulates two nodes sequentially rather than using a truly parallel multi-node chip. Thermal tuning is necessary in the silicon-photonic implementation. More generally, the papers state that performance depends on choosing good detuning and operating points, that feedback weakness degrades spectral selectivity, and that some residual electronic equalization is still required (Sozos et al., 15 Jul 2025).

A recurrent misconception is that ROSS is merely a variant of ordinary optical filtering. The cited work argues otherwise: recurrence, delayed interference, and pre-detection operation are precisely what distinguish ROSS from conventional static or feedforward slicers. A second misconception is that ROSS is intended to replace all DSP. The experimental papers consistently describe it instead as an optical accelerator or preprocessor that offloads the hardest part of equalization into photonics while leaving a simple FFE or modest linear DSP in place. Taken together, these studies suggest a research direction centered on lower-loss recurrent photonic hardware, multi-node integration, and continued exploration of grid-tolerant, distance-tolerant, and bandwidth-efficient optical preprocessing for IM/DD and self-coherent links (Teofilovic et al., 15 Jul 2025).