⚡ electrical engineering

On LLR Calculations for Soft-decision Decoding in Next-generation IM-DD Systems with Laser RIN

This paper proposes a low-complexity Log-Likelihood Ratio (LLR) approximation for soft-decision decoding in intensity-modulation direct-detection systems affected by laser relative intensity noise, which achieves no bit error rate penalty compared to optimal decoding and outperforms mismatched additive white Gaussian noise-based approaches.

Original authors: Felipe Villenas, Yunus Can Gültekin, Alex Alvarado

Published 2026-07-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Felipe Villenas, Yunus Can Gültekin, Alex Alvarado

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to send a secret message using a flashlight in a foggy room. The message is made of four different brightness levels (PAM-4), like dim, medium, bright, and super-bright. In a perfect world, the fog (noise) would be the same thickness no matter how bright the light is. But in the real world of next-generation data centers, the laser itself gets a bit jittery. This jitter, called Relative Intensity Noise (RIN), means the fog gets thicker whenever you turn the light up. The brighter the signal, the "fuzzier" the noise becomes.

The paper tackles a tricky problem: How do you decode these messages accurately when the noise changes depending on the signal?

The Wrong Way: The "One-Size-Fits-All" Guess

Traditionally, engineers have used a shortcut. They pretend the fog is uniform (like standard Additive White Gaussian Noise, or AWGN) and calculate a "confidence score" for each bit of data. In the paper's language, this is the mismatched AWGN-based LLR.

The authors ran simulations to see what happens if you use this shortcut in a laser-jittery environment. The result? It's like trying to read a map while wearing glasses that are slightly the wrong prescription. The system works okay at low speeds, but as you push the data rate to 400 Gb/s, 600 Gb/s, and 800 Gb/s, the error rate skyrockets. Specifically, at high signal strengths, using this wrong method caused the error rate to be 3 to 5 times worse than it needed to be. In some cases, it even pushed the system into a state where it couldn't meet the required safety standards (the KP4 threshold) at all. The paper explicitly rules out using this simple, mismatched method for high-speed systems with laser noise; it simply doesn't cut it.

The Perfect Way: The "Exact" Math

The "correct" way to calculate the confidence score (called the Log-Likelihood Ratio, or LLR) involves complex math that accounts for the fact that noise changes with signal strength. The paper shows that for this specific laser noise, the perfect calculation follows a quadratic polynomial relationship.

Think of this as a super-smart decoder that knows exactly how the fog thickens for every brightness level. It's accurate, but it's computationally heavy—like trying to solve a complex physics equation in your head while running a marathon.

The Paper's Solution: The "Smart Shortcut"

Here is the main finding: The authors proposed a low-complexity approximation that is much easier to calculate but works just as well as the perfect math.

They noticed that the most critical part of the decoding happens when the confidence score is near zero (the "zero-crossing" point). They created a piecewise linear approximation (basically, a few straight lines) that hugs the complex curve right where it matters most.

In their simulations, this "smart shortcut" performed virtually no differently than the perfect, complex math.

The Proof: When they tested it, the error rates (Post-FEC BER) were identical to the perfect method.
The Gain: By using this new method instead of the wrong "one-size-fits-all" method, they recovered the lost performance. For 400 Gb/s and 600 Gb/s systems, they could achieve the required error rate with 0.1 dB and 0.4 dB less signal power, respectively. For 800 Gb/s, the new method made reliable communication possible where the old method failed completely.

The Bottom Line

The paper doesn't just suggest this might work; in their simulations, it proved that you can swap the heavy, complex math for a simple, straight-line approximation without losing any accuracy. However, they also note that this is currently based on simulations. They suggest that future work will need to test this with real hardware and even stronger error-correction codes to see if it holds up in the physical world.

So, for the next generation of data centers, the takeaway is: Don't guess the noise; don't overcomplicate the math. Use this new "smart shortcut" to get the laser jitter right, and you can keep your data flowing fast and error-free.

Technical Summary: LLR Calculations for Soft-decision Decoding in Next-generation IM-DD Systems with Laser RIN

Problem Statement
The rapid scaling of data center intraconnects (DCI) toward 400 Gb/s/lane and beyond relies on intensity-modulation and direct-detection (IM-DD) systems using 4-ary pulse amplitude modulation (PAM-4). A critical bottleneck in these systems is relative intensity noise (RIN) originating from the laser source. Unlike additive white Gaussian noise (AWGN), RIN creates signal-dependent noise statistics, where the noise variance scales with the square of the optical signal amplitude.

Soft-decision (SD) decoding, essential for achieving high coding gains with concatenated Forward Error Correction (FEC) schemes (e.g., KP4 outer code + Hamming inner code), requires accurate Log-Likelihood Ratio (LLR) calculations. Standard LLR approximations, which assume a constant noise variance (AWGN model), fail to account for the signal-dependent nature of RIN. The paper identifies that using these mismatched AWGN-based LLRs in RIN-limited channels induces a significant penalty in post-FEC Bit Error Rate (BER), potentially preventing the system from meeting FEC thresholds at high data rates.

Methodology
The authors analyze an IM-DD system model where the channel output $Y$ is defined as $Y = X + Z\sigma(X)$ , with $X$ representing PAM-4 symbols and $\sigma(X)$ representing the signal-dependent noise standard deviation derived from thermal noise and RIN statistics.

Exact LLR Derivation: The paper first establishes the exact LLR expression for the bit position $k$ , which involves a summation of exponential terms weighted by the inverse of the signal-dependent noise variance $\sigma(x)$ . Analysis reveals that these exact LLRs follow a piecewise quadratic relationship with respect to the received value $y$ .
Mismatched Analysis: The study compares the exact LLRs against the standard AWGN-based LLRs (which assume a constant average variance).
Proposed Approximation: To reduce computational complexity while maintaining accuracy, the authors propose a low-complex, piecewise linear approximation. This method utilizes the "zero-crossing" (ZC) approximation around the point where the LLR is zero (the least reliable bits).
- The derivation employs the max-log approximation to simplify the summations within the logarithm.
- For the first bit position ( $k=1$ ), the approximation is a single linear function.
- For the second bit position ( $k=2$ ), the approximation is a piecewise linear function composed of two lines, corresponding to the two zero-crossing points identified in the exact LLR curve.

Key Contributions

Characterization of RIN-Limited LLRs: The paper demonstrates that for PAM-4 IM-DD systems with RIN, the exact LLRs behave as piecewise quadratic functions, distinct from the piecewise linear behavior seen in AWGN channels.
Quantification of Mismatch Penalty: It is shown that neglecting signal-dependent noise by using standard AWGN-based LLRs results in a non-negligible performance degradation in post-FEC BER.
Low-Complexity Approximation: The authors propose a specific piecewise linear approximation (ZC approximation) that is computationally simpler than the exact quadratic calculation.
Performance Validation: Through Monte Carlo simulations, the paper validates that the proposed approximation recovers the performance loss associated with mismatched LLRs, effectively matching the performance of the exact LLR calculation.

Numerical Results
Simulations were conducted for a concatenated KP4 + Hamming FEC scheme across net bitrates of 400, 600, and 800 Gb/s.

BER Performance: At high Optical Modulation Amplitude (OMA), where RIN dominates, the use of mismatched AWGN LLRs ( $L^{awgn}_k$ ) resulted in error floor increases of up to 5 times (for 400G), 3.3 times (for 600G), and 3 times (for 800G) compared to the exact LLRs ( $L_k$ ).
Approximation Efficacy: The proposed ZC approximation ( $L^{zc}_k$ ) successfully eliminated this penalty, achieving post-FEC BER performance virtually identical to the exact LLR calculation.
System Margin: For 400G and 600G systems, the proposed method allowed the system to meet the KP4 FEC threshold ( $2.26 \times 10^{-4}$ ) with 0.1 dB and 0.4 dB less OMA, respectively, compared to the mismatched approach. For 800G, the mismatched approach failed to meet the threshold (hitting an error floor), whereas the proposed approximation maintained a reliable operating margin.

Significance and Claims
The paper concludes that for next-generation IM-DD systems dominated by laser RIN, accurate modeling of signal-dependent noise is critical for SD decoding. The authors claim that their proposed low-complexity piecewise linear approximation offers a practical solution that avoids the performance penalties of mismatched AWGN models without incurring the computational cost of exact quadratic LLR calculations. This enables reliable error-free communication for high-speed transceivers (up to 800 Gb/s) using standard concatenated FEC schemes. The work suggests that future efforts should focus on extending these results to higher-order PAM formats and experimental validation with capacity-approaching FEC codes.

The Wrong Way: The "One-Size-Fits-All" Guess

The Perfect Way: The "Exact" Math

The Paper's Solution: The "Smart Shortcut"

The Bottom Line

More like this