Real-Time Low-Complexity Zak-OTFS with Superimposed Spread Pilots on GPUs

arXiv:2611.16177 · eess.SP · Submitted 2026-11-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Real-Time Low-Complexity Zak-OTFS with Superimposed Spread Pilots on GPUs".

Dev: Here's the gist of the paper. Orthogonal time frequency space (OTFS) modulation offers robust performance under high-mobility scenarios The paper presents a real-time GPU baseband processing pipeline for Zak-OTFS with superimposed spread pilots Exploiting the chirp structure of the spread pilot, we reduce the quasi-periodic cross-ambiguity exactly to a chirpmodulated 2D IDFT, lowering its complexity from O(M2N2) to O(MN log MN) Our GPU implementation eliminates cross-ambiguity as the processing bottleneck and achieves endto-end processing within a 1.03 ms frame time for (M, N) = (2053, 31) with one turbo iteration.

Rosa: System Model of Zak-OTFS The carrier waveform of Zak-OTFS modulation is a delay–Doppler (DD) pulse specified by the delay period τp and the Doppler period νp, which together define the period lattice Λp = 2πnτp, mνp n, m ∈ Z The frame spans M delay bins and N Doppler bins, carrying MN = BT symbols where T = Nτp and B = Mνp The transmitted TD waveform is obtained by the inverse Zak transform s(t) = Z−1 t (xdd(τ, ν)) The received TD waveform is mapped back to the DD domain by the Zak transform ydd(τ, ν) = Zt(r(t)) and subsequently filtered as ywrxdd(τ, ν) = (wrx ∗σ ydd)(τ, ν) The input–output relationship of the Zak-OTFS spread pilot system is ydd

k, l: = heff

k, l: ∗σ xdd

k, l: + ndd

k, l: , ∀k, l Low-Complexity Cross-Ambiguity Reduction The direct evaluation of cross-ambiguity (14) costs O(M2N2) The paper exploits the chirp structure of the OTFS spread pilot to reduce cross-ambiguity exactly to two Hadamard products, one 2D IFFT, and an index remapping This reduction leads to a closed form for the spread pilot xspdd

k, l: which is a 2D-chirp closed form The resulting cross-ambiguity in (34) reduces to bheff

k, l: = κ∗ √ρ ω − Θ(k,l) MN × F−1 2D n Y˜

k′, l′: ω k′a M ω l′b N This computation costs O(MN log MN).

Dev: Iterative Interference Cancellation The iterative pilot–data interference cancellation addresses the data contribution estimated in the previous iteration before channel sensing At the u-th turbo iteration, the estimation of the received pilot signal Yb (u)spdd is Yb (u)spdd = Ydd − vec−1 (p EdHb(u−1)dd x˘(u−1)dt,dd) Channel sensing is then repeated using Yb (u)spdd as bh(u)eff

k, l: = AYb (u)spdd,Xsp,dd

k, l: This process allows the receiver to achieve real-time processing within the frame duration T = 1.03 ms for frame sizes up to (M, N) = (2053, 31) with one turbo iteration.

Rosa: Performance and Trade-offs The performance evaluation shows that as M increases, the larger processing gain suppresses pilot–data interference, allowing one turbo iteration to capture most of the BER improvement at a lower PDR The end-to-end pipeline meets the 1.03 ms deadline up to (M, N) = (2053, 31) with only one iteration For M ≥ 509, the BER reaches the oracle after one iteration because the data interference power on each estimated channel tap decreases as 1/(ρMN) The BER-optimal PDR depends on both frame size and the number of turbo iterations.

Title and authors: Rosa: So we're moving on to the next paper. This one is called "Dual-Domain SSL Fusion for Unified All-Type Audio Deepfake Detection". It’s about audio deepfake detection, which is a really relevant topic right now because of how realistic audio synthesis is getting.

Dev: The authors are looking at unified all type audio deepfake detection, which aims to figure out if an input clip is real or fake regardless of whether the audio was generated by a specific type of model.

Taro: I mean, for someone in autonomy research, this could be useful because it deals with failure modes when the world misbehaves—say, if we’re relying on audio cues to navigate but those cues are faked.

Rosa: Right. And it sounds like they’re using some fusion technique to combine different domain information to make a single decision about whether the audio is real or fake regardless of what kind of synthetic audio was used.

Dev: They are focusing on combining SSL, which stands for self supervised learning, with other features to achieve this unified detection across all types of audio deepfakes.

Taro: I wonder how robust this fusion is when the world misbehaves; does it handle novel synthesis methods well or just the ones it was trained on?

Rosa: That’s a fair question. They are trying to make sure that their method doesn't just get stuck on known fake audio patterns, but can generalize to new synthesis techniques.

Dev: The complexity of combining all those domains can be high, but they need this to be fast enough for practical use, especially if we want real time detection in a real-world scenario.

Taro: If it’s too slow, it defeats the purpose of having a robust system because you can't react quickly enough when something goes wrong.

Rosa: The idea is to make the fusion efficient so that you can deploy this on edge devices where processing power isn't as massive as what you have in a big data center.

Dev: So they are balancing accuracy against computational cost, which is always a challenge in deep learning applications when you need low latency.

Taro: And for the listener, it means we’re getting systems that can identify fakes quickly without having to wait for massive cloud processing power to do the heavy lifting.

Rosa: That’s what matters when you think about autonomous systems relying on audio inputs—the ability to verify those inputs quickly in a dynamic environment.

Dev: It sounds like they are setting up a framework where different types of audio information can feed into each other iteratively to build a stronger verification signal.

The paper's summary: Taro: Okay, so looking at the summary for this dual domain SSL fusion for unified all type audio deepfake detection, they are essentially saying they use a dual domain approach to tackle the challenge of different audio synthesis methods.

Rosa: They are using two separate domains to look at the challenge from different angles: one angle is looking at one type of synthesis method, and the other angle is looking at another.

Dev: This dual domain approach allows them to cover a wider range of potential deepfake artifacts than just focusing on a single modality.

Taro: So instead of just checking for one kind of artifact, they are checking for multiple kinds simultaneously across different sources.

Rosa: That means that if you have a fake audio that uses method A, or if it uses method B, their fusion system should still be able to flag it as fake.

Dev: That’s the goal—unified detection across all types of audio deepfakes, not just one specific generation technique.

Taro: It sounds like they are building a system that is more robust against adversarial attacks because it has multiple layers of defense instead of one single check.

Rosa: So they are using self supervised learning to learn these complex patterns across these different domains.

Dev: The self supervised learning helps them extract the features that are most discriminative for distinguishing real from fake audio, even when the audio generation method is varied and changing.

Taro: That’s a key point for autonomy: if you're building a system that needs to handle unexpected inputs, you need that kind of generalization.

Rosa: It sounds like they are building something that can handle the variations in synthesis methods effectively through this fusion mechanism.

Dev: The authors are emphasizing the benefit of this unified detection over older single domain methods for handling diverse audio deepfake attacks.

The paper's improvements: Taro: Now, looking at the improvements section, they highlight how their approach is better than existing methods by showing that their dual domain method is more powerful.

Rosa: They show that this dual domain method leads to a better performance because it's more comprehensive in its coverage of the attack space.

Dev: They are demonstrating superior performance compared to existing single domain methods, which shows their fusion strategy isn't just an incremental step forward, but something genuinely different.

Taro: It means when the world misbehaves and you need a system that can handle unexpected inputs, this dual domain method is the one to look at because it covers more ground.

Rosa: So the improvement is primarily about expanding what kind of deepfakes they can detect, which is a more comprehensive coverage of the attack space.

Dev: And from an engineering standpoint, it shows that their fusion strategy actually works in practice when you test it on real audio data, not just in simulation.

Taro: That’s crucial because we need to know if this method can handle the messy reality of real audio inputs without breaking down when things get noisy or unexpected.

Rosa: So they’re claiming that their approach is more comprehensive in terms of what it can detect compared to existing methods.

Dev: And for deployment, it means you have a strategy for handling those diverse inputs without having to constantly retrain your model every time a new deepfake technique pops up.

Conclusion: Rosa: So summarizing the dual domain SSL fusion for unified all type audio deepfake detection, this paper shows that using two different domains to look at the challenge from different angles gives you a better performance because it's more comprehensive in its coverage of the attack space.

Dev: It’s about having multiple layers of defense instead of relying on one single check.

Taro: And for autonomy research, this means when the world misbehaves and you need a system that can handle unexpected inputs, this dual domain method is the one to look at because it covers more ground.

Rosa: So we’re getting systems that can identify fakes quickly without having to wait for massive cloud processing power to do the heavy lifting.

Dev: It’s about balancing accuracy against computational cost, which is always a challenge in deep learning applications when you need low latency.

Rosa: So that’s the core message of this dual domain SSL fusion for unified all type audio deepfake detection, focusing on comprehensive coverage over incremental improvement.

Rosa: That’s the core message of this dual domain SSL fusion for unified all type audio deepfake detection, focusing on comprehensive coverage over incremental improvement.

Chung-Hsuan Tung, Junyao Zheng, Vineetha Yogesh*, Mudasir Ahmad Sheikh*, Danish Nisar*, Robert Calderbank, Tingjun Chen

Department of Electrical and Computer Engineering, Duke University

eess.SP

Submitted: 2026-11-12

Updated: 2026-11-12

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Orthogonal time frequency space (OTFS) modulation offers robust performance under high-mobility scenarios The paper presents a real-time GPU baseband processing pipeline for Zak-OTFS with…

Key concepts

Zak-OTFS Modulation
A modulation scheme using a delay–Doppler (DD) pulse defined by delay period $\tau_p$ and Doppler period $\nu_p$. It maps data symbols into the time-delay and frequency-shift domain, allowing robust communication in high mobility environments.
Cross-Ambiguity Reduction
The direct calculation of cross-ambiguity is computationally expensive. The paper reduces this to a 2D IFFT with chirp operations, significantly lowering complexity from $O(M^2N^2)$ to $O(MN ext{ log } MN)$, making real-time processing feasible.
Iterative Interference Cancellation
This technique estimates and subtracts data interference from the received pilot signal in each turbo iteration. This iterative process refines channel sensing, allowing the receiver to achieve near-optimal performance within a single frame time.
GPU Acceleration
The entire baseband processing pipeline is implemented on a GPU. This hardware acceleration eliminates cross-ambiguity as the bottleneck and allows for end-to-end processing of large frames (like (2053, 31)) within the strict 1.03 ms latency budget.

Terminology

Summary

Orthogonal time frequency space (OTFS) modulation offers robust performance under high-mobility scenarios The paper presents a real-time GPU baseband processing pipeline for Zak-OTFS with superimposed spread pilots Exploiting the chirp structure of the spread pilot, we reduce the quasi-periodic cross-ambiguity exactly to a chirpmodulated 2D IDFT, lowering its complexity from O(M2N2) to O(MN log MN) Our GPU implementation eliminates cross-ambiguity as the processing bottleneck and achieves endto-end processing within a 1. p. 503 ms frame time for (M, N) = (2053, 31) with one turbo iteration. p. 0 p. 0

System Model of Zak-OTFS

The carrier waveform of Zak-OTFS modulation is a delay–Doppler (DD) pulse specified by the delay period τp and the Doppler period νp, which together define the period lattice Λp = 2πnτp, mνp n, m ∈ Z The frame spans M delay bins and N Doppler bins, carrying MN = BT symbols where T = Nτp and B = Mνp The transmitted TD waveform is obtained by the inverse Zak transform s(t) = Z−1 t (xdd(τ, ν)) The received TD waveform is mapped back to the DD domain by the Zak transform ydd(τ, ν) = Zt(r(t)) and subsequently filtered as ywrxdd(τ, ν) = (wrx ∗σ ydd)(τ, ν) The input–output relationship of the Zak-OTFS spread pilot system is ydd[k, l] = heff[k, l] ∗σ xdd[k, l] + ndd[k, l], ∀k, l

Low-Complexity Cross-Ambiguity Reduction

The direct evaluation of cross-ambiguity (14) costs O(M2N2) The paper exploits the chirp structure of the OTFS spread pilot to reduce cross-ambiguity exactly to two Hadamard products, one 2D IFFT, and an index remapping This reduction leads to a closed form for the spread pilot xspdd[k, l] which is a 2D-chirp closed form The resulting cross-ambiguity in (34) reduces to bheff[k, l] = κ∗ √ρ ω − Θ(k,l) MN × F−1 2D n Y˜ [k′, l′] ω k′a M ω l′b N This computation costs O(MN log MN).

Iterative Interference Cancellation

The iterative pilot–data interference cancellation addresses the data contribution estimated in the previous iteration before channel sensing At the u-th turbo iteration, the estimation of the received pilot signal Yb (u)spdd is Yb (u)spdd = Ydd − vec−1 (p EdHb(u−1)dd x˘(u−1)dt,dd) Channel sensing is then repeated using Yb (u)spdd as bh(u)eff [k, l] = AYb (u)spdd,Xsp,dd [k, l] This process allows the receiver to achieve real-time processing within the frame duration T = 1.03 ms for frame sizes up to (M, N) = (2053, 31) with one turbo iteration. p. 0

Performance and Trade-offs

The performance evaluation shows that as M increases, the larger processing gain suppresses pilot–data interference, allowing one turbo iteration to capture most of the BER improvement at a lower PDR The end-to-end pipeline meets the 1.03 ms deadline up to (M, N) = (2053, 31) with only one iteration For M ≥ 509, the BER reaches the oracle after one iteration because the data interference power on each estimated channel tap decreases as 1/(ρMN) The BER-optimal PDR depends on both frame size and the number of turbo iterations. p. 4

Conclusion

This paper presents a real-time GPU baseband processing pipeline for Zak-OTFS with superimposed spread pilots We demonstrate that, as M increases, the larger processing gain suppresses pilot–data interference, allowing one turbo iteration to capture most of the BER improvement at a lower PDR. The GPU implementation eliminates cross-ambiguity as the processing bottleneck and achieves endto-end processing within a 1. p. 503 ms frame time for (M, N) = (2053, 31) with one turbo iteration.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Cross-ambiguity reduction for channel sensing can be achieved by reducing it to a single 2D IFFT with element-wise chirp operations achieving over 7000x speedup compared to partially vectorized direct evaluation on GPUs. This allows for real-time processing of Zak-OTFS systems with superimposed spread pilots.

  2. Development of an FFT based formulation using singular value decomposition (SVD) of the phase term enables a complexity reduction from O(M2N2) to O(min(M, N)MN log MN). This method is suitable for arbitrary pilot structures by factorizing the twist matrix.

  3. Implementation of an end-to-end GPU accelerated ZakOTFS receiver supports frame sizes up to (M, N) = (2053, 31) with one turbo iteration within a 1.03 ms latency budget. This enables real-time baseband processing for high mobility scenarios in NextG wireless networks.

  4. The system demonstrates that as the number of delay bins M increases, the bit error rate (BER) converges rapidly with the first turbo iteration capturing most of the achievable gain. This shows that larger frames increase processing gain and allow lower pilot power and near converged BER after one turbo iteration.

  5. For a fixed frame size, one turbo iteration with a PDR around -9 dB is near-optimal for BER performance within the real-time budget. This provides an optimized operational point for system design under latency constraints.

Sources

Related papers