Joint Synchronization and Radar Parameter Estimation for OFDM-based DISAC Systems

arXiv:2606.17699 · eess.SP · Submitted 2026-06-16 · 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: "Joint Synchronization and Radar Parameter Estimation for OFDM-based DISAC Systems".

Title and authors: Rosa: So we're talking about this paper called "Joint Synchronization and Radar Parameter Estimation for OFDM-based DISAC Systems". It sounds like they are trying to figure out how to get all the timing errors, like time offset and frequency offset, plus the delay and Doppler information from a distributed system operating in these tricky channels.

Dev: Yeah, Rosa. The authors are Fuhrling, Rou, Ranasinghe, Gabreu and Gonzalez-Prelcic from Constructor University and UCSD. They propose a framework to do that joint estimation using something called bivariate Gaussian belief propagation or GaBP after linearizing the system model <ref:2606.17699#pg1>.

Taro: It seems like the main idea is they're taking all these different unknown parameters—the timing offsets, carrier frequency offsets, and the physical delay and Doppler values—and trying to estimate them all at once instead of estimating them separately <ref:2606.17699#pg1>.

Rosa: Exactly. The paper suggests exploiting the structure of how this system works to apply that GaBP algorithm jointly on all those parameters <ref:2606.17699#pg1>. It's about making the estimation process more efficient by looking at it as a single problem instead of many separate ones.

Dev: The summary says they use a two-stage procedure for this joint estimation, starting with a coarse initialization and then moving to a fine-tracking stage that refines everything through re-linearization <ref:2606.17699#pg1>. It sounds like they’re trying to get the right answer quickly before spending time on the really precise tuning.

Taro: And the way they do that refinement, it involves a first-order Taylor linearization of the observation model, which creates a separable structure that can then be solved using message passing on a factor graph <ref:2606.17699#pg7>. That sounds like they're setting up a clean mathematical structure to handle the nonlinearity.

Rosa: Right. And what’s really interesting about that linearization is how it separates the physical parameters from the synchronization parameters, which are coupled in a nonlinear way within the observation model <ref:2606.17699#pg8>. They identify physical and synchronization parameters as two distinct groups of variable nodes in their GaBP solver.

Dev: That separation is key because it allows them to treat those two groups differently during the estimation process, which is a big step for tackling this kind of coupled estimation problem <ref:2606.17699#pg13>. They use soft interference cancellation first, targeting the physical contribution before tackling the synchronization part <ref:2606.17699#pg14>.

Title and authors: Taro: When they get to that final stage, they have to handle those cross-domain terms left over in their approximation because they inflate the posterior variances, so they need a dedicated pass where each group is treated as a hard decision <ref:2606.17699#pg15>. That's the safety net for when the soft estimation gets messy.

Rosa: So what about how good this actually is in practice? The evaluation section says they benchmark their method against the Cramer Rao lower bound, which is a theoretical limit on how accurate an estimator can possibly be <ref:2606.17699#pg4>.

Dev: The simulation results show that their radar parameter estimates and synchronization parameter estimates both approach that Cramer Rao lower bound even at moderate-to-high signal-to-noise ratio regimes <ref:2606.17699#pg4>. That's a strong result, meaning it’s performing very well when the signal is good.

Taro: The text points out that the synchronization error for time offset falls below five decibels for TO, but the carrier frequency offset estimate only drops below zero decibels <ref:2606.17699#pg4>. That suggests they might be slightly better at nailing the timing than the frequency.

Rosa: And for those radar parameters—the range and velocity estimates—they match that lower bound for high signal-to-noise ratio, but then there's a short waterfall behavior before they stabilize again at ten decibels <ref:2606.17699#pg4>. That's a bit of a caveat about how quickly things settle down.

Dev: It’s important to remember that the paper is specifically focused on OFDM-based DISAC systems operating in doubly dispersive environments, which means they are dealing with both delay and Doppler shifts changing based on relative synchronization offsets <ref:2606.17699#pg2>. The complexity comes from those effective parameters defined by those relative offsets <ref:2606.17699#pg3>.

Taro: From my side, I wonder if this approach scales well when you move away from the two-stage procedure they describe, and how robust it is when the world misbehaves and the environment is much more chaotic than a simple doubly dispersive channel <ref:2606.17699#pg15>.

Rosa: That’s a fair question for real deployment. The paper does mention future work focusing on extending this to the full two-stage estimation process and testing it across different network topologies and waveforms <ref:2606.17699#pg5>. It shows they are thinking about making it more general than just the setup they tested.

Title and authors: Dev: Yeah, and for me, I'm interested in how this maps to real-time control loops. If we need low latency, the complexity of that full two-stage process might be a hurdle compared to a simpler approach <ref:2606.17699#pg15>. We have to consider the loop rate and the failure modes when things go wrong in those iterative refinement steps.

Taro: I think what’s most important is that they successfully treat the physical and synchronization parameters as distinct groups of variable nodes, which is a structural insight that should be useful for other estimation problems <ref:2606.17699#pg15>. It’s about exploiting the model's structure to make sense of the mess.

Rosa: So to wrap up this look at "Joint Synchronization and Radar Parameter Estimation for OFDM-based DISAC Systems," it proposes a framework that uses structural coupling via a linearized observation model and bivariate GaBP to jointly refine time offset, carrier frequency offset, delay, and Doppler estimates <ref:2606.17699#pg1>.

Dev: The results confirm that this estimator closely approaches the Cramer Rao lower bound for both radar parameter estimates and synchronization parameters at high signal-to-noise ratios <ref:2606.17699#pg4>. It’s a solid performance indicator, though we have to keep an eye on those waterfall behaviors before they stabilize <ref:2606.17699#pg4>.

Taro: For me, it means that for AI systems operating in distributed ISAC networks, we have a method that can achieve very tight accuracy for timing and channel parameters even when the underlying physical environment is complex and the nodes are unsynchronized <ref:2606.17699#pg1>.

Rosa: I think this paper gives us a concrete tool for when we need to be precise about synchronization in these distributed systems, which is something that happens outside of a perfect lab setting where everything is perfectly known <ref:2606.17699#pg4>.

Dev: So, the implication is that we can get better performance on those radar estimates than if we just estimated them one by one, provided the system structure allows for this kind of joint estimation via GaBP <ref:2606.17699#pg1>. We’ll see how much that holds up in a real-time control environment.

Taro: Exactly, and the future work they point to—extending this to full two-stage estimation across different network setups—is where we need to watch next to see if this translates beyond the simulation results <ref:2606.17699#pg5>.

The paper's summary: Rosa: So, to wrap up that last bit, this paper is proposing a method where we can figure out the timing errors like time offset and frequency offset right alongside the delay and Doppler values in these distributed systems.

Dev: Exactly. The core idea is they use something called bivariate Gaussian belief propagation after they linearize the observation model to handle how all those parameters are coupled together in a nonlinear way.

Taro: What I find interesting is how they split the problem into two groups of variable nodes, physical parameters and synchronization parameters, which lets them tackle them separately during the estimation <ref:2606.17699#pg15>. That structural insight is actually pretty useful for other kinds of coupled estimation problems.

Rosa: Right. And in terms of performance, they show that this approach gets very close to the Cramer Rao lower bound for both the radar parameters and the synchronization ones, even when the signal quality isn't perfect <ref:2606.17699#pg4>.

Dev: That’s a big deal because it means we can get really tight accuracy for range and velocity estimates in these distributed ISAC setups without having to estimate everything one piece at a time separately.

Taro: But there are caveats, right? The paper notes that the synchronization error for the time offset is good, but the carrier frequency offset estimate only gets below zero decibels <ref:2606.17699#pg4>. So you get better timing than you get frequency in this setup.

Rosa: And for those radar parameters, they match that lower bound at high signal-to-noise ratios, but then there's a short dip before they stabilize again around ten decibels <ref:2606.17699#pg4>. So it’s good, but you gotta watch the stabilization.

Dev: The implication is that for AI systems operating in these distributed networks, we have a tool that can get high accuracy on timing and channel parameters even when the underlying physical environment is complex and the nodes are unsynchronized <ref:2606.17699#pg1>.

Taro: It gives us a concrete way to handle those complex coupled errors, which moves beyond just looking at the simple doubly dispersive channel model they set up in section two <ref:2606.17699#pg4>.

Rosa: Exactly. So it shows how exploiting the model structure and using that bivariate Gaussian belief propagation can really help us get a better picture of what's happening in these wireless systems.

Dev: And for future work, they plan to look at extending this to the full two-stage estimation process and testing it across different network setups and waveforms <ref:2606.17699#pg5>. That’s where we need to see if this method is robust when the real world gets more chaotic.

The paper's improvements: Rosa: So, we've looked at how they solved the problem using bivariate Gaussian belief propagation, and now we need to talk about what they suggest as improvements for this estimation framework.

Dev: Right. The paper points out that while their current method handles time offset and frequency offset well, it relies on a two-stage procedure which might be too slow for some real-time loops <ref:2606.17699#pg15>.

Taro: They suggest focusing on making that full two-stage estimation process more efficient, so the system can get estimates faster when it's running in a live environment <ref:2606.17699#pg5>.

Rosa: It sounds like they want to move away from that iterative refinement and toward something quicker for deployment outside the lab setting <ref:2606.17699#pg4>. That's what we need when you're trying to deploy this in a field roboticist scenario.

Dev: I agree. The safety case for any real-time control loop depends on latency, so any improvement that reduces the computational burden of that iterative refinement is crucial <ref:2606.17699#pg15>. That’s a major engineering hurdle they’re trying to clear.

Taro: They also mentioned extending the testing across different network topologies and different waveforms, which means they want to see if this works when the world misbehaves in ways we haven't simulated yet <ref:2606.17699#pg5>. That’s what a real-world autonomy researcher needs to worry about.

Rosa: So, the improvement is really about making the method more general and faster so it can handle more complex scenarios without needing a lot of tweaking on the fly <ref:2606.17699#pg4>. It's moving from a specific lab result to something that actually works when things are messy.

Dev: And I’m interested in how they handle the noise floor during those final steps, because even with good estimates, if the noise floor is too high, everything collapses <ref:two thousand six hundred six point one seven six nine nine#pg15. The paper shows how they try to reduce that noise by hard-subtracting the contribution of one parameter group from another.

Taro: That reduction in noise floor is important because it lets you actually get a better final estimate than you would if you just used a standard Gaussian filter <ref:two thousand six hundred six point one seven six nine nine#pg15. It's about cleaning up the data before making the final call.

Rosa: So, the big idea here is to take this solid performance they showed against the Cramer Rao lower bound and make it practical for real-time use by simplifying those refinement steps <ref:two thousand six hundred six point one seven six nine nine#pg4. It’s about translating high-performance math into something a field robot can actually use.

Conclusion: Rosa: So, to wrap up this paper on "Joint Synchronization and Radar Parameter Estimation for OFDM-based DISAC Systems," we’re looking at a method that uses structural coupling to jointly estimate all those timing and radar parameters in a distributed network.

Dev: Right. It showed that by linearizing the observation model and using bivariate Gaussian belief propagation, you can get very accurate estimates for both synchronization errors and physical delay-Doppler parameters at moderate to high signal-to-noise ratios.

Taro: From my side, it’s interesting how they treat the physical parameters and the synchronization parameters as two separate groups of variable nodes in that solver <ref:2606.17699#pg15>. That separation is a good way to manage the complexity when things go wrong.

Rosa: Exactly. It means that for AI systems trying to operate reliably in distributed ISAC networks, they have a tool that can achieve very tight accuracy on timing and channel parameters even when the underlying physical environment is complex and the nodes are unsynchronized <ref:2606.17699#pg1>.

Dev: And for real-time systems, the fact that it approaches that Cramer Rao lower bound is a strong safety case, meaning you can trust the precision of those radar estimates when signal quality is good <ref:2606.17699#pg4>.

Taro: But we gotta remember the caveats they mentioned, like the synchronization error not dropping below zero decibels for carrier frequency offset <ref:2606.17699#pg4>. So you get better timing than frequency in this setup.

Rosa: And they flagged that there’s a short waterfall behavior before those radar estimates fully stabilize around ten decibels <ref:2606.17699#pg4>. That tells us we still have to watch the settling time when deploying this outside the lab setting.

Dev: The implication is that this framework gives us a concrete way to get better performance on those radar estimates than if we just estimated them one by one, provided the system structure allows for this kind of joint estimation via Gaussian belief propagation <ref:2606.17699#pg1>.

Taro: So it’s about exploiting that model structure to make sense of the mess, and that’s a useful technique we can look at elsewhere <ref:2606.17699#pg15>.

Rosa: Exactly. We're seeing how they use this method to get a better picture of what's happening in these wireless systems when things are messy, which is something that happens outside of a perfect lab setting <ref:2606.17699#pg4>.

Dev: So the paper "Joint Synchronization and Radar Parameter Estimation for OFDM-based DISAC Systems" gives us a solid method to joint estimate time offset, carrier frequency offset, delay, and Doppler parameters <ref:2606.17699#pg1>.

Taro: We’ll keep an eye on their future work focusing on extending this to the full two-stage estimation process and testing it across different network topologies and waveforms <ref:2606.17699#pg5>. That’s where we need to see if this method is robust when the world misbehaves in ways we haven't simulated yet.

School of Computer Science and Engineering, Constructor University, Bremen, Germany · Department of Electrical and Computer Engineering, University of California San Diego

eess.SP

Submitted: 2026-06-16

Updated: 2026-11-17

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

Importance score: 75/100

The gist: Joint synchronization and radar parameter estimation for OFDM-based DISAC systems proposes a novel approach to synchronize and jointly estimate time offset, carrier frequency offset, delay…

Key concepts

DISAC Systems
Distributed Integrated Sensing and Communication (DISAC) systems involve multiple single-antenna transmit nodes operating independently. These nodes use OFDM signals for both communication and sensing tasks, but they lack a global clock synchronization, meaning each node has its own unknown timing and frequency offsets.
Bivariate Gaussian Belief Propagation (BGaBP)
This is the core estimation algorithm used to solve the complex problem. It treats physical delay/Doppler parameters and synchronization parameters as two separate groups of variables within a factor graph, allowing for joint estimation by first linearizing the observation model into a separable bilinear form.
Cramer Rao Lower Bound (CRLB)
The CRLB represents the theoretical minimum variance achievable for parameter estimation. The simulation results show that the proposed method successfully estimates radar parameters (range and velocity) and synchronization parameters close to this lower bound, demonstrating excellent performance even in challenging high SNR conditions.

Terminology

Summary

Joint synchronization and radar parameter estimation for OFDM-based DISAC systems proposes a novel approach to synchronize and jointly estimate time offset, carrier frequency offset, delay, and Doppler parameters in distributed ISAC systems operating in doubly dispersive channels. p. 0 The proposed method exploits the structure of the system model to apply a bivariate Gaussian belief propagation algorithm that estimates synchronization parameters and radar parameters jointly. p. 0 Simulation results demonstrate that the radar parameter estimates and synchronization parameter estimates approach the Cramer Rao lower bound even at moderate-to-high signal-to-noise ratio regimes. p. 0

System Model Description

The system model considers a DISAC network comprising N single-antenna transmit nodes, each subject to an unknown time offset and carrier frequency offset. The baseband equivalent OFDM transmit signal block is generated at each node as defined by equation (1). p. 1 The channel model involves a continuous-time transfer function characterized by propagation delay and Doppler shifts, which results in a doubly dispersive fading environment. p. 1 The effective delay and Doppler become dependent on relative synchronization offsets, defined as equation (11) and (12). p. 2 The discrete DD channel operator is formulated by absorbing relative synchronization offsets into the path-dependent delay and Doppler parameters. p. 2

Joint Estimation Framework

The proposed framework targets the joint estimation of physical delay-Doppler parameters and synchronization parameters, which are coupled in a nonlinear manner within the observation model. The estimation is conceived as a two-stage procedure consisting of a coarse initialization stage followed by a fine-tracking stage that iteratively refines all estimates through re-linearization. p. 2 This refinement stage employs the bivariate Gaussian belief propagation (BGaBP) solver applied after a first-order Taylor linearization of the observation model, which yields a separable structure solvable via message passing on a factor graph. p. 3

Linearization and BGaBP Application

The observation model is linearized to yield a separable bilinear model given by equation (28). The Jacobians Hphys and Hsync isolate the sensitivity of the observation to delay, Doppler, and synchronization parameters respectively. p. 3 These Jacobians are populated by iterating over directed links, resulting in sparse matrices that are initialized to zero. p. 3 The Jacobians Hphys and Hsync are then scaled to unit column norm to precondition the system, which avoids numerical issues in the GaBP solver. p. 4

Performance Evaluation

The performance is evaluated over a DISAC network of N = 5 single-antenna nodes with K = 2048 QPSK pilots per block and P = 2 paths per link. The proposed method is benchmarked against the Cramer Rao lower bound rather than against a competing method. p. 0 Figure 2 illustrates the estimation performance for synchronization parameters, showing that the proposed method works well for high SNR, closely matching the CRLB. p. 5 Figure 3 demonstrates radar parameter estimation performance, where both range and velocity estimates match the CRLB for high SNR. p. 5 The synchronization error deviates from the CRLB below 5 dB for TO, while the CFO estimate only deviates below 0 dB. p. 5 The radar parameter estimates match the CRLB for high SNR, followed by a short waterfall behavior until they stabilize again at 10 dB. p. 5

Conclusion

The paper proposes a joint synchronization and radar parameter estimation framework that exploits structural coupling via a linearized observation model and bivariate GaBP to jointly refine TO, CFO, delay, and Doppler estimates. Simulation results confirm that the proposed estimator closely approaches the CRLB. p. 5 Future work targets the full two-stage estimation process and extended evaluation across varying network topologies and waveforms. p. 5 The framework successfully treats physical and synchronization parameters as distinct groups of variable nodes in the BGaBP solver. p. 4 The final estimates are retrieved by un-scaling the preconditioned estimates to yield the final estimate Θˆ (Imacro) after convergence. p. 4

REFERENCES

arXiv:2606.17699v2 [eess. p. 0SP] 24 Jun 2026

--- Page 1 ---

Joint Synchronization and Radar Parameter Estimation for OFDM-based DISAC Systems Niclas Fuhrling †, Hyeon Seok Rou†, Kuranage Roche Rayan Ranasinghe†, Giuseppe Thadeu Freitas de Abreu† and Nuria Gonzalez-Prelcic ‡ †School of Computer Science and Engineering, Constructor University, Bremen, Germany ‡Department of Electrical and Computer Engineering, University of California San Diego, La Jolla, CA, USA

--- Page 2 ---

Fig. 1: Illustration of a DISAC system with N distinct singleantenna nodes. p. 1 Next, the baseband equivalent OFDM transmit signal block with K samples in the time domain is generated at each n-th node as sn = F H Kxn ∈ C K×1, (1) where sn ∈ C K×1 is the transmit signal vector in the time domain and FK ∈ C K×K is the normalized K-point discrete Fourier transform (DFT) modulation matrix. p. 1

--- Page 3 ---

In a DISAC architecture, following the SotA [10], each node n operates with an independent clock characterized by an unknown timing offset ϵt,n ∈ R (seconds) and carrier frequency offset ϵf,n ∈ R (Hz). Since no global synchronization is assumed, only relative offsets between transmitter n′ and receiver n affect the received signal. p. 1 We assume that ϵt,n and ϵf,n are constant over one K-sample block. p. 1 We define the relative timing and frequency offsets as ∆ϵt,n,n′ = ϵt,n′ − ϵt,n (9) ∆ϵf,n,n′ = ϵf n′ − ϵf n (10) These relative offsets will be absorbed into the effective DD channel parameters. p. 1

--- Page 4 ---

In the presence of relative synchronization offsets, the effective delay and Doppler become τeff p,n,n′ = τp,n' + ∆ϵt n,n' (11) νeff p n',n′ = νp n' + ∆ϵf n', n' (12) Then, the corresponding normalized discrete delay and Doppler indices are leff p n', n′ = τeff p n', n' Ts (13) feff p n' = Kνeff p n'Ts (14) Next, the expression that follows generalizes the point-topoint DD channel in (6) to the distributed unsynchronized ISAC setting by absorbing relative synchronization offsets into the path-dependent delay and Doppler parameters. The discrete DD channel operator is therefore Hn,n′ = P Xn,n' p=1 hp n,n' V feff p n', n' Ψ(leff p n', n') (15)

--- Page 5 ---

Finally, by defining an equivalent noise vector w˜ n ≜ FKwn with the same statistics as wn and an equivalent channel matrix H¯n,n′ ≜ FKHn,n'F H K, the end-to-end input-output (I/O) relationship can be expressed as yn = X N n' =1 Hn,n'sn' + wn. (16) Next, demodulating the time-domain (TD) baseband signal yields yn = FKrn ∈ C K×1= X N n' =1 FKHn n'F H Kxn' + FKwn. (17) Finally, by defining an equivalent noise vector w˜ n ≜ FKwn with the same statistics as wn and an equivalent channel matrix H¯n n′ ≜ FKHn n'F H K, the end-to-end input-output (I/O) relationship can be expressed as yn = X N n' =1 Hbar n,n'xn' + w n. (18)

--- Page 6 ---

Since, to the best of our knowledge, no existing work has addressed the problem of joint synchronization and channel estimation in DISAC systems operating in doubly dispersive fading environments, we propose a novel method for the joint estimation of the synchronization and delay-Doppler parameters in OFDM-based DISAC systems. The rest of this paper is organized as follows. p. 0 First, the system model is offered in Section II. p. 0 Then, in Section III, the proposed method for the joint estimation of the synchronization parameters and the delay Doppler parameters is introduced. p. 0 Finally, a comparison of the proposed GaBPbased scheme with the Cramer Rao lower bound (CRLB) is presented in Section IV. p. 0

--- Page 7 ---

The key challenge in the joint estimation of physical and synchronization parameters is that they are coupled in a nonlinear manner within the observation model, as shown in (18) and (15). However, it is possible to leverage the structure of the observation model to decouple the estimation of the two groups of parameters. p. 3 To solve the resulting bilinear estimation problem, we propose a GaBP solver that can be applied after a first-order Taylor linearization of the observation model, which yields a linearized model with a separable structure solvable via message passing on a factor graph. p. 3

--- Page 8 ---

Since the observation depends on the unknowns only through the per-link effective indices leff p and feff p, which by (11)-(12) decompose additively into a physical and a synchronization contribution, the physical parameters θphys are shared by both directions of an edge through TDD reciprocity, whereas the synchronization parameters θsync enter anti-symmetrically across the two link directions. Thus, the observation is a known-gain nonlinear function of the stacked vector θ = [θ phys T, θ sync T] T, (23) which we linearize in the following, via a current estimate θˆ(i). p. 3

--- Page 9 ---

At iteration i, the reconstructed signal at node n is given by yˆ(i) n = X N n' =1 Hˆbar n,n' xn', (24) where the estimated effective channel of the link from n′ to n, evaluated at the known path gains is defined as Hˆbar n,n' = FK X P p=1 hp e Vˆeff,(i) p Ψ(leff p e) F H K. (25) Then, the complex residual at node n can be mapped to the real domain, in order to be compatible with the real-valued Gaussian message passing, which yields z(i) n ≜ " Re(yn − yˆ(i) n) Im(yn − yˆ(i) n) (26) and stacked over all nodes into the global residual given by z(i) ≜ [z(i) T 1, · · ·, z(i) T N] T ∈ R 2NK×1. Finally, to linearize the residual, we perform a first-order Taylor expansion, which yields the separable bilinear model given by z(i) ≈ Hphys ∆θphys + Hsync ∆θsync + w, (28)

--- Page 10 ---

with the Jacobians Hphys ∈ R 2NK×2P E and Hsync ∈ R 2NK×2(N−1) isolate the sensitivity of the observation to the delay, Doppler and to the synchronization parameters, respectively. Both Hphys and Hsync are sparse and are initialized to zero, after which each link l inserts its sensitivity blocks into the 2K rows of its receiver nl and the columns of the parameters it senses. p. 3 In particular, the real-valued 2K ×2 physical sensitivity block of path p of link l is given by J(l) phys,p ≜ Re(jτ,p) Re(jν,p) Im(jτ,p) Im(jν,p) (31) which is inserted into the rows of receiver nl and the two columns of θphys associated with path p of edge el.

--- Page 11 ---

By TDD reciprocity, the two directed links of edge el fill these same columns within the rows of their respective receivers, such that both transmissions jointly inform the shared physical parameters. Similarly, the synchronization sensitivities of link l are aggregated over its paths as J(l) rel,t ≜ Re(P p jτ,p) T Im(P p jτ,p) T (32a) and J(l) rel,f ≜ Re(P p jν,p) T Im(P p jν,p) T (32b).

--- Page 12 ---

Finally, because the delay and Doppler columns scale as 1/Ts and KTs and thus differ by many orders of magnitude, each Jacobian is scaled to unit column norm to precondition the system, in order to avoid numerical issues in the GaBP solver. The prior variances of the parameters are scaled accordingly, given by ϕ˜(·),k = ϕ(·),k w 2(·),k (34)

--- Page 13 ---

After linearization and preconditioning, the problem is solved via a bivariate GaBP that treats the physical and synchronization parameters as two distinct groups of variable nodes. In the first soft-interference cancellation (IC) step, the physical contribution is first cancelled, while the synchronization part is given by zsync = z(i) − H˜phys θ˜phys (35) such that the soft-IC residual targeting variable l is defined as ztilde m→l = [zsync]m - X j! p. 4=l [Htilde sync]m,j ˜θsync,j, (36)

--- Page 14 ---

Under the scalar Gaussian approximation (SGA), the factorto-variable precision and precision-weighted mean messages are defined as Λm→l = [Htilde sync]2 m,l v m→l, µm→l = [Htilde sync]m,l ztilde m→l v m→l (38a) and µm→l = [Htilde sync]m,l ztilde m→l v m->1 (38b) which are then used at the variable node to update the posterior mean and variance, where the posterior precision, which aggregates all factor messages and the prior, is given by Λl = 1/ϕ˜sync,l + 2 X NK m=1 Λm→l (39) and the mean and variance are updated with damping ρ˜θsync l ← ρ˜θsync l + (1 − ρ) 1/Λl X m µm→l, ψsync l ← ρ ψsync l + (1 − ρ) 1/Λl. For the physical parameter phase, the process is identical with the roles of the two blocks exchanged. p. 4

--- Page 15 ---

Next, after the outer loops, the cross-domain terms left in (37) inflate the posterior variances, such that a final dedicated pass is required that treats each converged group as a hard decision and removes the cross-domain term. For the physical variables, the synchronization estimate is hard-subtracted, which leads to z'phys = z(i) − H˜sync θ˜sync (43) with noise floor reduced to v pol m→k = N0 2 + X j!=k [Htilde phys]2 m,j ψphys,j (44) Similarly, for the synchronization variables, the physical contribution is hard-subtracted, which yields z'sync = z(i) − H˜phys θ˜phys (45) with noise floor reduced to v pol m→l = N0 2 + X j!=l [Htilde sync]2 m,j ψsync,j (46) Finally, the physical-unit corrections can be retrieved by un-scaling the preconditioned estimates given by equation (47a) and (47b). p. 4

--- Page 16 ---

The paper is titled Joint Synchronization and Radar Parameter Estimation for OFDM-based DISAC Systems. REFERENCES [1] T. p. 5 Wild, V. Braun, and H. p. 5 Viswanathan, “Joint design of communication and sensing for beyond 5G and 6G systems,” IEEE Access, vol. p. 5 9, pp. 30 845–30 857, 2021 [2] M. p. 5 U. F. Qaisar et al. p. 5, “The role of ISAC in 6G networks: Enabling next-generation wireless systems,” IEEE Trans. p. 5 Netw. Sci. Eng., 2026 [3] D Zhang et al. p. 5, “Integrated sensing and communications over the years: An evolution perspective,” arXiv:2504. p. 506830, 2026 [4] S Lu et al. p. 5, “Integrated sensing and communications: Recent advances and ten open challenges,” IEEE Internet Things J. p. 5, vol. 11, no. 11, pp. 19 094–19 120, 2024 [5] H Guo et al. p. 5, “Integrated communication, localization, and sensing in 6G D-MIMO networks,” IEEE Wireless Commun. p. 5, vol. 32, no. 2, pp. 214–221, 2025 [6] E C Strinati et al. p. 5, “Toward distributed and intelligent integrated sensing and communications for 6G networks,” IEEE Wireless Commun. p. 5, vol. 32, no. 1, pp. 60–67, 2025 [7] K Meng, C Masouros, A P Petropulu, and L Hanzo, “Cooperative ISAC networks: Opportunities and challenges,” IEEE Wireless Commun. p. 5, vol. 32, no. 3, pp. 212–219–[8] D Brunner et al. p. 5, “Bistatic OFDM-based ISAC with over-the-air synchronization: System concept and performance analysis,” IEEE Trans. p. 5 Microw. Theory Techn. p. 5, vol. 73, no. 5, pp. 3016–3029 [9] K Han, K Meng, and C Masouros, “Impact of synchronization error and phase noise on OFDM-based distributed ISAC systems,” in Proc. p. 5 IEEE RadarConf25, 2025 [10] ——, “Over-the-air time-frequency synchronization in distributed ISAC systems,” 2025. p. 5 [Online]. Available: https://arxiv. p. 5org/abs/2503. p. 5

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. Joint estimation of synchronization parameters (TO and CFO) alongside radar parameters (range and velocity) allows AI systems to operate reliably in distributed ISAC networks even when nodes have independent clocks, by exploiting the structure of the system model, which can be linearized.

  2. The proposed BGaBP refinement stage enables the system to jointly estimate physical delay-Doppler parameters and synchronization offsets by treating them as two distinct groups of variable nodes within a factor graph.

  3. The system can perform fine-tracking estimation by using a first-order Taylor linearization of the observation model, which yields a linearized model with a separable structure that can be efficiently solved via message passing on a factor graph.

  4. AI systems can achieve performance approaching the Cramer Rao lower bound (CRLB) for radar parameter estimates, such as range and velocity, even at moderate-to-high signal-to-noise ratio (SNR) regimes.

  5. The framework allows for the retrieval of physical parameters by un-scaling the preconditioned estimates, providing specific corrections like ∆ˆτp,e and ∆ˆνp,e to update the expansion point for iterative refinement.

Sources

Related papers