⚡ electrical engineering

Simultaneous Source Separation, Synchronization, Localization and Mapping for 6G Systems

This paper proposes a novel joint Bayesian framework using a sum-product algorithm on a factor graph to simultaneously perform source separation, synchronization, and mapping in cooperative 6G MP-SLAM systems, demonstrating that relaxing assumptions of perfect synchronization and orthogonal transmissions does not significantly degrade localization performance compared to state-of-the-art methods.

Original authors: Alexander Venus, Erik Leitinger, Klaus Witrisal

Published 2026-01-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alexander Venus, Erik Leitinger, Klaus Witrisal

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 in a large, echoey room with four loudspeakers (Base Stations) and three people walking around with microphones (Mobile Terminals). The speakers are all shouting the exact same message at the exact same time, but they aren't perfectly synchronized—some start a split-second earlier or later than others.

Your goal is to figure out two things simultaneously:

Where the people are walking.
Where the walls are in the room.

The Problem:
Normally, if the speakers are synchronized and shouting different messages, it's easy to tell which sound came from which speaker. But in this scenario, the speakers are shouting the same thing at the same time, and they are out of sync. The microphones hear a messy "soup" of sounds bouncing off the walls. It's like trying to figure out who threw a ball and where the walls are, just by listening to the echo, when four people are throwing identical balls at once, and they all started throwing at slightly different times.

The Solution:
The authors of this paper invented a new "smart ear" (an algorithm) that can untangle this mess. Here is how it works, using simple analogies:

1. The "Ghost" Map (Virtual Anchors)

Instead of just listening for direct sounds, the system listens for echoes. In physics, an echo off a wall looks like it's coming from a "ghost" speaker behind the wall. The system calls these Virtual Anchors.

The Analogy: Imagine looking in a mirror. You see a "ghost" version of yourself behind the glass. The system uses these ghosts to map out the walls. If it knows where the ghost is, it knows exactly where the wall is.

2. The "Mystery Guest" Party (Source Separation)

Since all the speakers are shouting the same thing, the microphones can't tell which echo came from which speaker just by listening.

The Analogy: Imagine a party where everyone is wearing the same mask and saying the same phrase. You can't tell who is who just by hearing them.
The Trick: The system looks at the timing and the angle of the sound. Even though the speakers are out of sync, the system treats the "time delay" as a secret code. By analyzing how the sound arrives from different angles, the system can deduce: "Ah, this specific echo must have come from Speaker A because of how it's delayed relative to the others." It separates the mixed-up voices without needing them to be synchronized first.

3. The "Group Chat" (Cooperative Mapping)

The three people walking around (Mobile Terminals) are all sharing what they hear.

The Analogy: Imagine three friends walking through a dark maze, each holding a flashlight. They shout out what they see to each other. Even if one friend is blind in one eye, the others can fill in the gaps. By combining their data, they build a complete, accurate map of the maze much faster than if they were alone.

4. The "Self-Correcting Clock" (Synchronization)

The system doesn't just map the room; it also figures out how much the speakers' clocks are off.

The Analogy: It's like a group of runners starting a race, but their stopwatches are all set to different times. As they run, they compare their positions. The system realizes, "Wait, Speaker A is always 0.5 seconds 'late' compared to the others," and it automatically corrects for that error while building the map.

What Did They Prove?

The authors ran computer simulations of this chaotic room. They compared their new "smart ear" method against the old "perfect world" method (where speakers are perfectly synchronized and don't interfere).

The Result:
Surprisingly, their new method performed just as well as the perfect method. Even though the speakers were shouting over each other and were out of sync, the algorithm successfully:

Found the locations of the walking people.
Mapped the walls accurately.
Figured out which echo belonged to which speaker.
Corrected the clock errors.

Why This Matters for 6G

The paper suggests this is a big deal for future 6G networks. Currently, 5G networks try to avoid interference by having speakers take turns (muting), which slows things down. This new method says, "We don't need to take turns. We can let everyone shout at once, and our algorithm will sort it out." This means faster, more accurate location tracking and better mapping of our physical environment, even in crowded, messy places like cities or big indoor stadiums.

In short: They built a mathematical tool that can listen to a chaotic, noisy room, figure out who is saying what, where the walls are, and what time it is on everyone's watch—all at the same time, without needing a conductor to keep everyone in sync.

Technical Summary: Simultaneous Source Separation, Synchronization, Localization and Mapping for 6G Systems

1. Problem Statement

Future 6G networks require robust location awareness to enable adaptive beamforming, efficient resource allocation, and integrated communication and sensing (ISAC). While Multipath-based Simultaneous Localization and Mapping (MP-SLAM) offers a promising solution by exploiting multipath components (MPCs) as virtual anchors (VAs) to overcome line-of-sight obstructions, existing cooperative MP-SLAM methods rely on restrictive assumptions. Specifically, they typically assume that base stations (BSs) are perfectly synchronized and transmit orthogonal sequences, rendering inter-BS interference negligible.

In practical scenarios, such as modern 5G systems using Positioning Reference Signals (PRS), BSs are often unsynchronized and may transmit on identical frequency bands. This leads to:

Inter-BS Interference: Mobile terminals (MTs) receive a superposition of signals from multiple BSs, making it difficult to distinguish which MPC originates from which source.
Synchronization Bias: Unknown clock differences between BSs and MTs introduce apparent delays that degrade localization accuracy.
Data Association Ambiguity: Without orthogonal sequences or synchronization, associating measured features (VAs) to their specific generating BSs becomes a complex probabilistic problem.

Current mitigation strategies, such as muting patterns in 5G, often fail to fully restore localization performance under these conditions. This work addresses the challenge of performing simultaneous source separation, synchronization, and mapping without assuming perfect BS synchronization or orthogonal transmission.

2. Methodology

The authors propose a novel Bayesian framework that jointly estimates the positions of multiple moving MTs, the locations of map features (BSs and VAs), and the synchronization biases of all nodes. The approach is built upon a factor graph representation and solved using a Sum-Product Algorithm (SPA).

Key Modeling Components:

System Model: The system considers $I$ unsynchronized MTs and $J$ unsynchronized BSs. Each BS transmits a pilot signal, and MTs observe a mixture of LOS and MPCs. The signal model accounts for clock biases ( $b_{bs,n}^{(j)}$ and $b_{mt,n}^{(i)}$ ) and antenna array geometries.
State Representation:
- MT States: Include position, velocity, orientation, and MT clock bias.
- BS States: Include clock biases.
- Map Features: Modeled as Potential Virtual Anchors (PVAs). A PVA represents a potential reflection point (VA) associated with a specific BS. The existence of a PVA is a binary random variable.
BS-Dependent Data Association: A core innovation is the modeling of the association between measurements and BSs. Since the source of a measurement is unknown, the algorithm treats the BS index as a discrete random variable. A new PVA is generated for each possible BS for every new measurement, and the algorithm probabilistically determines which BS generated the feature.
Synchronization Bias Modeling: The apparent synchronization bias observed at an MT is modeled as the difference between the BS clock bias and the MT clock bias. This allows the system to jointly estimate these offsets.
Sequential Update: The algorithm processes measurements sequentially across MTs and time steps. Legacy PVAs (previously detected) are updated, and new PVAs are introduced based on a Poisson process.
Inference: The joint posterior distribution is approximated using the SPA on a factor graph. This involves:
- Measurement Evaluation: Calculating likelihoods for new PVAs considering all possible BS associations.
- State Estimation: Computing Minimum Mean-Square Error (MMSE) estimates for MT positions, BS biases, and PVA locations.
- Pruning: A suboptimal pruning step removes PVAs with low existence probabilities to manage computational complexity, while retaining the primary LOS anchor for each BS.

3. Key Contributions

The paper outlines four primary contributions:

Novel Data Association Model: Introduction of a probabilistic source separation model for interfering BSs, enabling the algorithm to classify features by their generating BS without prior knowledge of transmission sequences.
Bias Modeling: A principled model for apparent synchronization bias as a difference between BS- and MT-related random variables, integrated into the joint estimation framework.
Joint Inference Framework: Derivation of the joint posterior distribution and a corresponding factor-graph-based SPA for the complete problem of simultaneous source separation, synchronization, and mapping.
Performance Analysis: A statistical analysis demonstrating that the proposed method achieves performance comparable to state-of-the-art cooperative MP-SLAM (which assumes synchronized and orthogonal BSs), even under unsynchronized and interfering conditions.

4. Experimental Results

The authors evaluated the algorithm in a simulated indoor environment with four BSs, four reflective surfaces, and three moving MTs over 400 time steps. The simulation included unsynchronized BSs with random clock biases and interfering transmission sequences.

Convergence: The algorithm required approximately the first 100 time steps to fully converge, after which the Root Mean Squared Error (RMSE) of MT positions and the Optimal Subpattern Assignment (OSPA) error for VAs matched the performance of the reference case (synchronized, orthogonal BSs).
Bias Estimation: The mean RMSE for BS bias estimation was 26 ps, and for MT bias estimation, it was 24 ps.
Association Accuracy: In the tested configuration, the algorithm achieved zero data association errors, correctly identifying the generating BS for all detected features.
Performance Degradation: The statistical analysis showed no significant performance degradation compared to the state-of-the-art cooperative MP-SLAM that assumes ideal synchronization and orthogonality.

5. Significance and Claims

The paper claims that the proposed method constitutes a robust and scalable solution for 6G ISAC systems where perfect synchronization and orthogonal resources cannot be guaranteed.

Principled Approach: The BS-dependent data association model is presented as a principled method for classifying features by arbitrary time-persistent properties (e.g., reflection order or feature type), not just source identity.
Scalability: The algorithm's runtime scales linearly with the number of PVAs, measurements, MTs, and particles. While the introduction of multiple BSs increases the number of potential PVAs (by a factor of $J$ ), the pruning mechanism effectively manages this complexity.
Flexibility: The SPA framework is described as flexible enough to incorporate additional sensing modalities (e.g., IMU or radar) for seamless integration into future networks.

The authors conclude that their work relaxes critical assumptions of existing MP-SLAM methods, enabling accurate localization and mapping in realistic, interference-prone 5G/6G environments without significant loss in performance. Future work is noted to include real-world experiments with 5G PRS data and extensions to dynamic network scenarios.