Cell-Free ISAC Network Model Overview

• Cell-free ISAC network model is a distributed, cooperative architecture that replaces traditional cell boundaries with multi-static sensing and joint communication.
• It integrates communication streams and sensing operations via centralized processing, joint beamforming, and flexible AP role assignment.
• Key challenges include synchronization, fronthaul constraints, interference management, and optimizing resource allocation.

Cell-free integrated sensing and communication (CF-ISAC) denotes a network-level perceptive wireless architecture in which geographically distributed access points or base stations jointly provide communication service and environmental sensing under centralized or coordinated control. In the surveyed and representative formulations, a central processing unit, central unit, central processor, or central server coordinates distributed APs/BSs, communication users, sensing targets, and fronthaul or backhaul links, while communication and sensing share spectrum, hardware resources, and often the same transmitted signals (Galappaththige et al., 27 Feb 2025). The defining structural feature is the replacement of cell boundaries and co-located monostatic sensing by distributed cooperative transmission and, in most formulations, multi-static sensing across multiple transmit–target–receive paths (Demirhan et al., 2024).

1. Architectural scope and recurring topologies

Representative CF-ISAC models differ mainly in how transmission and sensing reception are assigned across distributed nodes. A common architecture uses distributed APs that jointly serve users in the downlink while simultaneously illuminating a target and jointly processing echoes through a central unit (Demirhan et al., 2024). Other formulations split the infrastructure into transmitting APs and receiving APs, so that one set of APs serves communication users and illuminates targets while another set gathers echoes, yielding an explicitly multi-static sensing geometry (Guo et al., 6 May 2026). A third variant keeps distributed dual-function transmitters but moves sensing reception to a dedicated radar receiver physically separated from the transmitters (Du et al., 29 Sep 2025).

The following patterns recur in the literature.

Representative pattern	Distributed roles	Example papers
Joint Tx/Rx APs	Each AP transmits and receives for ISAC	(Demirhan et al., 2024, Rivetti et al., 2024)
Split Tx/Rx AP sets	Transmit APs serve users and illuminate targets; receive APs collect echoes	(Guo et al., 6 May 2026, Wang et al., 2024, Leyva et al., 17 Jan 2025, Wang et al., 23 Feb 2026)
Distributed Tx with dedicated radar receiver	All BSs transmit jointly; one separate radar receiver senses echoes	(Du et al., 29 Sep 2025)

Across these models, users are usually single-antenna, whereas APs/BSs are multi-antenna and employ ULA, UPA, URPA, or UCA arrays depending on the formulation (Guo et al., 6 May 2026). The cell-free property is not merely geographic distribution. It also implies cooperative service without fixed cell ownership, often with all APs or selected AP subsets serving the same users and, analogously, illuminating the same targets or target regions (Darabi et al., 6 May 2026). In survey-level formulations, this cooperative structure is further refined into user-centric and target-centric clustering, in which each user or target region is associated with selected transmitting and receiving AP subsets for scalability (Galappaththige et al., 27 Feb 2025).

A persistent point of clarification is that “cell-free” does not automatically imply decentralized optimization. Many models are geographically distributed but computationally centralized: APs are synchronized, connected to a CPU, and coordinated for joint beamforming design and sensing fusion (Liu et al., 2023). By contrast, later works explicitly reduce centralization through local beamforming, consensus updates, fronthaul compression, or split processing between APs and the CPU (Zafari et al., 1 Aug 2025).

2. Baseband signal models and propagation structure

The most common downlink baseband model superposes communication streams and sensing streams at each transmitting AP. In one standard formulation, the transmit signal of AP $m$ at symbol time $\ell$ is the sum of user streams and sensing streams, with per-AP beamforming vectors and per-AP power constraints (Demirhan et al., 2024). A closely related model writes the transmit signal of transmit AP $i$ as

$$\mathbf{x}_i=\sum_{k=0}^{K}\mathbf{f}_{i,k}s_k=\mathbf{F}_i\mathbf{s},$$

where $s_0$ is the sensing stream and $s_1,\dots,s_K$ are communication streams, all with unit power and mutual independence (Wang et al., 2024). Other formulations omit a dedicated sensing stream and instead use the communication waveform itself as sensing illumination (Leyva et al., 17 Jan 2025).

The communication receiver model is correspondingly standard but explicitly coupled to sensing. For user $k$, the received signal contains the desired communication term, multi-user interference from other communication streams, interference from sensing streams when a dedicated sensing stream is present, and additive noise (Wang et al., 2024). In many formulations the achievable communication rate is

$R_k=\log_2(1+\gamma_k^{(\mathrm c)}),$

with $\gamma_k^{(\mathrm c)}$ defined by the coherent sum of desired signals from multiple APs over interference and noise (Guo et al., 6 May 2026). This coherent aggregation is one of the main distinctions from cellular models with one serving BS.

Propagation assumptions vary. Narrowband block-fading models are common in beamforming-oriented work, where AP–UE channels are vector channels and target channels are rank-one steering-outer-product responses (Demirhan et al., 2024). OFDM-based CF-ISAC models instead work per subcarrier, often assuming a cyclic prefix long enough to absorb propagation delays so that each subcarrier is flat fading and sensing is represented through delay- and Doppler-dependent phase terms (Guo et al., 6 May 2026). In the OFDM network-level comparison between cell-free and multi-cell deployments, the bistatic target channel from transmitting AP $m$ to receiving AP $\ell$0 via target cell $\ell$1 is

$\ell$2

with steering, delay, and Doppler factors separated explicitly (Darabi et al., 6 May 2026).

Communication-channel statistics also differ by paper. Some studies use Rayleigh fading for AP–UE links (Demirhan et al., 2024), while others assume Rician fading, including 3GPP Urban Micro path loss and shadowing in channel-estimation work (Farzanullah et al., 7 Jun 2025). Several learning-based beamforming papers simplify further to line-of-sight steering-vector channels between APs and users (Wang et al., 2024). This suggests that CF-ISAC has no single canonical propagation model; rather, narrowband Rayleigh/Rician models, OFDM flat-subcarrier models, and geometric steering-vector models all coexist as accepted abstractions.

3. Sensing geometry, target modeling, and sensing observables

The sensing side is distinguished by multi-static geometry. In the simplest and most frequent case, the target is a single point reflector and the transmit–target–receive channel for transmit AP $\ell$3 and receive AP $\ell$4 is modeled as a rank-one matrix of the form

$\ell$5

or its multi-target analogue, with $\ell$6 capturing path loss and radar cross section and $\ell$7 denoting the array response (Demirhan et al., 2024). Several papers adopt the Swerling-I model, so the target channel gain remains constant over the sensing interval or over $\ell$8 symbols (Guo et al., 6 May 2026).

The sensing receiver model is then the superposition of echoes induced by all transmit APs, all active streams, and sometimes all targets. In a split Tx/Rx architecture, receive AP $\ell$9 observes

$i$0

with disturbance composed of clutter and thermal noise in clutter-aware formulations (Guo et al., 6 May 2026). In contrast, some models omit clutter and explicit full-duplex leakage, retaining only target echo and additive noise in the sensing observation (Demirhan et al., 2024). This difference is substantive rather than cosmetic: it changes the sensing metric from SNR to SCNR and determines whether direct-path cancellation or clutter rejection must be modeled explicitly.

There is likewise no single sensing-performance metric. Representative formulations use:

• Joint sensing SNR aggregated across AP receivers, often with both communication and sensing streams contributing useful illumination (Demirhan et al., 2024).
• SCNR when clutter is explicitly modeled as Gaussian spatially white clutter, with the objective to maximize expected echo energy over clutter-plus-noise (Guo et al., 6 May 2026).
• Minimum sensing-SNR constraints in rate-maximizing cooperative ISAC beamforming (Wang et al., 2024).
• CRB-based criteria when sensing is posed as angular estimation or location/velocity estimation rather than detection (Rivetti et al., 2024).
• GLRT-based sensing SNR in OFDM cell-free versus multi-cell comparisons, where target detection is formalized as a hypothesis test over virtual grid cells (Darabi et al., 6 May 2026).

These metrics are not interchangeable. CRB-based models emphasize parameter-estimation accuracy, SCNR-based models emphasize clutter-limited detection quality, and sensing-rate surrogates treat sensing through communication-like logarithmic utilities. Some works explicitly state that their sensing-capacity surrogate “does not carry direct physical meaning” and is used only as an indirect indicator (Du et al., 29 Sep 2025). A common misconception is therefore that “CF-ISAC sensing performance” has a uniform mathematical definition; the literature instead uses several task-dependent abstractions.

4. Communication–sensing coupling and optimization formulations

The shared-transmission structure creates two coupled effects that recur across the literature. First, sensing streams can degrade user SINR by appearing as interference in the communication denominator (Wang et al., 2024). Second, communication streams can contribute positively to sensing illumination, so they may increase sensing SNR, SCNR, or Fisher information rather than acting as pure interference (Demirhan et al., 2024). In some decentralized OFDM models this is stated explicitly as a “Comm-to-Sensing Contribution (C2SC)” term (Zafari et al., 1 Aug 2025).

Beamforming design is therefore usually formulated as a joint optimization over all AP-local beamformers. One influential weighted objective maximizes communication sum rate plus a logarithmic sensing utility: $i$1 subject to per-AP power constraints (Demirhan et al., 2024). Other formulations invert the priority and maximize sensing SCNR subject to minimum communication rates, leading to a non-convex QCQP that is lifted to an SDP and then solved via semidefinite relaxation (Guo et al., 6 May 2026). A third class maximizes communication sum rate subject to a minimum sensing-SNR requirement and per-AP power constraints, as in cooperative downlink beamforming with separate transmit and receive AP sets (Wang et al., 2024).

The design space expands further when mode selection or activation is included. In cooperative multi-BS ISAC, binary variables assign each BS to transmitter or receiver mode, with feasibility constraints such as at least one transmitter, at least one receiver, and enough transmit antennas to serve users (Liu et al., 2023). In radiation-footprint-controlled CF-ISAC, binary activation variables determine which BSs are active, while beamforming must satisfy communication QoS, sensing CRB limits, per-BS power, and spatial footprint constraints over protected sub-regions (Chen et al., 15 Apr 2025). This makes BS activation a first-class component of the network model rather than a post-processing detail.

Security-oriented variants modify the same coupled model by treating the sensing target as an eavesdropper and injecting artificial noise. There the waveform is optimized to minimize a CRB-based sensing objective while enforcing user SINR constraints, an Eve-SNR ceiling, and per-AP power budgets (Rivetti et al., 2024). This suggests that CF-ISAC network models can subsume physical-layer security without altering the underlying distributed multistatic architecture.

5. Graph, set, and learned abstractions of the network model

Recent work increasingly represents the CF-ISAC network model as a structured machine-learning object rather than only a system of coupled matrix equations. In heterogeneous-graph formulations, node types include APs, UEs, and a sensing target, while edge types encode AP–UE communication channels and AP–target sensing-direction information (Demirhan et al., 2024). A finer-grained variant pushes this to the antenna level, using transmit-AP antenna nodes, receive-AP antenna nodes, and user nodes, with communication edge features given by channel coefficients and sensing edge features given by entries of the target-response matrix (Wang et al., 2024). In both cases the graph mirrors the physical coupling structure: APs connect simultaneously to users and to sensing entities, making them the central shared agents of communication and sensing.

Set-based learning models abstract the same structure differently. In Set Transformer beamforming, the input is the unordered set of estimated user channels, and permutation-equivariant attention blocks are used to preserve the absence of meaningful ordering among users and AP contributions (Kulathunga et al., 24 Mar 2026). This is paired with three optimization regimes—sensing-centric, communication-centric, and joint ISAC—that correspond directly to the classical constrained formulations.

Learning-based CF-ISAC models also extend beyond beamforming. In multimodal channel estimation, sensing channel estimates and UE locations are treated as side information correlated with communication channels when users are near the sensing target, and a Conditional Denoising Diffusion Model refines noisy LS communication CSI accordingly (Farzanullah et al., 7 Jun 2025). In fronthaul-limited cooperative sensing, each receive AP encodes and vector-quantizes its local space–frequency–time sensing tensor, then sends only codeword indices to the CPU, which estimates target positions and velocities from fused latent representations and known transmit signals (Wang et al., 23 Feb 2026). These papers retain the underlying CF-ISAC network model—distributed APs, centralized fusion, multistatic sensing—but change the representation of state, measurements, and coordination.

6. Coordination, fronthaul, assumptions, and unresolved modeling tensions

A recurring modeling tension is between the geographic distribution of cell-free networks and the heavy centralization of many baseline formulations. Representative beamforming papers assume perfect time synchronization, perfect frequency synchronization, coherent transmission and reception, accurate communication CSI, known target angles or statistics, and often ideal direct-path cancellation (Guo et al., 6 May 2026). Surveyed challenges include synchronization, multi-target detection, interference management, and fronthaul capacity and latency (Galappaththige et al., 27 Feb 2025). The gap between these assumptions and practical deployments remains one of the defining open issues.

Fronthaul is therefore a major discriminator among CF-ISAC models. In centralized cooperative beamforming, fronthaul overhead may scale with the number of transmitting APs, users, and antennas; one SDR-based model states that the CPU sends $i$2 complex scalars per coherence interval (Guo et al., 6 May 2026). Two-stage distributed beamforming instead exchanges only equivalent scalar summaries and scalar combining weights, yielding fronthaul cost $i$3 complex scalars rather than a cost that scales with $i$4 (Leyva et al., 17 Jan 2025). In coordinated decentralized resource optimization, the reported per-AP signaling drops from 280 scalars in a centralized scheme to 4 scalars in SplitOpt, or $i$5 scalars in JointOpt (Zafari et al., 1 Aug 2025). In distributed sensing compression, codeword-index forwarding reduces fronthaul signaling overhead by 99% relative to raw centralized sensing (Wang et al., 23 Feb 2026).

Model idealizations are similarly recurrent. Many formulations assume a single point target (Demirhan et al., 2024). Several ignore clutter, multipath target spread, or residual self-interference despite simultaneous transmit/receive operation (Demirhan et al., 2024). Others include clutter explicitly but as Gaussian spatially white clutter rather than environment-dependent structured clutter (Guo et al., 6 May 2026). Some models use perfect CSI and known geometry; some add bounded or Gaussian channel uncertainty; some include target-presence uncertainty via Bernoulli variables (Du et al., 29 Sep 2025). This suggests a broad but still fragmented modeling landscape rather than a settled canonical standard.

The most stable core of the CF-ISAC network model is therefore architectural rather than metric-specific: distributed APs or BSs, cooperative user service without cell boundaries, multistatic sensing through multiple illumination and observation paths, and some form of central or coordinated fusion. Around that core, the literature varies in waveform domain (narrowband versus OFDM), role assignment (joint Tx/Rx APs, split Tx/Rx sets, or dedicated radar receiver), sensing objective (SNR, SCNR, CRB, detection, localization), and coordination model (centralized, split, decentralized, or compressed). That variability is not incidental; it is the present mathematical form of CF-ISAC research itself (Galappaththige et al., 27 Feb 2025).