Joint Precoding and Port Selection (JPPS)

• JPPS is a mixed discrete–continuous transceiver design problem that optimizes beamforming/precoding alongside selecting active transmission or reception ports.
• It appears in various architectures such as uplink switched-RF, fluid antenna systems, and cell-free massive MIMO, each with unique selection and hardware constraints.
• Key strategies include alternating optimization, surrogate relaxations, and learning-based methods to balance performance, computational complexity, and CSI overhead.

Searching arXiv for recent and foundational papers on joint precoding and port selection. Search query: joint precoding port selection antenna selection MIMO arXiv Joint precoding and port selection (JPPS) denotes a class of mixed discrete–continuous transceiver design problems in which continuous precoders or beamformers are optimized jointly with the activation of a subset of transmission or reception ports. The exact label appears in secure integrated sensing and communication with fluid antenna systems, but closely related formulations also appear as source-antenna selection in two-way relaying, receive-port activation through switching networks, beamspace-port selection in cell-free massive MIMO, sparse precoding with implicit antenna activation, and fixed-cardinality antenna subset selection under RF-chain limits (Salem et al., 30 Sep 2025, Wang et al., 2011, Ouyang et al., 2022, Zhang et al., 2023, Vu et al., 2021). Taken together, these works suggest that JPPS is best understood as an umbrella for coupled port-activation and precoder design rather than as a single canonical optimization template.

1. Scope and terminology

In the JPPS literature, the meaning of “port” is architecture-dependent. In uplink switched-RF systems, a port may be a receive antenna connected to one of $L$ RF chains through a selection matrix $\mathbf S$ or activity vector $\mathbf s$ (Ouyang et al., 2022). In beamspace cell-free massive MIMO, a port is one column of the spatial DFT matrix $\mathbf F$, so port selection means choosing beamspace components $\Lambda_{b,u}$ for each BS-user pair (Zhang et al., 2023). In fluid antenna systems, port selection means choosing a subset $\mathcal S\subseteq\{1,\dots,N_s\}$ of candidate radiating positions on a 2D surface (Salem et al., 30 Sep 2025). In sparse-vector formulations, the active ports are the nonzero coordinates of the transmit vector $\mathbf x$, so $\|\mathbf x\|_0$ is the active-port count (Ma et al., 2024). By contrast, some adjacent works replace ports by BS transmit points or helper-BS participation states, which is structurally similar but not identical to intra-array port selection (Ding et al., 2020, Wang et al., 19 Mar 2026).

This variation in terminology is not merely semantic. It determines whether selection occurs at the transmitter or receiver, whether it acts on physical antennas or effective beamspace dimensions, whether the selected objects are single branches or groups, and whether the continuous design variable is a beamforming matrix, a covariance matrix, or a single transmit vector.

Representative formulation	Selected object	Coupled continuous design
(Wang et al., 2011)	one source antenna at each source	source/relay precoders and receive filters
(Ouyang et al., 2022)	$L$ active receive antennas at the BS	user covariance or precoder design
(Zhang et al., 2023)	beamspace ports $\Lambda_{b,u}$	ZF precoding with reconstructed CSI
(Salem et al., 30 Sep 2025)	$\mathbf S$0 active FAS ports out of $\mathbf S$1	multiuser beamforming for secrecy and sensing

A common misconception is that JPPS must always refer to transmit-side antenna activation at a single BS. The literature is broader. Receive-port selection in uplink MU-MIMO is mathematically a JPPS variant because the switching matrix and user precoders are jointly optimized (Ouyang et al., 2022). Conversely, not every paper with sparse or sequential transmission design is strict JPPS: some works keep the selection set external to the main optimization or allow selection to emerge only implicitly (Lakshmana et al., 2015, Choi et al., 2019).

2. Canonical optimization structure

The recurring JPPS template combines a discrete cardinality or subset-selection variable with a continuous precoder under coupled physical constraints. In the statistical-CSI uplink switched-RF formulation, the optimization is

$\mathbf S$2

subject to

$\mathbf S$3

which is explicitly a joint receive-port selection and user covariance design problem (Ouyang et al., 2022). In secure ISAC with fluid antenna systems, the main JPPS problem is

$\mathbf S$4

subject to a radar output SINR constraint $\mathbf S$5, distinct selected columns of $\mathbf S$6, and a transmit-power constraint $\mathbf S$7 (Salem et al., 30 Sep 2025). In the two-way relay precursor, the objective is total-MSE minimization over source precoders $\mathbf S$8, relay precoder $\mathbf S$9, and receive filters $\mathbf s$0, with a special single-stream source-antenna-selection mode that selects one transmit antenna at each source (Wang et al., 2011).

These formulations suggest three recurrent ingredients. First, the objective is typically sum-rate, secrecy rate, or total-MSE. Second, the constraints combine a port budget with transmit-power, relay-power, or sensing-feasibility constraints. Third, the discrete and continuous variables are coupled through the effective channels themselves: changing the selected set alters both the feasible precoder dimension and the interference or noise structure. In direct sparse formulations this coupling is absorbed into the precoder variable; for example, the ideal vector-level port-budget problem enforces $\mathbf s$1 together with $\mathbf s$2 (Ma et al., 2024).

3. Representative model families

An early JPPS precursor is the two-way amplify-and-forward relay design in "Joint Source and Relay Precoding Designs for MIMO Two-Way Relaying Based on MSE Criterion" (Wang et al., 2011). Its main model is a half-duplex $\mathbf s$3 MIMO relay architecture in which two $\mathbf s$4-antenna sources exchange information via an $\mathbf s$5-antenna relay. The joint source-and-relay problem is non-convex because $\mathbf s$6, $\mathbf s$7, $\mathbf s$8, $\mathbf s$9, and $\mathbf F$0 are multilinearly coupled. The paper then introduces a source-antenna-selection (SAS) algorithm for the single-stream case, where one antenna is chosen at each source and the relay precoder and receive beamformers are optimized for every antenna pair. This is not generic port selection, but it is a clear example of discrete transmit-port choice embedded in a joint transceiver optimization.

A second model family is switched-RF and statistical-CSI design. In uplink MU-MIMO, the base station may have $\mathbf F$1 receive antennas but only $\mathbf F$2 RF chains, forcing selection through a rectangular permutation matrix $\mathbf F$3 or binary activity vector $\mathbf F$4 (Ouyang et al., 2022). The optimization is long-term: the switching network and user covariances are designed from channel statistics under the Weichselberger model, while only the selected effective channels are estimated per coherence block. In FDD cell-free massive MIMO, the selected objects are beamspace ports $\mathbf F$5, and the resulting reconstructed CSI determines the ZF precoder and achievable sum-rate (Zhang et al., 2023). Because the selected ports constrain the dimension and quality of acquired CSI, channel acquisition itself becomes part of the JPPS mechanism.

A third family is direct explicit JPPS. "Secure ISAC with Fluid Antenna Systems: Joint Precoding and Port Selection" formulates a downlink multiuser secrecy problem in which only $\mathbf F$6 of $\mathbf F$7 candidate FAS ports are active, the selected set modifies both communication channels and radar steering, and the precoder is jointly optimized to maximize the sum secrecy rate while satisfying a minimum radar SINR requirement (Salem et al., 30 Sep 2025). Here JPPS is not a precursor or an interpretation; it is the stated problem.

4. Algorithmic strategies

Alternating optimization is the most recurrent solution pattern. In the two-way relay precursor, the primal problem is decomposed into three blocks: receiver update, relay-precoder update, and source-precoder update. The receive step has a closed-form MMSE solution, the relay step is a convex KKT system with a scalar dual variable found by bisection, and the source step is a convex QCQP solved numerically (Wang et al., 2011). In the statistical-CSI switched-RF problem, alternating optimization cycles through fixed-point updates for deterministic-equivalent auxiliary variables, water-filling or MM-based covariance updates, and greedy receive-port selection (Ouyang et al., 2022). In secure FAS-enabled ISAC, fractional programming introduces auxiliary variables $\mathbf F$8, the beamforming block is convexified by SCA and solved with CVX, and the port set is updated through a relaxed or greedy utility-driven selection step (Salem et al., 30 Sep 2025).

A second family of methods replaces exact combinatorics by structured surrogates. In sparse massive-MISO precoding, the ideal $\mathbf F$9-constrained problem is relaxed to an $\Lambda_{b,u}$0-regularized convex program, and the resulting solution is further refined by entrywise thresholding; the asymptotic active-port fraction is then characterized by closed-form Gaussian-min-max-theorem expressions (Ma et al., 2024). In quantized RSMA with imperfect CSIT, antenna activation is induced by a smooth approximation to the indicator $\Lambda_{b,u}$1, the precoder is decomposed into direction and power-control variables, the direction block is solved through a generalized power iteration on a nonlinear eigenvalue problem, and inactive antennas are finally recovered by thresholding row norms (Sung et al., 7 Aug 2025). In beamspace cell-free MIMO, GS-JPS uses an analytically derived ZF-aware sum-rate approximation to score local port replacements, which converts a coupled JPPS problem into a tractable greedy search over candidate beamspace ports (Zhang et al., 2023).

The literature therefore suggests a methodological taxonomy: exact or exhaustive search for small subset spaces, greedy refinement when analytical surrogates are available, smooth sparsity relaxations when activation can be encoded by row or entry norms, and block-coordinate procedures when the continuous and discrete variables admit efficient conditional updates.

5. Learning-based and real-time JPPS

A central theme in more recent work is the transfer of combinatorial burden from online optimization to offline learning. In "Machine Learning-Enabled Joint Antenna Selection and Precoding Design: From Offline Complexity to Online Performance," the exact $\Lambda_{b,u}$2-out-of-$\Lambda_{b,u}$3 antenna subset is chosen offline by JASPD, then a DNN predicts a shortlist of promising subsets online, after which only those candidates are refined by the optimization-based precoder solver. The reported result is a reduction of the computation complexity by 95% while retaining more than 95% of the optimal performance (Vu et al., 2021).

Cell-free MIMO extends this idea to distributed inference. In "Learning-Based Joint Antenna Selection and Precoding Design for Cell-Free MIMO Networks," each BS runs a CNN for local antenna selection and a GNN for local precoding using only locally estimated CSI. Offline training is centralized, but online operation requires no CSI exchange. The reported average performance is 98.8% of centralized MMSE + IS, with 0.25% of the computational time and zero online information exchange (Wang et al., 2024). In FDD cell-free massive MIMO, DL-JPS learns the GS-JPS selection rule from long-term power patterns $\Lambda_{b,u}$4, allowing fast online port decisions when repeated analytical search is too slow (Zhang et al., 2023).

A more recent adjacent development is "BeamAgent: LLM-Aided MIMO Beamforming with Decoupled Intent Parsing and Alternating Optimization for Joint Site Selection and Precoding" (Wang et al., 19 Mar 2026). Its discrete variable is a candidate BS site rather than an intra-array port, so it is JPPS-adjacent rather than conventional JPPS. The technical significance lies elsewhere: semantic intent parsing is decoupled from numerical optimization, the site variable is updated by vectorized search, and the beamformer is refined by gradient descent under a thresholded dark-zone penalty. The reported performance is a bright-zone power of $\Lambda_{b,u}$5, $\Lambda_{b,u}$6 dB above exhaustive zero-forcing under the same dark-zone constraint, with the end-to-end system within $\Lambda_{b,u}$7 dB of the expert upper bound and completing in under $\Lambda_{b,u}$8 s on a laptop (Wang et al., 19 Mar 2026). This suggests a possible interface layer for future JPPS systems driven by natural-language constraints.

6. Boundary cases, adjacent formulations, and misconceptions

Several influential papers are best described as JPPS-adjacent rather than strict JPPS. In imperfect-backhaul joint transmission, helper BSs are ordered by participation probability $\Lambda_{b,u}$9, and their precoders are designed sequentially so that the system degrades gracefully if some helpers mute. The paper interprets this as an implicit joint precoding plus ordered candidate-port activation strategy, not an explicit sparse port-selection optimization (Ding et al., 2020). In limited-feedback CoMP, active BS-user links are fixed beforehand by relative thresholding based on long-term channel strength, and the precoder is computed only on that sparse transmission graph while omitted links are modeled statistically (Lakshmana et al., 2015). In multi-cell MU-MIMO with imperfect CSIT, GPIP reformulates joint user selection, power allocation, and precoding as the maximization of a product of Rayleigh quotients in a lifted space; selection then emerges through near-zero beamformer blocks rather than explicit binary variables (Choi et al., 2019).

These examples clarify two common misconceptions. First, discrete activation need not always be represented by explicit binaries. It may instead emerge from thresholded row norms, sparse vector entries, or near-zero blocks. Second, not every such formulation should automatically be called JPPS. When the selected object is a user, an active link, or a BS site, the analogy is strong but the physical semantics differ. The strictest JPPS reading is obtained when the selected object is an RF-fed transmission or reception port and the coupled continuous variable is the precoder defined on the active set.

7. Performance tradeoffs and unresolved directions

Across the literature, JPPS repeatedly exposes a three-way tradeoff among performance, computational complexity, and signaling or CSI-acquisition overhead. In the two-way relay precursor, the iterative joint precoder converges in roughly 10 iterations at low SNR, about 30 at medium SNR, and around 50 at high SNR for $\mathcal S\subseteq\{1,\dots,N_s\}$0, while the SAS alternative reduces signaling overhead and can outperform poorly initialized continuous beamforming despite its simpler selection structure (Wang et al., 2011). In the statistical-CSI switched-RF problem, the outer AO loop usually converges in only three or four iterations, and the design reduces signaling and CSI acquisition overhead because switching-network updates and covariance feedback are based on slowly varying statistics rather than per-block instantaneous CSI (Ouyang et al., 2022). In secure FAS-enabled ISAC, full JPPS converges in about 6 iterations; at $\mathcal S\subseteq\{1,\dots,N_s\}$1 dB it achieves about $\mathcal S\subseteq\{1,\dots,N_s\}$2 bps/Hz while the greedy ZF-based baseline achieves about $\mathcal S\subseteq\{1,\dots,N_s\}$3 bps/Hz, and the secrecy rate drops from about $\mathcal S\subseteq\{1,\dots,N_s\}$4 bps/Hz at $\mathcal S\subseteq\{1,\dots,N_s\}$5 to about $\mathcal S\subseteq\{1,\dots,N_s\}$6 bps/Hz for $\mathcal S\subseteq\{1,\dots,N_s\}$7, directly exhibiting the secrecy–sensing tradeoff (Salem et al., 30 Sep 2025).

Several scope limitations recur. The relay precursor selects only one source antenna at each source in the single-stream case and is specific to two-way AF relaying (Wang et al., 2011). The statistical-CSI switched-RF formulation optimizes receive-port selection in the uplink and would require reformulation for downlink transmit-port selection or hybrid beamforming architectures (Ouyang et al., 2022). The secure FAS paper assumes explicit CSI, a single-target LOS sensing model, and no explicit hardware switching-cost model (Salem et al., 30 Sep 2025). The site-selection framework chooses one transmit location rather than a subset of ports and optimizes a power-shaping objective rather than a standard sum-rate objective (Wang et al., 19 Mar 2026). These limitations indicate that JPPS remains fragmented across architectures.

The cumulative picture is nonetheless coherent. JPPS is the study of transceiver optimization when the active transmission or reception dimensions are themselves design variables. Whether posed through subset constraints, switching matrices, row sparsity, or thresholded activations, the essential problem is the same: port activation reshapes the channel seen by the precoder, and the precoder reshapes the value of the active port set. That mutual dependence is the defining feature of the field (Salem et al., 30 Sep 2025, Ouyang et al., 2022, Wang et al., 2011).