CSI-Free Multi-Antenna Beamforming Techniques

• CSI-free multi-antenna beamforming is a design paradigm that avoids explicit high-dimensional CSI acquisition by exploiting implicit observables, historical data, and structural spatial cues.
• Key methodologies include implicit-CSI design, worst-case robust beamforming, sensing-driven CSI construction, and blind passive beamforming to address unique system constraints and trade-offs.
• This approach enhances performance in mmWave, relay, wireless power transfer, and RIS systems while reducing complexity, scheduling overhead, and dependency on precise channel measurements.

Searching arXiv for recent and foundational papers on CSI-free multi-antenna beamforming to ground the article in the cited literature. CSI-free multi-antenna beamforming denotes a family of transmit, receive, hybrid, relay, and passive-array design strategies that avoid explicit or fresh high-dimensional channel state information while still exploiting spatial degrees of freedom. Across the literature, the term covers several distinct operating points: replacing explicit CSI by implicit observables such as coupling coefficients, eliminating instantaneous CSI for scheduling while retaining local per-node beam weights, constructing forward-link CSI proxies from sensing, robustifying codebooks against outdated CSI, using deterministic phase laws and antenna cycling in wireless energy transfer, optimizing secrecy without eavesdropper CSI, and configuring intelligent surfaces from received-power samples alone (Chiang et al., 2018, Zhou et al., 2020, Zhang et al., 2023, Thoota et al., 2023, López et al., 2020, López et al., 2021, Feng et al., 2024, Lai et al., 1 Jun 2025).

1. Conceptual scope and meaning of “CSI-free”

In this literature, “CSI-free” is not a single technical condition. In wideband hybrid mmWave MIMO, it means no explicit estimation of the high-dimensional channel matrices $H[k]$; instead, the transceiver measures coupling coefficients

$c_{k,i,j} = w_i^H H[k] f_j$

through beam sweeping and uses them as implicit CSI for analog beam selection and digital beamformer construction (Chiang et al., 2018). In buffer-aided relay systems, it means no instantaneous CSI for link selection or scheduling, even though each relay still uses local instantaneous CSI for MRC on $S\!\to\!R$ and MRT on $R\!\to\!D$ (Zhou et al., 2020). In wireless power transfer, it can mean eliminating ER pilots and feedback by replacing training with radar sensing at the AP (Zhang et al., 2023), or using fixed deterministic transmission laws such as AA-SS, AA-IS, SA, and rotary antenna beamforming with no CSI acquisition at all (López et al., 2020, López et al., 2020, López et al., 2021).

A closely related variant is low-CSI or stale-CSI beamforming. In initial access, the access point uses a historical CSI database plus an outdated estimate $\hat g$ to localize a CSI neighborhood and then designs a robust codebook that maximizes worst-case accumulated gain over that neighborhood, rather than relying on accurate instantaneous CSI (Thoota et al., 2023). In physical-layer security, the legitimate-user CSI is available, but Eve’s CSI is not; robustness is obtained by worst-case design over a geometric uncertainty set of potential Eve locations (Feng et al., 2024). In RIS/IS-aided MIMO, the strongest form appears: passive beamforming is performed without any CSI, using only random phase probes and received-power samples to infer per-element phase decisions (Lai et al., 1 Jun 2025).

A common misconception is that CSI-free beamforming is equivalent to beamforming without measurements. The surveyed methods contradict that interpretation. Some methods replace explicit CSI by lower-dimensional observables, some use local CSI but avoid global scheduling CSI, and some use sensing, historical databases, or power-only feedback rather than pilot-based channel estimation. This suggests that CSI-free beamforming is best understood as a reduction or re-parameterization of channel acquisition, not as a complete absence of channel-dependent information.

2. Principal system models and design paradigms

The literature spans OFDM hybrid arrays, relay networks, WPT systems, initial-access MISO links, RIS-aided MIMO, and movable-antenna secrecy systems. The unifying feature is that beam design is driven by observables or structural priors that are cheaper than explicit channel recovery.

Paradigm	CSI avoided	Core design variable
Implicit-CSI hybrid beamforming	Explicit $H[k]$ estimation	$$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$
Alternate distributed beamforming	Instantaneous CSI for scheduling	Fixed relay groups, local MRC/MRT
Training-free sensing-assisted WPT	ER pilots and feedback	$S_s, T_s, S$
Data-driven robust IA	Accurate instantaneous CSI	Codebook $\{f_\ell\}_{\ell=1}^L$
CSI-free WET and RAB	CSI acquisition entirely	AA-SS, AA-IS, SA, rotation, $p_j$
Blind passive beamforming	IS-side CSI	RIS phases $c_{k,i,j} = w_i^H H[k] f_j$0

In frequency-selective hybrid beamforming, the transmitter and receiver employ $c_{k,i,j} = w_i^H H[k] f_j$1- and $c_{k,i,j} = w_i^H H[k] f_j$2-element ULAs, analog RF beamformers $c_{k,i,j} = w_i^H H[k] f_j$3 and $c_{k,i,j} = w_i^H H[k] f_j$4, and digital baseband beamformers $c_{k,i,j} = w_i^H H[k] f_j$5 and $c_{k,i,j} = w_i^H H[k] f_j$6 over OFDM subcarriers. The effective channel for a candidate analog pair is

$c_{k,i,j} = w_i^H H[k] f_j$7

and the whitened version is

$c_{k,i,j} = w_i^H H[k] f_j$8

The design problem is then transferred from the full $c_{k,i,j} = w_i^H H[k] f_j$9 channel to low-dimensional $S\!\to\!R$0 effective channels assembled directly from coupling measurements (Chiang et al., 2018).

In alternate distributed beamforming, a half-duplex dual-hop DF network partitions $S\!\to\!R$1 relays into two fixed groups. In odd slots one group receives a common packet from the source while the other transmits a buffered common packet to the destination via distributed beamforming; in even slots the roles swap. The receiving and transmitting groups are fixed and do not vary with the channel state. The relay-level beamformers are

$S\!\to\!R$2

and the per-hop SNRs are

$S\!\to\!R$3

Hence CSI-free scheduling coexists with local instantaneous CSI for distributed receive and transmit beamforming (Zhou et al., 2020).

In training-free WPT, the AP first performs monostatic radar sensing and estimates $S\!\to\!R$4, then constructs

$S\!\to\!R$5

and solves a fairness-oriented energy beamforming SDP over the transmit covariance $S\!\to\!R$6 with $S\!\to\!R$7 (Zhang et al., 2023). In blind RIS beamforming, the surface phases $S\!\to\!R$8 are optimized from random probes and received power

$S\!\to\!R$9

without access to $R\!\to\!D$0, $R\!\to\!D$1, $R\!\to\!D$2, or the cascaded channel (Lai et al., 1 Jun 2025).

3. Core methodologies for beam design without explicit CSI

A central methodology is implicit-CSI design. In the mmWave OFDM setting, beam sweeping over codebooks $R\!\to\!D$3 and $R\!\to\!D$4 yields

$R\!\to\!D$5

For a candidate analog beam pair, the subcarrier-wise measurement matrix is

$R\!\to\!D$6

and the effective-channel estimate is

$R\!\to\!D$7

The low-SNR surrogate objective is

$R\!\to\!D$8

which motivates analog beam selection by aggregate Frobenius norm rather than by explicit mutual-information optimization (Chiang et al., 2018).

A second methodology is worst-case robust beamforming over a localized uncertainty set. In data-driven initial access, the AP computes the normalized correlation

$R\!\to\!D$9

against a historical CSI database, selects entries with $\hat g$0, expands them by $\hat g$1 temporal neighbors, and constructs a neighborhood $\hat g$2. It then solves

$\hat g$3

using a successive convex approximation of the epigraph form. The method is explicitly minimax and robust to CSI aging rather than CSI-free in the strictest sense, but it operationalizes beamforming when accurate instantaneous CSI is unavailable (Thoota et al., 2023).

A third methodology is sensing-driven CSI construction. In WPT, the sensing stage minimizes $\hat g$4 under CRLB constraints

$\hat g$5

with $\hat g$6, and the energy stage solves

$\hat g$7

The fundamental trade-off is between sensing duration $\hat g$8 and energy-transmission time $\hat g$9, with the estimation threshold chosen to balance alignment accuracy against usable charging time (Zhang et al., 2023).

A fourth methodology is blind power-based passive beamforming. In RIS-aided MIMO, the effective channel is

$H[k]$0

and for fixed active beamformer $H[k]$1, the capacity is

$H[k]$2

The paper replaces discrete capacity maximization by a sum-power surrogate $H[k]$3, where $H[k]$4, and then recovers the same per-element linear-search decisions blindly via conditional sample means: $H[k]$5 The convergence statement is

$H[k]$6

which is a particularly explicit sample-complexity guarantee within CSI-free beamforming (Lai et al., 1 Jun 2025).

4. Deterministic, diversity-oriented, and geometry-driven CSI-free beamforming

The WET literature studies deterministic multi-antenna transmission laws that require no CSI or feedback. In AA-SS, all antennas transmit the same signal; in AA-IS, all antennas transmit independent signals; in SA, one antenna transmits at a time over sub-blocks. Under correlated Rician fading, the AA-SS RF energy is

$H[k]$7

while AA-IS uses

$H[k]$8

and SA averages the nonlinear harvested energy over antennas,

$H[k]$9

A central result is that AA-IS and SA cannot take advantage of multiple antennas to improve the average provided energy, but their dispersion can be significantly reduced; AA-SS provides the greatest average energy but also the greatest energy dispersion, and consecutive antennas must be $$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$0 phase-shifted for optimum average energy performance under AA-SS (López et al., 2020).

In the massive low-power IoT setting, a CSI-based MRT reference beamformer is close to optimum whenever the farthest node experiences at least $$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$1 dB of power attenuation more than the remaining devices, whereas the adopted CSI-free mechanism provides greater WET/WIT diversity with lower energy requirements. The comparative outcome depends strongly on traffic. The CSI-free scheme performs favorably under periodic traffic conditions, but it may be deficient in case of Poisson traffic, specially if the setup is not optimally configured. The same work emphasizes that a minimum mean square error equalizer is preferable to zero-forcing for uplink information decoding in the low-power regime (López et al., 2020).

Rotary antenna beamforming extends deterministic CSI-free WET by combining a fixed equal-gain beamformer

$$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$2

with continuous or periodic mechanical rotation of the ULA. The instantaneous normalized gain is

$$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$3

and the time-averaged harvested power over one rotation is

$$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$4

Its azimuthal average gain is

$$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$5

This produces average EH gains that scale as $$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$6, and a rotation-specific linear program

$$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$7

enables max–min fairness and direct incorporation of SAR constraints (López et al., 2021).

The relay literature offers a different geometry-driven form of CSI-free beamforming. In ADB, the source-to-relay bottleneck is the minimum decoder SNR within the receiving group, while the relay-to-destination hop is the coherent sum of relay norms. The overall throughput is

$$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$8

with one term for each alternating mode. This framework is CSI-free only at the scheduling layer, but it demonstrates that a deterministic grouping policy plus buffers and distributed MRT can remove per-slot global CSI decisions (Zhou et al., 2020).

5. Performance characteristics, approximation regimes, and comparative behavior

The performance of CSI-free beamforming depends on what information is removed and what structural surrogate is introduced in its place. In hybrid mmWave beamforming, the aggregate Frobenius metric

$$F_{\mathrm{RF}}, W_{\mathrm{RF}}, F_{\mathrm{BB}}[k], W_{\mathrm{BB}}[k]$$9

is a low-complexity surrogate for the sum rate

$S_s, T_s, S$0

and it is reported to be very accurate up to about $S_s, T_s, S$1 dB. In the reported setup $S_s, T_s, S$2, $S_s, T_s, S$3, $S_s, T_s, S$4, and DFT codebooks of size $S_s, T_s, S$5, using $S_s, T_s, S$6 shortlisted analog pairs allowed the proposed method to achieve higher rates than a reference hybrid method based on explicit CSI with least-squares digital design, because the digital design used the optimal SVD of the effective channel (Chiang et al., 2018).

In ADB, numerical results show that fixed scheduling achieves a significant improvement in achievable throughput over existing link selection policies, and that increasing the number of antennas equipped at each relay is better than increasing the number of relays equipped with a single antenna when the total number of relay antennas is fixed. Symmetric grouping $S_s, T_s, S$7 is reported as often best in practice under i.i.d. fading, because it balances the worst-relay penalty on $S_s, T_s, S$8 and the beamforming gain on $S_s, T_s, S$9 (Zhou et al., 2020).

In sensing-assisted WPT, numerical results show that the training-free design performs close to the perfect-CSI upper bound and outperforms isotropic transmission and a non-joint baseline with equal time split between sensing and energy transmission. The gain is tied to a nontrivial optimum of the CRLB threshold $\{f_\ell\}_{\ell=1}^L$0: tighter sensing accuracy increases $\{f_\ell\}_{\ell=1}^L$1 and reduces $\{f_\ell\}_{\ell=1}^L$2, whereas looser sensing degrades beam placement (Zhang et al., 2023).

In environment-specific initial access, the robust max–min-sum codebook consistently outperforms eigen-beamforming and MRT under CSI aging. The reported behavior is monotone in neighborhood size: as the neighborhood grows, worst-case normalized sum gain decreases for all methods. The data further show that the proposed MMS achieves EBF’s $\{f_\ell\}_{\ell=1}^L$3 performance with only $\{f_\ell\}_{\ell=1}^L$4 beams, while MRT with $\{f_\ell\}_{\ell=1}^L$5 still fails to match MMS’s $\{f_\ell\}_{\ell=1}^L$6 robust gain in the zig-zag setting (Thoota et al., 2023).

In WET, the comparison between CSI-free and CSI-based transmission is nuanced. CSI-based MRT may be near-optimal under strong near–far disparity, but CSI-free SA offers diversity, lower downlink CSI energy cost, and favorable overall outage under periodic traffic. The same line of work reports that AA-IS and SA do not improve average incident energy, but they reduce dispersion and improve reliability; AA-SS increases average energy yet also amplifies energy variance, with strong sensitivity to LOS phase offsets and spatial correlation (López et al., 2020, López et al., 2020). RAB shifts that trade-off by generating average gains above isotropic CSI-free baselines, and the paper states that RAB performance approaches quickly, or surpasses for scenarios with sufficiently large number of EH devices or with the proposed power control, the performance of a traditional full-CSI based transmit scheme (López et al., 2021).

For RIS/IS beamforming, field tests in commercial 5G networks are especially notable. With $\{f_\ell\}_{\ell=1}^L$7 probes, the proposed blind method achieved $\{f_\ell\}_{\ell=1}^L$8 dB RSRP gain and $\{f_\ell\}_{\ell=1}^L$9 Mbps rate indoors, and $p_j$0 dB RSRP gain and $p_j$1 Mbps rate outdoors, outperforming beam training, rank beam training, and rank CSM baselines. The paper attributes the smaller outdoor gain to higher path loss at $p_j$2 GHz and a smaller IS aperture-to-distance ratio (Lai et al., 1 Jun 2025).

In secrecy beamforming with movable antennas, simulation results demonstrate that MA-enabled PLS significantly enhances the secrecy rate compared to conventional FPA-based schemes. The reported trends are non-monotonic in the number of moved antennas and in the path-loss exponent $p_j$3, indicating that spatial reconfiguration is not simply equivalent to increasing aperture or power; rather, it changes the constructive and destructive phase patterns seen by Bobs and the worst-case Eve region (Feng et al., 2024).

6. Assumptions, limitations, and persistent technical issues

The surveyed literature is heterogeneous in assumptions. Hybrid mmWave designs commonly assume single-user MIMO, ULA arrays with half-wavelength spacing, codebook-based RF beams, and analog beams held constant across the OFDM symbol (Chiang et al., 2018). ADB assumes no direct $p_j$4–$p_j$5 link, negligible inter-relay interference, infinite buffers, stationary and ergodic block fading, and synchronization across transmitting relays (Zhou et al., 2020). Training-free sensing-assisted WPT assumes LoS single-path channels per ER, narrowband propagation, monostatic sensing, quasi-static frames, and known calibration constants for amplitude reconstruction (Zhang et al., 2023). RAB uses far-field ULA steering, LoS-dominant or Rician channels, and mechanically feasible rotation (López et al., 2021). Blind passive beamforming requires the channel to remain sufficiently stationary during the $p_j$6-probe window and assumes reliable received-power measurements (Lai et al., 1 Jun 2025).

Several limitations recur across domains. Codebook design matters in implicit-CSI hybrid beamforming; coarse or non-orthogonal codebooks degrade both analog scoring and effective-channel conditioning (Chiang et al., 2018). Fixed-group distributed beamforming shifts complexity from scheduling to synchronization and buffer management (Zhou et al., 2020). Sensing-based WPT is sensitive to mobility within the frame, unknown or time-varying radar cross section, and sensing noise in the FIM/CRLB model (Zhang et al., 2023). Historical-CSI robust IA depends on repeated trajectories and environment stationarity, and the database creation and maintenance overhead are not analyzed (Thoota et al., 2023). Deterministic WET schemes rely on linear or simplified nonlinear EH models and typically neglect spatial correlation or hardware impairments in their main analyses (López et al., 2020, López et al., 2020). RAB introduces mechanical constraints such as rotational inertia, vibration, and rotor power consumption (López et al., 2021). MA-enabled secrecy assumes field-response models and feasible small-scale antenna movement with minimum separation constraints (Feng et al., 2024). Blind RIS beamforming remains sensitive to channel dynamics because the probing budget must fit within coherence time (Lai et al., 1 Jun 2025).

A second persistent issue is the precise interpretation of robustness. In several cases, CSI-free beamforming replaces explicit channel estimation with another uncertainty-handling mechanism: subcarrier aggregation in implicit CSI, ergodic buffering in ADB, CRLB-constrained sensing, minimax codebooks over CSI neighborhoods, or worst-case geometry for Eve. This suggests that CSI-free design is rarely “model-free”; it is more often a deliberate substitution of explicit CSI by structural priors, statistical averaging, or power-based sufficient statistics.

A plausible implication is that future work will continue to blur the boundary between CSI-free and low-CSI beamforming. The surveyed papers already point in that direction through hierarchical beam training and subcarrier thinning in mmWave hybrid systems, active-set and exploratory beam tracking, cooperative sensing and joint beamforming across multiple APs, adaptive tuning of neighborhood parameters $p_j$7 and $p_j$8, hybrid electronic and mechanical steering, artificial-noise extensions for MA-enabled secrecy, and adaptive sampling for blind RIS optimization (Chiang et al., 2018, Thoota et al., 2023, Zhang et al., 2023, López et al., 2021, Feng et al., 2024, Lai et al., 1 Jun 2025). The technical core remains unchanged: beamforming gains can still be realized when explicit CSI is absent, provided the design exploits some lower-overhead observable, structural regularity, or deterministic spatial law.