ELAAs: Near-Field Propagation & Beamforming

• Extremely Large-Scale Antenna Arrays (ELAAs) are antenna systems with vast apertures that enable near-field propagation where spherical wavefronts replace conventional planar-wave models.
• ELAAs exploit joint angle-distance processing to achieve enhanced beam focusing, improved user separation, and integrated sensing, vital for next-gen 6G applications.
• Advanced measurement and estimation techniques in ELAAs leverage spherical-wave modeling and sparse recovery methods to overcome challenges like spatial non-stationarity and variable path visibility.

Searching arXiv for recent ELAA papers to ground the article. arXiv search query: "extremely large-scale antenna arrays near-field ELAA" Extremely large-scale antenna arrays (ELAAs) are antenna systems with very large physical apertures and, often, hundreds or thousands of elements, developed to support high-frequency 6G communications, sensing, and integrated sensing-and-communication functions. Their defining technical consequence is that practical users, scatterers, and targets frequently lie in the radiative near field rather than the classical far field, so the usual planar-wave, angle-only, and spatially stationary assumptions of conventional massive MIMO cease to be reliable. In this regime, propagation is governed by spherical wavefronts, array response depends jointly on angle and distance, beamforming becomes beam focusing in the polar or spatial domain, and channel behavior can become spatially non-stationary across the aperture (Fan et al., 2024, He et al., 2024).

1. Field-regime transition and defining characteristics

The fundamental distinction of ELAAs is not merely a larger antenna count, but a sufficiently large aperture $D$ that the near-field boundary expands to practically relevant distances. Several works use the Rayleigh or Fraunhofer distance

$R=\frac{2D^2}{\lambda}$

as the conventional far-/near-field boundary, and also distinguish a Fresnel-type boundary such as

$F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$

Because $R$ grows with the square of aperture and inversely with wavelength, high-frequency ELAA deployments at mmWave, THz, and even FR3 can push ordinary communication or sensing ranges into the near field (Wang et al., 2024, Cui et al., 2021, Wei et al., 31 Dec 2025).

In the far field, array response is predominantly angle-parameterized and a beam is steered toward a direction. In the near field, exact element-to-point distance matters, the wavefront is spherical, and the array can focus energy at a specific location in angle and distance. This introduces an additional spatial degree of freedom that can separate terminals that are close in angle but different in range; it also alters channel rank, localization identifiability, and beam-management procedures. A perspective treatment of ELAA-enabled near-field ISAC explicitly describes this as a transition from angle-only processing to angle-distance processing, with implications for communication, sensing, localization, security, and beam training (He et al., 2024).

A direct empirical illustration is provided by a mid-band ELAA measurement in which a $720$-element virtual uniform circular array operating at $16\!-\!20$ GHz had a far-field distance of $133$ m, while the transmitter–receiver line-of-sight distance was only $6.7$ m. Under those conditions the channel was unambiguously near-field, and the measured impulse responses showed curved, “s”-shaped trajectories across the aperture rather than plane-wave behavior (Fan et al., 2024).

A second defining characteristic is spatial non-stationarity. In conventional channel models, a path is often assumed visible over the whole aperture with nearly uniform large-scale behavior. In ELAAs, some scatterers illuminate only part of the array, path power may vary across visible elements, and visibility can exhibit path birth and death over the aperture. This linked pair of effects—partial visibility and power variation—appears repeatedly in ELAA measurements and models, and is treated as inherent to near-field ELAA channels rather than as a secondary perturbation (Fan et al., 2024).

2. Propagation models and channel representations

Near-field ELAA modeling replaces plane-wave steering by spherical-wave geometry. For a ULA, one representative near-field steering vector is

$$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$

with

$$r^{(n)} = \sqrt{r^2 + (n-1)^2d^2 - 2r(n-1)d\sin(\theta)}.$$

A second-order approximation yields a decomposition into an angular term and a quadratic phase term through the effective distance

$R=\frac{2D^2}{\lambda}$0

which becomes central to near-/far-field hybrid sparse representations (Wang et al., 2024).

For UPAs, the near-field model becomes two-dimensional and range-coupled. One formulation writes the UPA channel as

$R=\frac{2D^2}{\lambda}$1

with exact spherical distances embedded in $R=\frac{2D^2}{\lambda}$2. A later approximation reformulates the UPA near-field steering matrix as the outer product of two ULA near-field steering vectors, one for each planar dimension, yielding a modified 2D-DFT dictionary and a 2D block-sparse representation that supports specialized Bayesian recovery (Chen et al., 15 Mar 2026).

A mid-band ELAA measurement-and-modeling framework proposes a compact hierarchical representation for spatially non-stationary near-field channels,

$R=\frac{2D^2}{\lambda}$3

where $R=\frac{2D^2}{\lambda}$4 denotes the paths’ CFR at a reference point, $R=\frac{2D^2}{\lambda}$5 is the array manifold, $R=\frac{2D^2}{\lambda}$6 is the spatial non-stationarity matrix, and $R=\frac{2D^2}{\lambda}$7 denotes element-wise multiplication. Setting $R=\frac{2D^2}{\lambda}$8 to an all-one matrix yields a near-field stationary model; replacing $R=\frac{2D^2}{\lambda}$9 by a plane-wave manifold yields the far-field approximation. This provides a direct hierarchy from far-field to near-field stationary to near-field spatially non-stationary modeling (Fan et al., 2024).

A broader cross-field formulation further argues that ELAA links are often neither purely far-field nor purely near-field. Instead, different parameters and different path segments can belong to different regimes simultaneously. By refining each NLoS path into a first-bounce scatterer (FBS), a virtual FBS–LBS link, and a last-bounce scatterer (LBS), the model allows delay and angles to be classified separately on the transmitter and receiver sides: $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$0 This “cross-field” viewpoint implies that azimuth, elevation, and delay need not share the same field classification for a given path (Wang et al., 2024).

3. Measurement, sounding, and empirical characterization

The experimental characterization of ELAAs has advanced from theory-driven modeling toward dedicated channel sounding platforms. For mid-band FR3 measurements, one work develops a distributed modular 2-port vector network analyzer architecture with radio-over-fiber, phase compensation, and a virtual antenna array implementation based on an Anritsu distributed Modular 2-port VNA, ShockLine ME7868A series. The receiver-side ELAA is synthesized by rotating a vertically polarized omnidirectional biconical antenna on a turntable with a $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$1 cm offset from the rotation center, producing a $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$2-element virtual uniform circular array with $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$3 step size over $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$4. For each virtual element, the channel frequency response is measured over $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$5 GHz with $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$6 frequency points, and the study focuses on the $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$7 GHz subset with $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$8 points and $F_r \triangleq 0.62\sqrt{\frac{D^3}{\lambda}}.$9 Hz IF bandwidth (Fan et al., 2024).

After inverse Fourier transform, the measured channel impulse responses show pronounced “s”-shaped delay trajectories across the aperture, which are interpreted as direct evidence of spherical wavefront curvature. Some trajectories extend over the full aperture, whereas others are incomplete, indicating spatially non-stationary visibility. The same study reports that $R$0 paths are extracted within a $R$1 dB power dynamic range, including both stationary paths and a spatially non-stationary path interpreted physically as an incomplete reflection from an elevator surface. The authors validate channel reconstruction visually under far-field, near-field stationary, and near-field spatially non-stationary assumptions; among these, the near-field SnS reconstruction most closely reproduces the measured CIRs, especially for dominant paths (Fan et al., 2024).

At much higher frequency, cross-field measurements at $R$2 GHz provide direct evidence that cluster parameters vary across an ELAA aperture even within a single indoor environment. A measurement system using a Ceyear 3672C VNA, a displacement platform, and a rotator operates over $R$3 GHz with $R$4 GHz bandwidth, giving $R$5 ps time resolution and $R$6 mm space resolution. In one campaign, an $R$7 virtual UPA with $R$8 mm spacing and $R$9 mm aperture has a Rayleigh distance of $720$0 m in an environment of about $720$1 m $720$2 $720$3 m, so reflections are in the near field of the array. The measured dominant wall-reflected ray exhibits more than $720$4 mm distance variation across the array, about $720$5 arrival-angle variation, and about $720$6 dB gain variation; cluster delay, intra-cluster delay spread, and intra-cluster angular spread also vary across elements (Wang et al., 2024).

A second $720$7 GHz campaign with larger path-length range reports $720$8 clusters in one case and $720$9 in another. The corresponding Tx-side effective degrees of freedom constrained by spatial correlation are $16\!-\!20$0 and $16\!-\!20$1 for a $16\!-\!20$2-element virtual transmitter, close to the maximum of $16\!-\!20$3, which indicates low inter-element correlation under near-/cross-field propagation (Wang et al., 2024).

4. Channel estimation and structured inference

Near-field ELAA estimation differs from classical massive-MIMO estimation because the parameter set must include range-like information and, in some formulations, spatial non-stationarity. A mid-band FR3 estimator is organized in three stages: a spherical-wave model with an SnS vector per path, a robust beamformer for coarse estimation, and an MLE-based refinement exploiting parameter orthogonality to decompose a high-dimensional search into several lower-dimensional searches. The explicitly estimated parameters are azimuth angle of arrival, elevation angle of arrival, propagation delay, power or complex path gain, spherical distance, and an SnS coefficient vector whose entries lie in $16\!-\!20$4, with $16\!-\!20$5 denoting invisibility and values in $16\!-\!20$6 denoting element-dependent power variation (Fan et al., 2024).

For modular ELAA architectures under hardware impairments, a separate near-field LOS estimation framework considers multiple subarrays, each with its own distributed BBU, and models receiver nonlinearities with a third-order quasi-memoryless LNA model. The distorted received pilot can still be written in the form

$16\!-\!20$7

so the near-field geometry remains exploitable. On that basis, the paper proposes a hierarchy of estimators—LS, CM-LS, RS-LS, CM-RS-LS, and 2D-DFT-masked variants—culminating in DFT-CM-RS-LS, which combines constant-modulus structure, reduced geometry-induced subspace, and 2D spatial-frequency compaction. In the reported simulations, only about $16\!-\!20$8 of local 2D-DFT coefficients need to be forwarded, depending on subarray size, and the retained fraction decreases as subarray size increases (Demir et al., 20 Sep 2025).

A complementary line of work develops sparse representations tailored to near-field propagation. For ULA-based hybrid near-/far-field estimation, one paper constructs an orthogonal dictionary

$16\!-\!20$9

and proves that the representation coefficients are block sparse. The number of significant coefficients grows only sublinearly, with nonzero fraction decreasing on the order of $133$0; for multipath channels with $133$1 paths, the paper states that the nonzero fraction is at most of order $133$2. The resulting block-sparse recovery formulation replaces large, coherent polar dictionaries by a unitary near-field-aware basis (Wang et al., 2024).

For extremely large UPAs at mmWave/THz, another study derives a separable approximation of the near-field steering matrix and a modified 2D-DFT dictionary, then proves that the coefficient matrix is 2D block sparse. The estimation problem is solved by 2D Pattern-Coupled Sparse Bayesian Learning (2D-PCSBL), whose nominal complexity remains $133$3 under a GAMP-based approximation. In the reported results, the proposed method reaches approximately $133$4 dB NMSE at SNR $133$5 dB, while BOMP needs about $133$6 dB SNR to reach the same NMSE; at SNR $133$7 dB and $133$8, it still achieves around $133$9 dB NMSE (Chen et al., 15 Mar 2026).

These estimation results collectively indicate that ELAA inference increasingly relies on structure beyond conventional angular sparsity: spherical-wave manifolds, visibility-aware weighting, subspace restriction, and block or pattern-coupled sparsity in suitably modified transform domains. This suggests that “near-field estimation” in the ELAA literature is not a single algorithmic family, but a convergence of geometry-aware maximum-likelihood methods, structured compressed sensing, and architecture-aware low-complexity inference.

5. Beamforming architectures and wideband effects

Near-field ELAA beamforming departs from classical phase-only angular steering in two distinct ways. First, it must focus in both range and angle. Second, under wide bandwidth, the focus itself becomes frequency dependent. A wideband THz study identifies a “near-field beam split” phenomenon in which different frequencies focus on different physical points, so a phase-shifter-only beamformer designed at $6.7$0 suffers large gain loss at $6.7$1. To mitigate this, the paper partitions the array into $6.7$2 subarrays of $6.7$3 elements and uses a delay-phase precoding architecture with one time delayer per subarray and phase shifters within each subarray. The constructive phase-delay focusing design is

$6.7$4

which aligns inter-subarray spherical-wave mismatch through time delays while letting phase shifters compensate intra-subarray planar mismatch (Cui et al., 2021).

The same work defines an effective Rayleigh distance based on beamforming gain loss rather than classical phase error: $6.7$5 In the reported example with $6.7$6, $6.7$7 GHz, and $6.7$8, the classical Rayleigh distance is about $6.7$9 m while the effective Rayleigh distance is about $$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$0 m, which more closely matches the point where far-field beamforming performance deteriorates (Cui et al., 2021).

At the architecture level, dynamic hybrid beamforming has been proposed specifically for near-field ELAA communications. In that design, each antenna is either connected to exactly one RF chain or deactivated, so the analog beamformer belongs to

$$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$1

with non-overlapping connectivity

$$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$2

This permits adaptive aperture shaping in response to antenna-dependent path loss and radiation gain. The paper develops both a real-time design with instantaneous CSI and a two-timescale design with long-timescale analog optimization and short-timescale digital precoding. In the strongest near-field setting reported, about $$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$3 of antenna elements are switched off; more broadly, the proposed dynamic schemes outperform conventional fully connected and fixed-subarray hybrid beamforming, and avoid the sum-rate deterioration those baselines can exhibit as the antenna count grows (Liu et al., 2024).

These results reinforce a broader conclusion already implicit in the propagation models: for ELAAs, beamforming is no longer only a matter of synthesizing a spatial phase slope. It becomes an exercise in range-aware focusing, bandwidth-aware compensation, and, in some architectures, adaptive aperture selection.

6. Sensing, localization, and integrated functionalities

Near-field ELAA sensing exploits the same spherical-wave geometry that complicates communication. A tutorial-style synthesis emphasizes that spherical propagation adds a distance dimension to the array manifold, enabling beam focusing, single-anchor localization through curvature of arrival, and finite Cramér–Rao bounds for distance estimation, whereas the distance-estimation CRB is described as infinite under far-field planar-wave propagation (He et al., 2024).

At the algorithmic level, scalable passive localization has been developed through array partitioning. In that approach, a UE is assumed to be in the near field of the whole ELAA but in the far field of each subarray. The array is divided into $$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$4 subarrays satisfying

$$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$5

local AoAs are inferred at each subarray, and the AoAs are fused through a Bayesian geometric model. The resulting APLE algorithm has complexity $$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$6, while the refinement stage E-APLE has overall complexity

$$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$7

so both scale linearly with the number of BS antennas. In the reported simulations, E-APLE achieves less than $$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$8 cm RMSE for $$a(\theta,r)\triangleq \frac{1}{\sqrt{N}} \left[ e^{-j\frac{2\pi}{\lambda}(r^{(1)}-r)}, \ \cdots,\ e^{-j\frac{2\pi}{\lambda}(r^{(N)}-r)} \right]^T,$$9 m and SNR $$r^{(n)} = \sqrt{r^2 + (n-1)^2d^2 - 2r(n-1)d\sin(\theta)}.$$0 dB, and remains tractable in array sizes where OMP and MUSIC run out of memory (Yuan et al., 2023).

For integrated sensing and communication, near-field ELAA covariance design has been studied under exact 3D spherical-wave models. One formulation jointly serves multiple communication users and localizes multiple point targets by optimizing the transmit covariance subject to user SINR constraints. Three design criteria are considered: minimizing the sum CRB, maximizing the minimum target illumination power, and maximizing the minimum target echo signal power. Although the problems are non-convex, they admit globally optimal SDR-based solutions with low-rank structures tied to the sensing and communication subspaces. In a special case with one collocated target/CU toward the middle of a symmetric UPA, all three optimal designs become identical to the SINR-maximization design and admit a closed form; in general they differ, and the CRB can first decrease and then increase as the target/CU moves away from the transmitter/receiver (Hua et al., 2024).

Near-field ELAA also reshapes the optimization landscape of ISAC beamforming. For multi-user, multi-target ELAA-ISAC with a ULA at $$r^{(n)} = \sqrt{r^2 + (n-1)^2d^2 - 2r(n-1)d\sin(\theta)}.$$1 GHz, one study formulates sum-rate maximization subject to sensing beampattern-gain constraints and develops a Riemannian stochastic gradient descent-based augmented Lagrangian manifold optimization method. Its emphasized result is computational: for $$r^{(n)} = \sqrt{r^2 + (n-1)^2d^2 - 2r(n-1)d\sin(\theta)}.$$2 BS antennas the method is reported to be $$r^{(n)} = \sqrt{r^2 + (n-1)^2d^2 - 2r(n-1)d\sin(\theta)}.$$3 faster than the SDR/SCA benchmark, and for $$r^{(n)} = \sqrt{r^2 + (n-1)^2d^2 - 2r(n-1)d\sin(\theta)}.$$4 it is $$r^{(n)} = \sqrt{r^2 + (n-1)^2d^2 - 2r(n-1)d\sin(\theta)}.$$5 faster, while achieving similar beampattern gains and sum rates (Galappaththige et al., 29 Jan 2025).

Integrated sensing, communication, and powering extends the same logic. In a multi-cell near-field ISCAP system, each BS employs an ELAA to support downlink communication, WPT, and sensing under ER location uncertainty. The design objective is to maximize worst-case sensing power or detection probability proxy under CU SINR, ER harvesting, and power constraints. The SDR is proved tight, and a notable corollary is

$$r^{(n)} = \sqrt{r^2 + (n-1)^2d^2 - 2r(n-1)d\sin(\theta)}.$$6

because the optimized dual-purpose covariance can lie in the null space of the associated CU channel, eliminating intra-cell dual-purpose interference at the transmitter side (Guo et al., 5 Jan 2026).

Security provides another application in which range selectivity is explicitly beneficial. In a near-field ISCAP system with one CU, one sensing target, and multiple ERs treated as potential eavesdroppers, the BS transmits one information beam and a dedicated sensing/energy covariance that also functions as artificial noise. The secrecy-rate maximization problem is solved by SDR, fractional programming, and a 1D search, with additional ZF- and MRT-based alternatives. The reported numerical behavior is closely tied to near-field ELAA geometry: when the CU and target have the same angle but different distances, secrecy rate is zero only when the target shares the CU’s distance; as the target’s range moves away, secrecy becomes positive again, which the paper presents as a sharp contrast to far-field beam steering (Ren et al., 2024).

Finally, the estimation-theoretic limits of near-field sensing have been revisited for narrow-band ELAA systems with moving targets. Closed-form FIMs and CRBs are derived for joint position, velocity, and RCS estimation under spherical-wave, element-dependent slow-time Doppler models. The resulting approximations show that far-field CRBs become increasingly unreliable as arrays grow, whereas the near-field approximations remain accurate across representative ELAA operating regimes (Wei et al., 31 Dec 2025).

7. Architectures, tradeoffs, and open problems

Several architectural themes recur across the ELAA literature. Centralized ultra-massive MIMO, distributed or cell-free realizations, multi-station cooperation, RIS-based large apertures, modular arrays with local BBUs, and hybrid analog-digital transceivers all appear as viable forms of ELAA realization. What unifies them is not a single hardware topology, but the fact that each creates large electromagnetic apertures whose propagation is no longer well captured by array-global plane-wave models (He et al., 2024, Demir et al., 20 Sep 2025).

The practical tradeoffs are equally recurrent. Sequential virtual arrays are useful for proof-of-feasibility measurements but are unsuitable for dynamic channels. Modular ELAA processing reduces hardware integration burden and can lower fronthaul, but requires architecture-aware estimation and, in practical systems, will face synchronization, quantization, and calibration constraints. Dynamic hybrid architectures can exploit antenna-dependent usefulness in the near field, but their switch optimization is combinatorial. Near-field ISAC offers richer spatial control and range-angle selectivity, yet communication, sensing, and powering objectives remain tightly coupled and can compete for the same transmit covariance (Fan et al., 2024, Liu et al., 2024, Guo et al., 5 Jan 2026).

Several open problems are stated explicitly. Measurement-driven work notes the lack of full stochastic parameter laws for delay, range, visibility-region length, and SnS coefficients; channel-sounding papers emphasize the need for explicit numerical validation metrics and benchmark comparisons; sparse-estimation papers identify multiuser, multipath, wideband, and hardware-aware extensions; wideband beamforming work points to multi-antenna users, UPAs, and RIS extensions; ISAC and ISCAP studies call for broader hybrid-field models, imperfect-CSI robustness, hardware-aware designs, continuous-aperture arrays, and more complete radar signal models with clutter, Doppler, and extended targets (Fan et al., 2024, Chen et al., 15 Mar 2026, Cui et al., 2021, He et al., 2024).

Taken together, these studies portray ELAAs as a field transition rather than a mere scaling of massive MIMO. The central shift is from angle-only, plane-wave, aperture-stationary reasoning to spherical-wave, distance-aware, and often non-stationary reasoning. This shift affects channel modeling, sounding, estimation, beamforming, localization, sensing theory, secure transmission, and network-level coordination. A plausible implication is that future ELAA systems will be judged less by raw antenna count than by how effectively they exploit, estimate, and control the geometry-dependent structure induced by extremely large apertures.