Concurrent Multi-Beam Digital Predistortion Using FFT Beamforming and Virtual Arrays Authors: Björn Langborn, Christian Fager, Rui Hou, Thomas Eriksson Published: 2026 ABSTRACT A digital predistortion (DPD) scheme for concurrent multi-beam transmission in fully digital multiple-input, multiple-output (MIMO) systems, using Fast Fourier Transform (FFT) beamforming and so-called virtual-array processing, is proposed. In a MIMO array with nonlinear power amplifiers (PAs), transmitting multiple beams concurrently yields intermodulation products that end up in both user and non-user directions. In the setting with few users in a large array, the array dimension will typically be much larger than the number of generated intermodulation products. At the same time, linearization per PA is excessively costly for large arrays. This work shows that it is instead possible to linearize the system by producing predistorted user beams, and non-user intermodulation products, through DPD processing in a virtual array of a much smaller dimension than the physical array. Theoretical derivations and simulation examples show how this approach can lead to manyfold reductions in DPD complexity. 1. Introduction Multi-antenna transmitters offer the possibility to spatially direct multiple beams (The concept of a beam in this paper will sometimes be used interchangeably with a precoded data stream, generalizing statements from line-of-sight to other channel conditions.) concurrently, which is of interest for many applications. For communication, it allows base stations to serve multiple users in different locations concurrently. For localization and sensing, utilizing multi-beam schemes can improve localization accuracy [1]. To utilize these possibilities fully in future radio systems, the number of transmit branches is expected to increase dramatically. Simultaneously, operating PAs with as high efficiency as possible always comes with a compromise in their linearity. However, generating distortion due to nonlinear operation is problematic for multiple reasons. Firstly, distortion will be emitted in-band in the directions of the users, deteriorating communication link quality. Secondly, distortion will be generated both in-band and out-of-band into non-user directions and can interfere with other systems. As such, nonlinear distortion must be considered to fulfill transmission requirements. To this end, DPD is widely adopted. Numerous approaches have been proposed for linearizing multi-antenna transmitters, as reviewed in Section 1.1. In this work, we focus on systems employing fully digital beamforming. Prior contributions have targeted either distortion cancellation specifically in the user directions [2,3] or full-angle distortion suppression when a single dominant beam is transmitted [4]. Other recent state-of-the-art research has extended these ideas to scenarios with multiple simultaneous beams through so-called beam-domain DPD (BD-DPD) [5,6]. BD-DPD for multi-beam linearization is similar to multi-dimensional DPDs from the concurrent multi-band setting, such as 2D and 3D-DPD [7,8,9,10]. These linearization approaches take multiple user data streams as input to a DPD, and combine the streams to generate a set of multi-dimensional basis functions. As such, they suffer the drawback of a drastically increasing number of multi-dimensional basis functions as the number of user streams increases. One alternative approach to this in the multi-band setting is frequency relocation [11], which involves relocating bands to lie closer in frequency, and generating all the necessary basis functions for linearization through a single nonlinearity before inverse frequency relocation. No similar technique to this exists in the multi-beam setting, to the authors’ best knowledge. 1.1. Previous Work Handling distortion emitted from multi-input, multi-output (MIMO) transmitters can be done differently depending on system requirements, available resources, and number of users to be served simultaneously. This subsection provides a brief overview of two main categories of prior work: distortion suppression through beamforming strategies and distortion mitigation through linearization. A first line of work addresses nonlinear distortion by shaping the beamforming pattern such that distortion is nulled in specific directions, as demonstrated in [12]. In [12], a single-user scenario is studied, with reference to [13] that intermodulation will tend towards being radiated isotropically once multiple beams, or many channel taps, are considered. This glosses over the case when, e.g., there are multiple, but few, beams in relation to the array size. Additionally, a nullforming approach can prevent distortion from being transmitted in specific directions, but it will not be able to prevent distortion from being generated at the PA outputs and subsequently transmitted in many directions. For that, distortion compensation is needed. A second approach to handling nonlinear distortion is thereby through linearization. In a fully digital beamforming MIMO system, the most straightforward approach to linearization is adopting one DPD per PA in the transmitter chain. While effective, this solution becomes costly in hardware, energy consumption, and computational complexity as the number of transmit branches grows. Consequently, several alternative linearization strategies have been proposed. For a MIMO transmitter with a single beam, so-called full-angle DPD [4] has been proposed. Here, instead of linearizing L PAs with L equal-size DPDs, it is proposed to use a common DPD prior to the precoder, and to perform only minor tuning in the branches to consider branch-wise variations. It is shown to perform well, and to be able to effectively cancel distortion from the transmitter. It is worth noting that this approach still requires tuning for all the branches but one, which is used as reference. From their results, it can be seen that the tuning boxes are not insignificant in size, as their common DPD had 21 coefficients whilst the tuning boxes each had 3–10 coefficients. Furthermore, the DPD scheme considers only a single beam, while in practice it is of great interest to transmit multiple beams concurrently. Several works have addressed multi-beam linearization. In [3], the goal is to cancel inter-user distortion arising from concurrent multi-beam transmission, and the approach is validated for two users. A related technique for three concurrent users is presented in [14]. However, neither of these studies considers distortion radiated in non-user directions. In [5], BD-DPD was adopted, applying linearization prior to the beamformer and operating on user data streams rather than on per-antenna signals. With this approach, the DPD complexity scales in relation to the number of users, K, and user stream intermodulation products, rather than the number of antennas, L. For systems where K≪L, this can yield large complexity savings. However, as in 2D- and 3D-DPD for multi-band signals, the number of intermodulation products grows rapidly with K, leading to multi-dimensional basis functions such as x1x2x3*. This is significant for practical implementations that rely on lookup tables (LUTs) to reduce arithmetic computations: multi-dimensional basis functions require multi-dimensional LUTs, increasing both memory usage and interpolation/extrapolation complexity, as discussed in [15]. Lastly, BD-DPD assumes equal PAs across branches. In the presence of PA variations, the performance of BD-DPD is expected to deteriorate. Other recent contributions include neural network-based DPD architectures applied prior to the precoder [16], as well as DPD strategies for hybrid beamforming systems [2]. Although the neural network approach in [16] obtains good performance and is supported by experimental validation, it does not offer new theoretical or architectural insights. Furthermore, it incurs high memory requirements and substantial floating-point computational costs. As for hybrid beamforming architectures, they are highly relevant to study. However, subarray-based approaches suffer reduced beamforming gain, whereas fully connected architectures entail significant complexity due to crosstalk effects that must be incorporated into the DPD unless model-order reductions are applied [2,17]. 1.2. Key Contributions and Paper Organization In this paper, a novel approach to DPD processing using a so-called virtual array is proposed for concurrent multi-beam linearization. The core idea is to process the user data streams through a low-dimensional virtual transmit array, and a corresponding virtual channel representation, whose operation reflects that of the actual array. In the physical transmitter, the L branches with nonlinear PAs generate intermodulation (IM) distortion that is radiated over the air (OTA) in both user and non-user directions. In the virtual array, by contrast, Lv nonlinear DPD branches seek to generate the IM terms required to cancel the distortion from the physical array—while using far fewer branches, with Lv≪L. This concept has similarities to how previous work performed dimensionality reduction for the multi-band setting, utilizing frequency domain sparseness, via frequency relocation [11] or non-uniform downsampling [18]. However, this work performs dimensionality reduction for the multi-beam setting by utilizing so-called beam-domain sparseness. The specific contributions of this paper are as follows:A novel approach to multi-beam DPD for MIMO transmitters performing DPD in a small virtual antenna space is proposed and exemplified. It is shown how such an approach can lead to manyfold reductions in complexity.Practical considerations in implementing the proposed scheme are accounted for through simulation and discussion, considering multi-user cases, different channels, and lookup table (LUT) realizations.Simulation results verify the theoretical findings, show the applicability of the proposed scheme, and demonstrate how the DPD scales in computational complexity in relation to the number of users. A novel approach to multi-beam DPD for MIMO transmitters performing DPD in a small virtual antenna space is proposed and exemplified. It is shown how such an approach can lead to manyfold reductions in complexity. A novel approach to multi-beam DPD for MIMO transmitters performing DPD in a small virtual antenna space is proposed and exemplified. It is shown how such an approach can lead to manyfold reductions in complexity. Practical considerations in implementing the proposed scheme are accounted for through simulation and discussion, considering multi-user cases, different channels, and lookup table (LUT) realizations. Practical considerations in implementing the proposed scheme are accounted for through simulation and discussion, considering multi-user cases, different channels, and lookup table (LUT) realizations. Simulation results verify the theoretical findings, show the applicability of the proposed scheme, and demonstrate how the DPD scales in computational complexity in relation to the number of users. Simulation results verify the theoretical findings, show the applicability of the proposed scheme, and demonstrate how the DPD scales in computational complexity in relation to the number of users. This paper is organized as follows: The assumed system model description is expanded upon in Section 2. Firstly, in Section 2.1, the model description is expanded upon in order to comprehend how distortion is spatially distributed in the beam domain during concurrent multi-beam operation. After that, in Section 2.2, Fast Fourier Transform (FFT) beamforming is described. This is followed in Section 3 by an example where the orthogonality of FFT beamforming highlights the sparsity of the beam domain after over-the-air (OTA) transmission. Section 4 goes on to describe how processing in a virtual array, smaller than the actual physical one, can produce the same linearization signals as per-PA (PP) DPD but at a lower computational cost. Lastly, numerical simulation results are presented in Section 5 to verify the proposed approach, alongside a conclusion in Section 6. 2. System Modeling The system modeling of this work focuses on a discrete time-domain, baseband description of concurrent multi-beam transmission in a fully digital multi-branch transmitter setting, firstly by detailing how intermodulation products arising from nonlinear PAs can be understood in such a description, and then relating them to FFT beamforming to demonstrate the suitability of the proposed approach. Two key concepts used throughout this work are the beam domain and the antenna space. The beam domain refers to the domain in which signals have already propagated through a channel and formed beams as a result of beamforming. In Figure 1, this is illustrated by the colored beams (red and orange), which are observed over the air in various directions as signals y(1),…,y(Q). Some beams carry the user streams x1,…,xK, while others contain only IM distortion. In contrast, the antenna space refers to the signals on the different transmit branches, prior to channel propagation. In Figure 1, the antenna space encompasses the signals u1,…,uL, z1,…,zL and also fPA,1(z1),…,fPA,L(zL). The beam domain and antenna space are related through multiplication with a channel or beamforming matrix. For a channel H∈CQ×L, a signal z∈CL×N in the antenna space, where N is the number of baseband data samples, the corresponding beam-domain signal is Y=Hz∈CQ×N. This relation is of particular interest when Q=L, and H is an orthogonal matrix, as will be later seen in the subsection on Fast Fourier Transform beamforming. Additionally, whilst this paper initially expands upon the relation between the antenna space and beam domain in the context of line-of-sight (LOS) channels, it is later addressed for non-LOS conditions as well. To limit the scope of the paper, all signals x1,…,xK listed in this work are considered to be single-band signals. This can however be generalized to multi-band signals as well, with appropriate modification to the involved DPD units. As a final comment, since this work considers fully digital beamforming, the numbers of antennas, transmit branches, and PAs are all equal and denoted by L. 2.1. Concurrent Transmission of Multiple Beams Consider the schematic setup of Figure 1, but start by letting the DPDs be bypassed such that zl=ul, l=1,…,L. In this setup, the beamformer W is formed to direct user streams x1,…,xK to directions 1,…,Q in matrix form as(1)W=w(1)…w(K), with ||w(k)||2=1, yielding PA input signals(2)z1z2⋮zL=[w(1),…,w(K)]x1x2⋮xK. Alternatively, the signal zl passed to PA l can be written as(3)zl=wl(1)x1+…+wl(K)xK. Now, it is known that Volterra series are well-suited to model a broad range of nonlinear functions, and as such, it is commonly applied to model saturation, soft clipping, and other effects in PAs. For practical application, different finite truncations focused on the lower-order nonlinear terms of the Volterra series are frequently used. This can be justified by the fact that lower-order nonlinear terms typically dominate the induced distortion. For qualitative analysis here, consider simply a third-order polynomial baseband PA model as(4)fPA,l(zl)=c1,lzl+c3,lzl|zl|2. where fPA,l refers to the nonlinear function of the PA on branch l, and c1,l,c3,l are modeling coefficients. Then, inserting (3) into (4) yields(5)fPA,l(zl)=c1,l(wl(1)x1+…+wl(K)xK)+c3,l(wl(1)x1+…+wl(K)xK)|wl(1)x1+…+wl(K)xK|2=c1,l(wl(1)x1+…+wl(K)xK)+c3,l(wl(1)wl(1)wl(1)*x1|x1|2+wl(1)wl(2)wl(2)*x1|x2|2+wl(1)wl(1)wl(2)*x12x2*+…). First, assume that all PAs are equal, such that c1,l and c3,l are the same for each branch l. Then, let each PA output propagate through a channel vector hl=hl(1)…hl(Q) from branch l to OTA location q, as (6)y(q)=h1(q)…hL(q)fPA,1(z1)…fPA,L(zL)T. Consider the first-order linear terms(7)y1st(q)=h1(q)c1(w1(1)x1+…+w1(K)xK)+…+hL(q)c1(wL(1)x1+…+wL(K)xK). Then it is known from beamforming theory that if x1 is to go in direction (q), then beamforming weights can be chosen such that wl(1)=(hl(q))*. This yields constructive interference c1∑hl(q)wl(1)=c1. Then, by studying the third-order terms in (5), it can be seen that since wl(1)wl(1)*,wl(2)wl(2)* will be real constants, x1|x1|2 and x1|x2|2 will also go in direction (q). However, the summation of the IM products last listed in (5), c3∑lhl(q)wl(1)wl(1)wl(2)*, typically has constructive interference in a direction other than (q). Generally, it can be seen that the terms of the form w(i)w(j)w(k)*, where k≠i,j, will give rise to new beam directions of constructive interference, which are referred to as off-beams. Two interesting observations can be made from (5), still assuming c3,l=c3∀l. Firstly, the generated off-beam terms and directions can be deterministically computed from the intended transmit directions. To provide an example, off-beam terms for a static nonlinearity of order P=3 and for K=2,3,4 concurrent users are listed in Table 1. Secondly, the nonlinear products of the input streams are of a different form in the off-beams than for the in-beam. Thus, one needs to be aware of all beam directions and manage to generate the correct IM terms in the respective directions in order to linearize the system. This can readily be understood to grow more complex with more beams. Simplified assumptions have been made in this subsection for qualitative analysis. The simple beamformer in (1) will later be further elaborated on in Section 4.3. Furthermore, the validity of the proposed method when PAs vary on different branches is investigated by simulation in Section 5.5, addressing the assumption of equal PAs. Lastly, it is well-known that PAs will often exhibit higher-order nonlinear behavior than order 3, as in (4), and that memory effects will be present unless the signal bandwidth is very narrow. Section 4.4 comments on introducing higher-order terms in the scheme, and whilst the analysis here was done for static PAs, memory can be added without any loss of generality. 2.2. Fast Fourier Transform Beamforming FFT beamforming is one common way to efficiently limit the beamforming matrix to an orthogonal set of beam directions, by first defining(8)ω=e−j2π/L. Then, a one-dimensional discrete Fourier transform (DFT) beamformer matrix can be formulated as(9)WDFT=1L111…1ωω2…1ω2ω4…⋮⋮⋮⋮=1Lω(b1)ω(b2)ω(b3)…, where bi∈[1,L] are beam indices indexing the beams on the DFT beamgrid. In implementation though, no matrix multiplication of WDFT with data is needed, as would otherwise be indicated by (2). Instead, FFT algorithms can perform this computation at lower complexity. It is worth noting that when the main beams are limited to a DFT beamgrid, bi∈[1,L], the off-beam directions will be restricted to the same beamgrid as well, given equal PAs. Furthermore, extension to two-dimensional beamforming is straightforward, by implementing FFT along both horizontal and vertical directions, but it is omitted from the derivations for brevity. All the analysis following in this paper can then be applied to each dimension, horizontal or vertical, separately. With the system model established, the following section details how per-PA (PP) DPD produces predistorted streams and how intermodulation beam placement is affected by aliasing due to the antenna array being finite and discrete. 3. Per-PA DPD This section details the processing steps for transmitting K user streams using per-PA (PP) DPD, as illustrated in Figure 1. A three-user example is used to demonstrate two important aspects: (i) the inherent sparsity of the resulting beam-domain representation for large L and (ii) an effect referred to as array aliasing, which appears due to the finite and discrete nature of the antenna array. 3.1. Per-PA DPD Example for Three Concurrent Users Let the antenna array be a linear, one-dimensional array in space, and restrict the beamformer to a DFT beamgrid. As such, this subsection follows the notation of Figure 1, in which x1,…,xK represent user streams and L represents the number of transmit branches, whilst the beamformer W is represented by WDFT from (9). Consider that there are L=16 transmit branches (While L=16 is chosen here for clarity of exposition, practical systems would typically employ significantly larger arrays), transmitting x1,x2,x3 to K=3 users in different LOS directions. Being restricted to a DFT beamgrid, the users can be located at beam indices {b1,b2,b3}∈[1,L]. To exemplify, let b1=1,b2=2,b3=5 such that(10)z=WDFTx=1L111…1ωω2…1ω2ω4…⋮⋮⋮⋮x1x200x30⋮=1Lx1+x2+x3x1+ωx2+ω4x3x1+ω2x2+ω8x3⋮=ω(b1)x1+ω(b2)x2+ω(b3)x3. Assume for analytical brevity that each DPD is represented by the same nonlinear function f(·). Applying the nonlinear DPDs to each of these beamformed data streams, with a third-order nonlinearity f(x)=θ1x+θ3x|x|2 as the DPD function, then(11)f(WDFTx)=1Lθ1(x1+x2+x3)+1Lθ3(x1+x2+x3)|x1+x2+x3|2θ1(x1+ωx2+ω4x3)+1Lθ3(x1+ωx2+ω4x3)|x1+ωx2+ω4x3|2θ1(x1+ω2x2+ω8x3)+1Lθ3(x1+ω2x2+ω8x3)|x1+ω2x2+ω8x3|2⋮︸L×1, is obtained. How these signals will be spatially distributed OTA can be seen by transformation to the beam domain, achieved through an inverse Fourier transform. The user streams will then be clearly separated in the beam domain, like the beams in Figure 1. With the matrix WIDFT=LWDFT*, where * denotes the Hermitian transpose, factors L outside the matrix cancel out, resulting in data streams(12)LWDFT*f(WDFTx)=θ1x1+1Lθ3x1(|x1|2+|x2|2+|x3|2)θ1x2+1Lθ3x2(|x1|2+|x2|2+|x3|2)1Lθ3x22x1*1Lθ3x1x3x2*θ1x3+1Lθ3x3(|x1|2+|x2|2+|x3|2)1Lθ3x2x3x1*01Lθ3x32x2*1Lθ3x32x1*0001Lθ3x12x3*1Lθ3x1x2x3*1Lθ3x22x3*1Lθ3x12x2*. Each entry at a separate index here represents a signal going in a unique beam direction. Two observations are worth noting. The first observation is that only 12 out of the 16 entries are non-zero. This is unsurprising, but important, as Table 1 shows that there can only be up to nine off-beam terms going in potentially different directions, in addition to the K=3 user directions, for a third-order nonlinearity. This is also true when L is a very large value. Thus, for large L, there are very few low-order signal components in relation to the array size. As for the second observation, concerning where each signal component ends up on the beamgrid, it is detailed in the following subsection. 3.2. Array Aliasing In (12), each intermodulation (IM) term maps to a beam index determined by adding or subtracting the corresponding precoder beam indices. For example, the intermodulation term (ω(b2)x2)2(ω(b1)x1)*=(ω(b2)ω(b2)ω(b1)*)x22x1* indeed produces the intermodulation term x22x1* up on beam index b2+b2−b1=2+2−1=3. However, this simple index addition does not always yield a valid beam index in the range [1,L]. For instance, the IM product (ω(b1)x1)2(ω(b2)x2)* corresponds to the index 2b1−b2, which would fall outside [1,L]. Instead, this term is aliased back in to the beamgrid range [1,L] according to(13)bi1+bi2−bi3¯=bi1+bi2−bi3−L⌊bi1+bi2−bi3−1/2L⌋, for {bi1,bi2,bi3}∈[1,L], {i1,i2,i3}∈[1,…,K], and where the bar notation represents the aliasing operation. With this in mind, it can be noted that whilst the example of Section 3.1 produced 12 uniquely placed data streams of main beam and intermodulation beam data, aliasing can, for a different beam configuration {b1,b2,b3}, end up overlapping intermodulation terms to go in the same beam directions. This will lead to distortion being concentrated in fewer beam directions. 4. Virtual-Array DPD In this section, virtual-array DPD (VA-DPD) processing, as shown in Figure 2, is proposed to reduce the complexity compared to applying a DPD per PA. As previously described in Section 1.2, the VA-DPD aims to produce predistorted user streams and off-beam IM terms in a low-dimensional virtual array, and use these to linearize the larger, real transmit array. This processing conceptually follows four main steps. First, user streams are precoded to the antenna space of a virtual array, by placement on a beamgrid and through an FFT. The beam placement in the virtual array depends on the precoding to be done in the real array, which will be detailed later. Second, nonlinear DPDs are applied to the data on each of the Lv virtual antenna space branches. Third, an IFFT is performed, acting as a virtual channel and to separate user streams and off-beam IM terms in the virtual beam space. Fourth, and last, the predistorted streams that have been generated are precoded to the real array. The mapping in the virtual domain is chosen such that the predistorted terms generated by the virtual array match those needed to cancel the distortion produced in the physical array. In the following subsections, the derivations of Section 3.1 will be used as a starting point to give mathematical motivation as to how a smaller virtual array, in place of the original array, can be used for performing DPD. This is presented first with an example comparison to the per-PA case, followed by presenting special beam configuration cases in which the complexity can be reduced even further. Finally, generalizations to more general channels and higher-order modeling are discussed. 4.1. Virtual-Array DPD Example for Three Concurrent Users Revisiting the example of Section 3.1, consider again the beam indices {b1,b2,b3}={1,2,5}, but now replace the physical array size L=16 with a virtual-array size of Lv=13. The processing steps then follow directly from the VA-DPD structure of Figure 2: precoding into the virtual antenna space, nonlinear DPD processing, and transformation into the virtual beam domain via an IFFT. This yields(14)LvWDFT*f(WDFTx)=θ1x1+1Lvθ3x1(|x1|2+|x2|2+|x3|2)θ1x2+1Lvθ3x2(|x1|2+|x2|2+|x3|2)1Lvθ3x22x1*1Lvθ3x1x3x2*θ1x3+1Lvθ3x3(|x1|2+|x2|2+|x3|2)1Lvθ3x2x3x1*01Lvθ3x32x2*1Lvθ3x32x1*1Lvθ3x12x3*1Lvθ3x1x2x3*1Lvθ3x22x3*1Lvθ3x12x2*. instead of (12). Since Lv