WorksEngineering
Geometry-Aware Channel Aging in MIMO Systems: When Is Kronecker Separability Enough?
Abstract
This letter investigates when Kronecker-separable spatial correlation approximates the covariance of multiple-input-multiple-output (MIMO) channels. We propose a geometry-aware channel model in which the same propagation environment determines both the Weischelberger spatial coupling matrix and the temporal correlation coefficient. The propagation environment follows a 3GPP-inspired clustered channel model. We consider a MIMO system affected by symbol-level channel aging caused by the user equipment's mobility. We schedule both uplink and downlink pilots, thereby deriving the respective spectral efficiencies (SEs) in the presence of imperfect channel state information. Numerical results show that channel aging's magnitude is strongly geometry-dependent, with diffuse propagation inducing greater temporal decorrelation and SE degradation than concentrated scattering. The results also identify regimes where Kronecker separability can be applied and regimes where geometry-aware Weischelberger coupling is needed.
Introduction
Multiple-input multiple-output (MIMO) systems rely on accurate channel state information (CSI) to perform multi-user (MU)-precoding and interference suppression. In practice, CSI is acquired through pilot transmission and becomes progressively outdated over time due to user equipments (UEs) mobility and the dynamic evolution of the propagation environment. This phenomenon, commonly referred to as channel aging [1], becomes prevalent in high-mobility scenarios, where the channel variations within the frame may become non-negligible [2], resulting in a degradation of the achievable spectral efficiency (SE). The channel’s temporal correlation decay is often modeled using Jake-like isotropic scattering models, which produce correlation coefficients mainly determined by the UE velocity magnitude [3]. Though mathematically tractable, this model fails to capture the effect of the propagation geometry, as UEs with the same velocity magnitude may exhibit different correlation decays depending on the positions of their dominant multipath clusters.
The propagation geometry is represented by the channel spatial covariance matrix. Thanks to its mathematical tractability, the Kronecker-separable model is among the most widely adopted [4]. This model represents the MIMO channel covariance as the product of one-sided base station (BS) and UE correlation matrices. This separability assumption neglects the joint correlation between BS and UE, which is especially important in MU-MIMO networks. The Weischelberger [5] model overcomes this limitation by introducing a coupling matrix that describes how power is distributed across BS-UE eigenmode pairs [6]. Nevertheless, this coupling is often treated as an abstract entity rather than being explicitly linked to the propagation geometry [7].
The need for a unified framework where the same propagation environment gives rise to both the spatial and the temporal correlation of the MIMO channel motivates the main question in this letter: When is the Kronecker-separable model sufficiently accurate, and when is joint-correlation coupling modeling necessary to capture the propagation geometry?
Contributions
This letter proposes a geometry-aware MIMO channel-aging model that links spatial eigenmode coupling and temporal correlation through the same propagation geometry. The propagation geometry is inspired by the 3GPP-standardized cluster delay line channel model. The propagation environment is indeed made of propagation clusters, whose positions are randomly generated, even though it is worth mentioning that the same framework could be driven by ray tracing in practical deployments We construct a Weischelberger coupling matrix by projecting the cluster-level covariance contributions onto the BS and UE eigenmodes. A cluster-dependent temporal correlation coefficient is then derived in which each cluster Doppler shift is weighted by its power, UE-side angle, and Doppler spread. To this end, we propose a Doppler-spread model that directly links the uncertainty in the angular domain to that in the Doppler domain. Finally, we evaluate the symbol-level aging impact in a time-division duplexing (TDD) MIMO system with explicit uplink (UL) and downlink (DL) pilots, covariance-based channel estimation, and imperfect-CSI SE expressions. Numerical results identify regimes where the Kronecker approximation is accurate and regimes where geometry-aware coupling is needed, showing that diffuse propagation induces stronger temporal decorrelation and larger SE losses than concentrated scattering.
System Model
We consider a MU-MIMO system, where a BS, located at the origin and equipped with antennas, serves UEs, equipped with antennas, located at . The BS performs orthogonal frequency division multiplexing
(OFDM) transmission around the carrier frequency , over active subcarriers with a subcarrier spacing of . The subcarriers are partitioned into blocks of subcarriers, indexed by the block index : the channel is assumed constant across subcarriers within the same block. Here is the symbol duration, with being the cyclic prefix duration. We define the symbol time index , with time unit . The frequency-domain UL communication channel between UE and BS is denoted by . TDD and UL/DL channel reciprocity are assumed, i.e., . We account for non-negligible time correlation between symbol times: the correlation decay is modelled as a first-order autoregressive (AR) process with a scaled identity state-transition matrix [1]. The channel’s time correlation is thus defined as , where captures the correlation decay. The frequency domain UL observation at the BS at time on a generic subcarrier belonging to block is defined as
where is the precoding matrix adopted by UE at time for all the subcarriers in block . Here, accounts for the path loss and large-scale fading of UE . Here are the transmitted symbols adopted by UE in its UL transmission. We assume that . The BS receiver noise is . Conversely, the frequency-domain DL observation at UE is defined as
where is the precoder matrix adopted for every subcarrier in the -th block. The transmitted symbols are defined as , with . Lastly, is UE receiver noise.
Weischelberger correlation model
We here take the Weichselberger Rayleigh fading model [5] as a reference and use it to study when the Kronecker-separable approximation is accurate. This model introduces an eigenmode coupling matrix, here denoted by , specifying how the channel power is distributed across BS–UE eigenmode pairs, while the Kronecker model is obtained as a separable special case, where the BS and UE eigenmodes are perfectly uncoupled. Let us define the one-sided correlation matrices of the UL channel between UE and BS as and , with their respective eigendecompositions and . The UL channel covariance is defined as
Where . The DL channel covariance follows by reciprocity:
For simplicity, the traces of both matrices are normalized to one. It should be noted that the coupling matrix yields a Kronecker-separable covariance model. The covariance mismatch between the Kronecker-separable model and the jointly-correlated one is measured as
Propagation geometry
The propagation environment between the BS and the UEs is composed of a set of scattering clusters. Cluster is centered at and is seen by the BS at the angle and at the angle by UE . The angles of the rays scattered by each cluster follow a normal distribution with side-dependent angle spreads, denoted by and . Assume that the size of each cluster is described by its effective radius . The angle spreads of those clusters are then computed as
The rationale behind this model is that closer clusters scatter rays within a wider angular interval than farther ones. The marginal covariances and are computed according to the model in [8]: under the assumption of small angle spreads, we can write , where
The definition of follows from the same model and is omitted for brevity. Here represents the power of cluster as seen from the BS; conversely, represents the cluster power seen by UE , while is the power of said cluster in the coupling matrix. Each of those powers is generated by a geometry-based adaptation of the standardized power delay profile [9]. The clusters BS side, UE side, and joint delays are defined as , , and , with denoting the speed of light. Each of those delays is then re-normalized such that the closest cluster has zero delay, i.e., . The cluster powers are then computed as [9 p. 54]
where and are the root mean square (RMS) delay spread and the delay proportionality factor. The per-cluster shadowing term is denoted by . These powers are then normalized such that , . We propose the following coupling matrix
Proposition 2. Let and denote the -th largest eigenvector of and the -th largest eigenvector of , then the geometry-based coupling matrix
This coupling matrix maps the propagation environment to eigenmode coupling: is large when both eigenmodes are excited by one or more clusters with high power. How much an eigenmode is excited by a cluster is measured by projecting the cluster contribution to the marginal covariance (e.g., ) onto the corresponding eigenvector (e.g., ). A Sinkhorn-like scaling is applied to [10] to match the row and column sums to the BS and UE marginal eigenvalues, thus ensuring a fair comparison with .
Temporal correlation
In our previous works, the temporal correlation coefficient followed a Jake’s model. While analytically convenient, this model averages different propagation directions and does not account for the fact that the UE’s Doppler shift is not isotropic, meaning that clusters at different positions age at different rates. Let denote UE velocity with , with indicating the velocity’s direction. therefore, the Doppler shift seen by cluster is defined as . Following the model presented in [11 Lemma 1], the correlation coefficient is
where represents the Doppler spread of said cluster, which can be related to the propagation geometry according to the following lemma.
Lemma 1. Let us consider a cluster seen by UE at an angle with an angle spread of . Assuming a Gaussian intra-cluster angular deviation, the cluster Doppler spread as seen by UE , denoted by , is defined as
where , , and .
Proof: Each ray scattered by cluster is seen by UE at an angle , where represents the deviation from the nominal angle. Then, the Doppler shift of ray is . Keeping in mind that and using the trigonometric identity we compute . Furthermore, by using the identity , we are able to compute . The lemma follows by computing
Channel estimation and spectral efficiency
Figure 1 shows the adopted pilot transmission pattern. Both UL and DL pilots are scheduled to provide to UEs and BS with estimated CSI for data detection and precoding. Each frame is composed of UL data slots and downlink ones, bringing the total number of data slots to . The precoding matrices adopted during pilot transmission, denoted by , , are assumed to be fixed and known by both parties1. Let us now focus on the UL phase: each UE transmits its pilot matrix in subcarriers; these matrices are mutually orthogonal, i.e., . The considerations that follow apply to all blocks; we thus omit the index for brevity. Let us consider the arbitrary pilot slot , the observed pilot matrix is . After pilot de-spreading, the vectorized observation is defined as
Figure 1. Representation of the pilot and data symbol transmission structure in the time-frequency domain. The communication channel is estimated at the BS and at each UE thanks to the respective UL and DL pilots.
The receiver uses pilot observations before the present one, bringing the total number of observations used for channel estimation to : We adopt the minimum mean square error (MMSE) channel estimation presented in [13], that is
where
The channel estimate is . On the other hand, The channel estimation error is distributed as . We now derive achievable UL and DL SE, assuming imperfect CSI and known precoders. We consider a generic subcarrier belonging to the block . The BS adopts the arbitrary combining matrix for the detection
of , while UE adopts the combining matrix to detect . Assuming successive interference cancellation, UE’s UL SE at data symbol time is [6 Cor. 1]
where and is defined in (16). Here has been obtained by stacking into a matrix. The DL channel estimator is not explicitly reported for brevity, as it follows the same covariance-based approach of (15) and is reported in [11]. Assuming known BS precoders at the UEs, the DL SE can be modelled as
where and is defined in (17). Therefore, the average UL SE, and analogously the DL SE, over the frame, aggregated over the subcarriers and UEs, is defined as .
Precoding strategy
The adopted precoding strategy differs between data slots, where CSI is available, and pilot slots, where the previous estimate has drifted too far from its actual value. In the absence of CSI, the BS implements eigen-beamforming, that is , while . In the presence of CSI, the BS adopts MMSE precoding [14 Ch. 5] while the UEs precoders are , where is the left singular matrix of . Precoder[s] are normalized to ensure an equal power allocation between subcarriers and streams. We compare zero-forcing (ZF) and MMSE combining at the BS: the ZF combining matrix is , with , while the MMSE combining matrix is defined in (18). The UEs adopt the combining matrix , defined in (17).
Numerical results
The default value of the simulation parameters is reported in Table I. The effective radii are drawn independently from the log-normal distribution shown in Table I: is clipped between 2 and 30 m to avoid unrealistically large radii. The resulting angle spreads are then also clipped between and . The number of clusters is divided into dominant clusters and background one[s]; this allows us to control the propagation environment by
| Parameter | value |
| BS and UE antennas | 16, 4 |
| UE positions | m |
| UE pathloss | 3GPP street canyon model [15] |
| UE velocities | m/s |
| Frequency blocks, | 2 |
| Subc. per block | 20 |
| Carrier frequency, Symbol time | 4 GHz, 53 s |
| UL and DL data slots | 5 |
| UL and DL pilot length | 20 |
| BS and UEs total power | 30, 20 dBm |
| Number of clusters | 25 |
| -th cluster effective radius | |
| RMS delay spread | 150 s |
| proportionality constant | 2.4 |
| Shadowing variance | 3 |
| UEs and BS noise variance | dB |
Table I. Simulation parameters default values
tuning . The dominant clusters are divided among the UEs as and their position is randomly generated as . Here , while is a unit vector orthogonal to . On the other hand, the background clusters positions , with , and . The proposed Weiscelberger model is denoted by “Geo”, while the Kronecker-separable baseline is labeled as “Kron”. Figure 2 shows the empirical Cumulative distribution function (CDF) of the covariance mismatch between the Kronecker separable model and the jointly correlated one: the solid curves show the nominal geometry while the dashed ones represent setups where the dominant-clusters centroid is shifted by . We observe that the accuracy of a Kroneckerseparable approximation is strongly favored by a concentrated scattering environment (i.e., ), exhibiting a 50% probability of being below 25%. A dominant-cluster centroid shift makes the Kronecker separable approximation much less accurate and less dependent on , as the gap between the dotted curves shrinks substantially. This indicates that the Kronecker accuracy is favored by a strongly concentrated propagation environment where the clusters lie on the line between BS and UE. We now move to investigating the impact of a Kronecker-separable approximation on the achievable SE: Figure 3 shows the average UL and DL SEs evolution for increasing UE mobility in diffusive (i.e. ) and concentrated (i.e. ) scattering conditions. Figs 3a and 3b show the UL SE with MMSE and ZF BS combining, respectively, while Fig. 3c shows the DL SE. Increasing degrades all the SEs due to an increased aging rate of the channel. A concentrated scattering environment generates higher SEs: concentrated scattering geometries produce smaller Doppler spreads and their is less sensitive to UE mobility. The DL SE in Fig. 3c is more sensitive to than the UL, as the UL SE benefits from centralized BS-side detection, and shows a low-to-moderate Geo-Kron gap. On the contrary, the UL SE is less sensitive to UE mobility and exhibits a large Geo-Kron gap. Interestingly, with ZF BS combining, the Geo-Kron gap becomes smaller in concentrated propagation environments.

Figure 2. CDF of the covariance mismatch between the Kronecker separable model and the Jointly correlated one. The CDF is computed over 500 independent geometric setups.

Figure 3. Average SE in the UL, for both MMSE (a) and ZF (b) BS combining, and in the DL (c). We obtain the reported curves by averaging 20 Monte Carlo realizations of the propagation environment; for each environment, we average the SE over 500 channel and precoder realizations.
Conclusions
We proposed a geometry-aware channel model where the same clustered propagation environment determines both the Weichselberger eigen-mode coupling and the temporal correlation decay. Channel aging is now determined by the UE motion relative to the scattering geometry, rather than its velocity magnitude alone. Using explicit UL and DL pilots, we evaluated channel estimation and imperfect-CSI SE. The results show that Kronecker separability is an adequate approximation in highly concentrated propagation regimes, but can miss relevant eigenmode coupling in diffuse environments. Moreover, diffuse propagation induces stronger Doppler diversity, causing stronger channel aging and larger SE degradation at high UE velocities. This motivates eigenmode-coupled models for mobility-sensitive MIMO and ISAC systems.
References
- [1]S. Fodor, G. Fodor, D. Gürgünoğlu, and M. Telek, “Optimizing pilot spacing in MU-MIMO systems operating over aging channels,” IEEE Transactions on Communications, vol. 71, no. 6, pp. 3708–3720, 2023.arxiv.org/abs/2204.00213
- [2]K. T. Truong and R. W. Heath, “Effects of channel aging in massive MIMO systems,” Journal of Communications and Networks, vol. 15, no. 4, pp. 338–351, 2013.arxiv.org/abs/1305.6151
- [3]C.-X. Wang and M. Patzold, “Efficient simulation of multiple cross-correlated Rayleigh fading channels,” in 14th IEEE Proceedings on Personal, Indoor and Mobile Radio Communications, 2003. PIMRC 2003., vol. 2. IEEE, 2003, pp. 1526–1530.DOI
- [4]M. Bengtsson and P. Zetterberg, “Some notes on the Kronecker model,” EURASIP Journal on Wireless Communications and Networking, 2006.
- [5]W. Weichselberger, M. Herdin, H. Özcelik, and E. Bonek, “A stochastic mimo channel model with joint correlation of both link ends,” IEEE Transactions on wireless Communications, vol. 5, no. 1, pp. 90–100, 2006.DOI
- [6]Z. Wang, J. Zhang, B. Ai, C. Yuen, and M. Debbah, “Uplink performance of cell-free massive MIMO with multi-antenna users over jointly-correlated Rayleigh fading channels,” IEEE Transactions on Wireless Communications, vol. 21, no. 9, pp. 7391–7406, 2022.arxiv.org/abs/2110.04962
- [7]K. Dovelos, M. Matthaiou, H. Q. Ngo, and B. Bellalta, “Massive mimo with multi-antenna users under jointly correlated ricean fading,” in ICC 2020-2020 IEEE International Conference on Communications (ICC). IEEE, 2020, pp. 1–6.DOI
- [9]M. Series, “Guidelines for evaluation of radio interface technologies for imt-2020,” Report ITU, vol. 2512, no. 0, 2017.
- [10]R. Sinkhorn and P. Knopp, “Concerning nonnegative matrices and doubly stochastic matrices,” Pacific Journal of Mathematics, vol. 21, no. 2, pp. 343–348, 1967.
- [11]S. Rivetti, G. Fodor, E. Björnson, and M. Skoglund, “On the impact of channel aging and doppler-affected clutter on OFDM ISAC systems,” arXiv preprint arXiv:2601.05032, 2026.arxiv.org/abs/2601.05032
- [12]E. B. Kama, J. Kim, and E. Björnson, “Downlink pilots are essential for cell-free massive MIMO with multi-antenna users,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2024, pp. 1–6.arxiv.org/abs/2404.18516
- [13]S. Rivetti, G. Fodor, E. Bjornson, and M. Skoglund, “Cram”er-rao bound analysis of bistatic isac under partial symbol knowledge and clutter,” arXiv preprint arXiv:2604.11426, 2026.arxiv.org/abs/2604.11426
- [14]Ö. T. Demir, E. Björnson, L. Sanguinetti et al., “Foundations of user-centric cell-free massive MIMO,” Foundations and Trends® in Signal Processing, vol. 14, no. 3-4, pp. 162–472, 2021.DOI
- [15]European Telecommunications Standards Institute (ETSI), “Study on channel model for frequencies from 0.5 to 100 GHz: 3GPP TR 38.901 version 15.0. 0 release 15,” 2018.
