How to Mobilize mmWave: A Joint Beam and Channel Tracking Approach

by   Jiahui Li, et al.

Maintaining reliable millimeter wave (mmWave) connections to many fast-moving mobiles is a key challenge in the theory and practice of 5G systems. In this paper, we develop a new algorithm that can jointly track the direction and channel coefficient of mmWave beams in phased antenna arrays. Despite the significant difficulty in this problem, our algorithm can simultaneously achieve fast tracking speed, high tracking accuracy, and low pilot overhead. In static scenarios, this algorithm can converge to the minimum Cramér-Rao lower bound (CRLB) of beam tracking with high probability. In dynamic scenarios, this algorithm greatly outperforms existing algorithms in simulations. Even at SNRs as low as 5dB, our algorithm is capable of tracking a mobile moving at an angular velocity of 5.45 degrees per second and achieving over 95% of channel capacity with a 32-antenna phased array, by inserting only 10 pilots per second.


page 1

page 3

page 5


Super Fast Beam and Channel Tracking in 2D Phased Antenna Arrays

Millimeter wave (mmWave) is an attractive candidate for high-speed mobil...

Robust Adaptive Beam Tracking for Mobile Millimetre Wave Communications

Millimetre wave (mmWave) beam tracking is a challenging task because tra...

Fast Accurate Beam and Channel Tracking for Two-dimensional Phased Antenna Arrays

The sparsity and the severe attenuation of millimeter-wave (mmWave) chan...

Fast Accurate Beam and Channel Tracking for Two-dimensional Phased Antenna Array

The sparsity and the severe attenuation of millimeter-wave (mmWave) chan...

Joint Beam and Channel Tracking for Two-dimensional Phased Antenna Arrays

Millimeter wave (mmWave) is an attractive candidate for high-speed mobil...

Near-field Tracking with Large Antenna Arrays: Fundamental Limits and Practical Algorithms

Applications towards 6G have brought a huge interest towards arrays with...

A Robotic Antenna Alignment and Tracking System for Millimeter Wave Propagation Modeling

In this paper, we discuss the design of a sliding-correlator channel sou...

I Introduction

Millimeter-wave (mmWave) communication is promising to support the vastly growing data traffic for future wireless systems [1, 2, 3]. In the mmWave band, only several distinctive propagation paths exist, i.e., the line-of-sight path and a few relatively strong reflected paths [4, 5]. Therefore, the directional beamforming with large antenna arrays is necessary to provide sufficiently strong received signal power.

To overcome the hardware limitation on the number of radio frequency (RF) chains with large array size and high carrier frequency, analog beamforming with phased antenna arrays was proposed [6, 7, 8, 9, 3]. A phased array can receive the signal that is projected onto a certain spatial subspace, with a cost of requiring much more pilots than the fully digital arrays to find the rare and precious paths. When users move quickly, it is needed to track the dynamic paths and even more pilots are required. Hence, one fundamental challenge is how to accurately track a large number of dynamic paths from many high-mobility terminals/reflectors using limited pilots, e.g., in V2V/V2I, high-speed railway, and UAV scenarios [10].

The compressed sensing based algorithms (e.g., [11, 12, 13]) were proposed for phased arrays, which can reduce pilot overhead and make beam direction acquisition faster. However, these algorithms are designed for static or quasi-static scenarios, and will encounter performance deterioration under high-mobility scenarios. To cope with high-mobility scenarios, the algorithms in [14, 15, 16]

use the prior information to track the dynamic beam directions. However, these solutions do not optimize the tracking scheme with the optimal training beamforming vectors, which leads to poor tracking accuracy.

Since the tracking of a large number of dynamic paths can be decoupled into tracking each path with low pilot overhead, we have proposed a beam tracking algorithm in [17, 18] to optimize both the training beamforming vectors and tracking scheme. However, it assumes known channel coefficients, while both channel coefficient and beam direction might be unknown and time-varying in a real mobile system. In this paper, we further develop a recursive beam and channel tracking (RBCT) algorithm to jointly track the dynamic beam direction and channel coefficient. In static scenarios, the Cramér-Rao lower bound (CRLB) of beam direction is derived, which is a function of the training beamforming vectors. We also obtain the minimum CRLB by optimizing these training beamforming vectors, and establish three theorems to verify that the RBCT algorithm can converge to the minimum CRLB with high probability. Simulations reveal that the RBCT algorithm can achieve much faster tracking speed, lower tracking error, and lower pilot overhead than several existing algorithms.

We use the following notations: is a matrix, is a vector, is a scalar. is the 2-norm of . , and are ’s transpose, Hermitian and inverse, respectively. denotes expectation and () obtains the real (imaginary) part. The natural logarithm of is .

Ii System Model

Consider a phased array in Fig. 1, where omnidirectional antennas are placed on a line, with a distance between two neighboring antennas. Each antenna is connected through a phase shifter to the same RF chain. In time-slot , the pilot symbols arrive at the array from an angle-of-arrival (AoA) . The channel vector is given by


where is the sine of the AoA , is the steering vector of the arriving beam, is the wavelength, and is the complex channel coefficient.

Fig. 1: System model.

To track the beam direction and channel coefficient simultaneously, at least two observations using different beamforming vectors are needed. Hence, we assume that two pilot symbols are applied in each time-slot. To receive the -th () pilot symbol, let be the beamforming vector in time-slot , denoted by


which is assumed to have the same form as the steering vector. Combining the output signals of the phase shifters yields


where is the pilot symbol that is known by the receiver, and is an i.i.d.

circularly symmetric complex Gaussian random variable. Given


, the conditional probability density function of

is given by


A beam and channel tracker determines the beamforming matrix

, and provides an estimate

of the channel coefficient and the sine of the AoA. Let be a beam and channel tracking policy. In particular, we consider the set of causal beam and channel tracking policies: The estimate of time-slot and the beamforming matrix of time-slot are determined by using the history of beamforming matrices and channel observations .

Iii Joint Beam and Channel Tracking Problem

Our goal is to develop a joint beam and channel tracking algorithm to minimize the beam tracking error. For any time-slot , the joint beam and channel tracking problem is given by


where the constraint ensures that is an un-biased estimate of .

Problem (5) is a constrained sequential control and estimation problem that is difficult to solve optimally, where the beamforming matrix is the control action. First, the system is only partially observed through the channel observation . Second, both the beamforming matrix and the estimate need to be optimized: On the one hand, the optimization of is a non-convex optimization problem of in (2), which is discussed in Section III-A. On the other hand, as will be discussed in Section V, the optimization of is also non-convex and has multiple local optimal estimates.

Iii-a Cramér Rao Lower Bound of Beam Tracking


Now, we try to establish a lower bound of the MSE in (5) in static scenarios, where the ground true of beam direction and channel coefficient is invariant for all time-slot , i.e., . Given the beamforming matrices of the first time-slots, the MSE in (5) is lower bounded by the CRLB as follows [19]:


where obtains the matrix element in row and column , and is the Fisher information matrix, i.e., [20]


where , , and . By optimizing the beamforming matrices on the RHS of (6), we obtain the minimum CRLB as below:


where because the linear additive property of Fisher information matrix [20], the optimal are the same, and from (7), we can get


Problem (III-A) is non-convex with respect to and , which makes it too hard to obtain the analytical solution. However, we can still use numerical method to find the solution, which yields the optimal beamforming matrix as below:


where , and when , is very close to . In Fig. 2, the optimal receiving beam directions are depicted by plotting vs. and , where , and the signal-to-noise ratio (SNR) is . It can be observed that is almost the same as and there are two symmetric optimal solutions. Therefore, we will set in the proposed RBCT algorithm in Section IV.

Fig. 2: Optimization of Problem (III-A) using numerical method.

Iv Recursive Beam and Channel Tracking

Fig. 3: Frame structure.

We propose a two-stage algorithm to approach the minimum CRLB in (III-A), which is given below:

Recursive Beam and Channel Tracking (RBCT):

1) Coarse Beam Sweeping: pilots are used successively (see Fig. 3). The beamforming vector to receive the -th observation is set as . Obtain the initial estimate by


where , , , the size of determines the estimation resolution, and .

2) Beam and Channel Tracking: In time-slot , two pilots are received at the beginning (see Fig. 3) using beamforming vectors and , given by


and the estimate is updated by (13) on the top of the page, where , , , and .

In Stage 1, the exhaustive sweeping is used, and the initial estimate is obtained in (11) by using the orthogonal matching pursuit method (e.g., [13]). This ensures that the initial beam direction is within the mainlobe set, i.e.,


In Stage 2, the recursive tracker is motivated by the following maximization likelihood problem:


where is subject to (2). We propose a two-layer nested optimization algorithm to find the solution of (15). In the outer layer, we use the stochastic Newton’s method to update the estimate , given by [19]


where is the Hessian matrix, can be calculated by using (7), is the step-size that will be specified later, and


with and . Plugging and (17) in (IV), we get (13). In the inner layer, it is equivalent to minimize the CRLB to update , i.e.,


which results in (12).


Different from the beam tracking algorithm in [17, 18], the RBCT algorithm uses two pilots and jointly updates the beam direction and channel coefficient in each time-slot.

V Asymptotic Optimality Analysis


There are multiple stable points for (13), which correspond to the local optimal estimates for Problem (5) [21]. Hence Problem (5) is non-convex for the estimate . To study these stable points, we rewrite (13) as follows:


where is defined in (20), is defined in (21), with , , , , and .

A stable point should satisfy: 1) , and 2) is a negative definite matrix. Let


denote the stable points set at time-slot . Then, we can verify as below:

  • When , we have Hence, .

  • From (20), we can get


    Then, the derivative can be obtained by (24). Similar to 1), the first term in (24) is when . Moreover, the partial derivative in the second term is equal to when . Therefore, when , we have


Note that except for the real direction , the antenna array gain is quite low at other local optimal stable points in . Hence, one key challenge is how to ensure that the RBCT algorithm converges to the real direction , instead of other local optimal stable points in .

In static beam tracking, where and , we adopt the diminishing step-sizes, given by [19, 21, 22]


where and . We use the stochastic approximation and recursive estimation theory [19, 21, 22] to analyze the RBCT algorithm.To support the more general joint beam and channel tracking scenario than [17, 18], three new theorems are developed to resolve the challenge mentioned above:

Theorem 1 (Convergence to Stable Points).

If is given by (26) with any and , then converges to a unique point within with probability one.


See the detailed proof in Appendix A. ∎

Hence, for general step-size parameters and in (26), converges to a stable point in .

Theorem 2 (Convergence to the Real Beam Direction ).

If (i) , (ii) is given by (26) with any , then there exist and such that


See the detailed proof in Appendix B. ∎

By Theorem 2, if the initial point is in the mainlobe , the probability that does not converge to decades exponentially with respect to . Hence, one can increase the transmit SNR and reduce the step-size parameter to ensure with high probability.

Theorem 3 (Convergence to with the Minimum MSE).

If (i) is given by (26) with and any , and (ii) , then


See the detailed proof in Appendix C. ∎

Theorem 3 tells us that should not be too small: If , then the minimum CRLB on the RHS of (III-A) is achieved asymptotically with high probability.

Vi Numerical Results

Fig. 4: MSE vs. time-slot number in static scenarios.

We compare the RBCT algorithm with three reference algorithms: the compressed sensing algorithm [13], the IEEE 802.11ad algorithm [14], and the beam tracking algorithm [18]. The first two algorithms have the same configuration as that in Section VI of [18]. The third one uses the same training beamforming vectors as the RBCT algorithm, i.e., in each time-slot, it receives two pilots with the beamforming vectors in (12), and the beam direction is tracked by using both observations. Moreover, its channel coefficient is obtained with a least square estimator by using these observations. Consider the system model in Section II with antennas, and the antenna spacing is . The pilot symbol is , and the transmit SNR is set as . To ensure fairness, we assume that 2 pilot symbols are received in each time-slot, hence all the algorithms have the same pilot overhead.

In static scenarios, we set the step-size as . The real AoA

is randomly generated by a uniform distribution on

in each realization, and the results are averaged over 10000 random realizations. Figure 4 plots the MSE over time. It can be observed that the MSE of the RBCT algorithm converges to the minimum CRLB in (III-A), which is much smaller than the reference algorithms.

Fig. 5: MSE vs. angular velocity in dynamic scenarios.
Fig. 6: Data rate vs. angular velocity in dynamic scenarios.

In dynamic scenarios, we set the step-size as a constant value, i.e., . The channel variation is modeled as: The AoA where , denotes the rotation direction, and is a fixed angular velocity. The rotation direction is chosen such that varies within . The channel coefficient is subject to Rician fading with a K-factor , according to the channel model proposed in [23]. In Fig. 5 and 6, one can observe that the RBCT algorithm can support much higher angular velocities and data rates than other algorithms. According to Fig. 6, the RBCT algorithm can achieve 95% of channel capacity when the angular velocity is 0.19rad (1.09 degrees) per time-slot. If 5 time-slots last for one second, i.e., 10 pilots per second received, then the RBCT algorithm is capable of tracking a mobile moving at an angular velocity of 5.45 degrees per second and achieving over 95% of channel capacity.

Vii Conclusion

We have developed a joint beam and channel tracking algorithm for mmWave phased antenna arrays, and established its convergence and asymptomatic optimality. Our simulation results show that the proposed algorithm can achieve much faster tracking speed, lower beam tracking error, and higher data rate than several state-of-the-art algorithms, with the same pilot overhead.


  • [1] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadband systems,” IEEE Commun. Mag., vol. 49, no. 6, Jun. 2011.
  • [2] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, “Five disruptive technology directions for 5G,” IEEE Commun. Mag., vol. 52, no. 2, Feb. 2014.
  • [3] R. W. Heath, N. González-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, “An overview of signal processing techniques for millimeter wave MIMO systems,” IEEE J. Sel. Top. Signal Process., Apr. 2016.
  • [4] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5G cellular: it will work!” IEEE Access, vol. 1, May 2013.
  • [5] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wideband millimeter-wave propagation measurements and channel models for future wireless communication system design,” IEEE Trans. Commun., vol. 63, no. 9, Sep. 2015.
  • [6] S. Sun, T. S. Rappaport, R. W. Heath, A. Nix, and S. Rangan, “MIMO for millimeter-wave wireless communications: Beamforming, spatial multiplexing, or both?” IEEE Commun. Mag., vol. 52, no. 12, Dec. 2014.
  • [7] S. Han, C. L. I, Z. Xu, and C. Rowell, “Large-scale antenna systems with hybrid analog and digital beamforming for millimeter wave 5G,” IEEE Commun. Mag., vol. 53, no. 1, Jan. 2015.
  • [8] A. Puglielli, A. Townley, G. LaCaille, V. Milovanović, P. Lu, K. Trotskovsky, A. Whitcombe, N. Narevsky, G. Wright, T. Courtade, E. Alon, B. Nikolić, and A. M. Niknejad, “Design of energy- and cost-efficient massive MIMO arrays,” Proc. IEEE, vol. 104, no. 3, Mar. 2016.
  • [9] A. F. Molisch, V. V. Ratnam, S. Han, Z. Li, S. L. H. Nguyen, L. Li, and K. Haneda, “Hybrid beamforming for massive MIMO-a survey,” IEEE Commun. Mag., vol. 55, no. 9, Sep. 2017.
  • [10] G. Brown, O. Koymen, and M. Branda, “The promise of 5G mmWave - How do we make it mobile?” Qualcomm Technologies, Jun. 2016.
  • [11] J. Wang, Z. Lan, C.-W. Pyo, T. Baykas, C.-S. Sum, M. A. Rahman, J. Gao, R. Funada, F. Kojima, H. Harada, and S. Kato, “Beam codebook based beamforming protocol for multi-Gbps millimeter-wave WPAN systems,” IEEE J. Sel. Areas Commun., vol. 27, no. 8, Oct. 2009.
  • [12] A. Alkhateeb, O. E. Ayach, G. Leus, and R. W. Heath, “Channel estimation and hybrid precoding for millimeter wave cellular systems,” IEEE J. Sel. Top. Signal Process., vol. 8, no. 5, Oct. 2014.
  • [13] A. Alkhateeb, G. Leusz, and R. W. Heath, “Compressed sensing based multi-user millimeter wave systems: How many measurements are needed?” in IEEE ICASSP, Apr. 2015.
  • [14] IEEE standard, “IEEE 802.11ad WLAN enhancements for very high throughput in the 60 GHz band,” Dec. 2012.
  • [15] J. Palacios, D. De Donno, and J. Widmer, “Tracking mm-Wave channel dynamics: Fast beam training strategies under mobility,” IEEE INFOCOM, 2017.
  • [16] X. Gao, L. Dai, Y. Zhang, T. Xie, X. Dai, and Z. Wang, “Fast channel tracking for Terahertz beamspace massive MIMO systems,” IEEE Trans. Veh. Technol., vol. 66, no. 7, Jul. 2017.
  • [17] J. Li, Y. Sun, L. Xiao, S. Zhou, and C. E. Koksal, “Analog beam tracking in linear antenna arrays: Convergence, optimality, and performance,” in 51st Asilomar Conference on Signals, Systems, and Computers, 2017.
  • [18] ——, “Super fast beam tracking in phased antenna arrays,” arXiv preprint arXiv:1710.07873, 2017.
  • [19] M. B. Nevel’son and R. Z. Has’minskii, Stochastic approximation and recursive estimation, 1973.
  • [20] H. V. Poor, An introduction to signal detection and estimation.   New York, NY, USA: Springer-Verlag New York, Inc., 1994.
  • [21] H. Kushner and G. G. Yin, Stochastic approximation and recursive algorithms and applications.   Springer Science & Business Media, 2003, vol. 35.
  • [22] V. S. Borkar, Stochastic approximation: a dynamical systems viewpoint, 2008.
  • [23] M. K. Samimi, G. R. MacCartney, S. Sun, and T. S. Rappaport, “28GHz millimeter-wave ultrawideband small-scale fading models in wireless channels,” in 2016 IEEE VTC Spring, May 2016.
  • [24] J. M. Holte, “Discrete Gronwall lemma and applications,” in MAA-NCS meeting at the University of North Dakota, vol. 24, 2009.

Appendix A Proof of Theorem 1

Before providing the proof, let us provide some useful definitions. In static beam tracking, where and , recall the recursive procedure (19):


where and are given in (20) and (21) separately. From (21), we have


where , and is the covariance matrix of calculated by (31). In (31), the step can be obtain as follows:

  • Since consists of two i.i.d. circularly symmetric complex Gaussian random variables, we get



  • By splitting the real part and imaginary part, we obtain

  • Combining (32), (33) and (34) yields