Online convex optimization with periodic updates for downlink multi-cell MIMO wireless network virtualization

Abstract

A method and network node for online coordinated multi-cell precoding are provided. According to one aspect, a method includes receiving from each of the plurality of service providers a virtual precoder matrix determined by the corresponding service provider. The method also includes determining a precoder matrix by minimizing a precoding deviation from a virtualization demand of the network subject to at least one power constraint, the virtualization demand being based at least in part on a product of a channel state matrix and a virtual precoder matrix. The method further includes applying the determined precoder matrix to signals applied to a plurality of antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

Description

TECHNICAL FIELD

This disclosure is related to wireless communication and in particular, to online convex optimization (OCO) with periodic updates for downlink multi-cell multiple input multiple output (MIMO) wireless network virtualization.

BACKGROUND

OCO has emerged as a promising solution to many signal processing and resource allocation problems in the presence of uncertainty. In OCO, at the beginning of each time slot, an agent makes a decision from a known convex set. At the end of each time slot, the system reveals the convex loss function to the agent, and then the agent's loss is realized. The convex loss function can change arbitrarily at each time slot with unknown statistics. It is thus not possible for an OCO algorithm to make an optimal decision at each time slot due to the lack of in-time information of the convex loss function. Thus, a goal of an OCO algorithm is to make a sequence of decisions to minimize the regret, i.e., the performance gap of the accumulated loss to an optimal offline fixed strategy with all convex loss functions known in hindsight. An OCO algorithm should provide sub-linear regret, i.e., the gap from the optimal offline performance benchmark, in terms of time-averaged loss, tends to zero as time approaches infinity. With the bounded gradient or sub-gradient assumption, an online projected gradient descent algorithm achieves an O(√T) regret in a seminal work of OCO, where T is the total time horizon.

OCO with long-term constraints has been considered, in which short-term constraints are relaxed to long-term constraints to circumvent the computational challenge of the projection operator. An OCO algorithm with long-term constraints should provide sub-linear long-term constraint violations, i.e., the time-averaged violations of the long-term constraints tend to zero as time approaches infinity.

Wireless network virtualization (WNV) as an extension of wired network virtualization to the wireless domain has been proposed to share a common network infrastructure among several virtual networks. In a virtualized wireless network, an infrastructure provider (InP) owns the network infrastructure and virtualizes the network resources into virtual slices. The service providers (SPs) lease these virtual slices and serve their subscribing users under their own management and requirements, without the need to be aware of the underlying physical network infrastructure. WNV reduces the capital and operational expenses of wireless networks and makes it easier and faster to migrate to new network technologies, while ensuring the existing services are unaffected. Different from wired network virtualization, WNV concerns the sharing of both the wireless infrastructure and the radio spectrum. Due to the broadcast and random nature of the wireless medium, spectrum sharing in WNV brings new challenges in guaranteeing the isolation of virtual networks.

Existing works on OCO with long-term constraints use the standard per-time-slot settings, where the decision is made in each time slot and the convex loss function is revealed at the end of each time slot. OCO with long-term constraints has been considered in which a saddle point algorithm was proposed to achieve lower computational complexity than another known method and provides O(√T) regret and O(T{circumflex over ( )}(¾)) long-term constraint violation. A trade-off between the regret and the long-term constraint violation has been shown in a follow-up work. Variations of the online saddle point methods have been studied for distributed online problems with consensus and proximity constraints. A virtual-queue-based algorithm has been proposed and provides O(√T) regret and O(1) long-term constraint violation. The above works consider time-invariant constraints that are known in advance. Virtual-queue-based OCO algorithms have been proposed for time-varying constraints with static and dynamic regret analysis. A modified online saddle point algorithm has been proposed to deal with time-varying constraints. However, many practical systems do not allow a decision to be made at each time slot, while the system itself evolves over time.

While the known methods assume that all convex loss function feedbacks are delayed for one time slot, there are existing works on OCO with multiple-time-slot or adversarial feedback delays. However, all of these works assume decisions are made at each time slot. Algorithms dealing with a known fixed feedback delay have been studied. The standard online gradient descent algorithm has been extended to deal with a known fixed feedback delay. A delay-adaptive online gradient descent algorithm has been proposed to accommodate adversarial feedback delays, under the assumption that the feedback delays do not change the order of the received feedbacks. A unified framework has been proposed to extend the standard OCO algorithms with full-information feedbacks to handle adversarial feedback delays. Regret bound analysis of online gradient descent algorithm with adversarial feedback delays has been provided.

Inherited from wired network virtualization, most existing MIMO WNV works enforce strict physical isolation. The system throughput and energy efficiency maximization problems have been studied for WNV in orthogonal frequency division multiplexing (OFDM) massive MIMO systems. The cloud radio networks (C-RAN) and non-orthogonal multiple access (NOMA) techniques have been combined with uplink MIMO WNV in some known methods. Exclusive sub-channels are allocated to each SP through a two-level hierarchical auction architecture and antennas are allocated through pricing among the SPs for virtualized massive MIMO systems in some known works. The physical isolation approach does not fully utilize the benefit of spatial spectrum sharing enabled by MIMO beamforming. In contrast, all antennas and channel resources are shared among the SPs in some works, through stochastic robust precoding for massive MIMO WNV.

Although the above MIMO WNV works focus on offline problems, the Lyapunov optimization technique and OCO technique have been utilized in various online studies for non-virtualized MIMO systems. Under the framework of Lyapunov optimization, online power control for wireless transmission with energy harvesting and dynamic transmit covariance design for point-to-point MIMO systems have been studied. Online projected gradient descent and matrix exponential learning have been used for MIMO uplink covariance matrix design. Online MIMO WNV schemes under accurate and inaccurate CSI have been studied without considering channel sate feedback delay. Furthermore, known methods all are under the standard per-time-slot settings.

Under standard OCO settings, decision making and convex loss function feedback are in a strict per-time-slot fashion, which is restrictive for many real applications. In a practical long term evolution (LTE) network, there is only one reference or pilot symbol inserted in every six or seven transmission symbols to allow channel estimation. MIMO precoding relies on channel state feedback and is designed every six or seven symbol durations, while the underlying channel state keeps varying over those symbol durations. Existing OCO algorithms cannot be directly applied to such a system, where decisions and system states are updated per period that could last for multiple time slots.

SUMMARY

Some embodiments advantageously provide a method and system for online convex optimization with periodic updates for downlink multi-cell multiple input multiple output (MIMO) wireless network virtualization.

An OCO algorithm with periodic updates is provided. The algorithm is applied to an online downlink multi-cell MIMO WNV problem, where precoding design and channel state feedback are in a per-update-period fashion.

Also provided is a way to apply the OCO algorithm to wireless network virtualization (WNV) in the downlink coordinated multi-cell setting. In some embodiments, in the most general setting, an agent makes a decision, taken from a known convex decision space, at the beginning of each update period that can last for multiple time slots. The gradient or sub-gradient information of some convex loss functions is then revealed to the agent at the end of each update period. The convex loss function can change arbitrarily at each time slot with unknown statistics. The gradient or sub-gradient feedbacks may be delayed for multiple time slots, received out of order, and partly missing within each update period. The sequence of decisions is subject to both long-term and short-term constraints. Based only on the past gradient or sub-gradient information, the agent makes a sequence of decisions to provide updates to sub-linear regret and long-term constraint violation parameters.

In contrast to known methods, some embodiments of the algorithm allow gradient or sub-gradient feedbacks to be delayed for multiple time slots, received out of order and partly missing within each update period, while a decision is made only at the beginning of each update period.

In some embodiments, SPs are allowed to share all antennas and wireless spectrum resources simultaneously through the spatial virtualization approach.

In some embodiments, an online downlink coordinated multi-cell MIMO WNV algorithm is provided, where under the general setting a precoding design and channel state feedback are in a per-update-period fashion which, for example, is used in standard Long-Term Evolution (LTE).

In some embodiments, the OCO algorithm is implemented to design an online coordinated multi-cell precoding scheme at the infrastructure provider (InP) for downlink MIMO WNV, in which the wireless network is comprised of multiple service providers (SP) with unknown channel distribution information (CDI) and delayed channel state information (CSI). In the WNV framework discussed herein, in some embodiments, the InP communicates the corresponding channel state to each service provider (SP) after obtaining the global channel state. Each SP is allowed to utilize all antennas and wireless spectrum resources provided by the InP and design its own virtual precoding matrix in each cell, without the need to be aware of its own users in other cells and the users of other SPs. In some embodiments, at the beginning of each demand-response period, the InP coordinates the cells at the precoding level to meet the current and future virtualization demands from the SPs, based only on the past channel states and virtual precoding matrices. The InP aims to minimize the accumulated precoding deviation from the virtualization demand, accommodating both long-term and short-term transmit power constraints.

An OCO algorithm with periodic updates is described below. The OCO algorithm makes a decision at the beginning of each update period that can last for multiple time slots. The convex loss function can change arbitrarily at each time slot with unknown statistics. The gradient or sub-gradient feedbacks of the convex loss functions are allowed to be delayed for multiple time slots, received out of order, and partly missing within each update period. A sequence of decisions is made based only on the past gradient or sub-gradient information, subject to both long-term and short-term constraints. In some embodiments, the OCO algorithm provides O(√T) regret defined with partial feedbacks and O(1) long-term constraint violation.

An online downlink coordinated multi-cell MIMO WNV algorithm, with unknown CDI and delayed CSI under standard LTE transmission frame settings is disclosed. In the WNV framework, the InP may communicate the corresponding channel state to each SP after receiving the global channel state feedback. Each SP is allowed to utilize all antennas and wireless spectrum resources and design its own virtual precoding matrix in each cell, without the need to be aware of its own users in other cells and the other SPs. At the beginning of each demand-response period, in some embodiments, the InP coordinates the cells at the precoding level to meet the current and future virtualization demands based on the past channel states and virtual precoding matrices. In some embodiments, the InP optimizes MIMO precoding to minimize the accumulated precoding deviation from the virtualization demand, accommodating both long-term and short-term transmit power constraints. In such a case, the online coordinated multi-cell precoding problem has a distributed closed-form solution with low computational complexity.

Extensive simulation studies have been performed to validate the performance of the proposed online downlink coordinated multi-cell MIMO WNV algorithm under standard LTE network settings. Simulation studies show a fast convergence of time-averaged performance. The algorithm is able to track channel variations and is applicable to massive MIMO. Comparisons with an optimal offline fixed precoding strategy and OCO algorithm with long-term constraints under the standard settings demonstrate system performance advantage of the method proposed herein. Performance studies under different demand-response periods, precoding strategies adopted by the SPs, long-term transmit power limits, numbers of cells and antennas, and per-antenna transmit power constraints are also provided.

According to one aspect, a method of online coordinated multi-cell precoding for a network node configurable at least in part by an infrastructure provider is provided. The network node is configured to facilitate sharing of wireless network infrastructure resources by a plurality of service providers. The method includes obtaining a global channel state, the global channel state being based at least in part on a channel state in each cell of a plurality of cells and communicating a channel state to a corresponding service provider. The method also includes receiving a virtual precoding matrix from each service provider of the plurality of services providers, each virtual precoding matrix being based at least in part on the channel state communicated to the corresponding service provider and being further based at least in part on a condition that each service provider of the plurality of services providers is allowed to use all of a plurality of available antennas and available wireless spectrum resources. The method further includes executing an optimization procedure to periodically determine a multiple input multiple output, MIMO, precoding matrix that minimizes an accumulated precoding deviation from a virtualization demand subject to constraints, the virtualization demand and the MIMO precoding matrix being based at least in part on the virtual precoding matrices received from the plurality of service providers. The method includes applying the determined MIMO precoding matrix to signals applied to the plurality of available antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of the throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

According to this aspect, in some embodiments, the optimization procedure includes periodically determining a gradient of a convex loss function and providing the determined gradient as feedback to the determination of the MIMO precoding matrix. In some embodiments, feedback in the optimization procedure is allowed to be delayed for multiple time slots, received out of order and/or partly missing within an update period. In some embodiments, the optimization procedure includes an online projected gradient ascent algorithm that provides O(√{square root over (T)}) regret and O(1) long term constraint violation, where T is a total time horizon over which multiple updates of the determined MIMO precoding matrix occur. In some embodiments, the accumulated precoding deviation is determined according to f_t(x_t) where T is a time horizon, x_t is a decision in a sequence of decisions made by the network node, f_t(x_t) is a convex loss function, f_t(x_t) is an accumulated loss and x^o arg min_x∈X0f_t(x) is the argument of f_t(x) that produces a minimum value of f_t(x). In some embodiments, the method further includes dividing a total time horizon T into update periods, each update period having a duration of T_o time slots, T_o being at least one time slot, and updating the MIMO precoding matrix at a beginning or end of each update period. In some embodiments, at a beginning of each update period, a decision is taken from a known convex decision space and a loss is determined by an end of the duration of T_o time slots based at least in part on the decision, the loss being based at least in part on a convex loss function. In some embodiments, the virtualization demand is further based at least in part on past channel states. In some embodiments, the constraints include long term transmit power constraints and short term transmit power constraints. In some embodiments, the MIMO precoding matrix is determined to provide sub-linear T-slot regret with partial feedback on the accumulated precoding deviation from the virtualization demand, where T is a total time horizon.

According to another aspect, a network node configured for online coordinated multi-cell precoding, the network node configurable at least in part by an infrastructure provider is provided. The network node is configured to facilitate sharing of wireless network infrastructure resources by a plurality of service providers. The network node includes processing circuitry configured to: obtain a global channel state, the global channel state being based at least in part on a channel state in each cell of a plurality of cells and communicate a channel state to a corresponding service provider. The processing circuitry is further configured to receive a virtual precoding matrix from each service provider of the plurality of services providers, each virtual precoding matrix being based at least in part on the channel state communicated to the corresponding service provider and being further based at least in part on a condition that each service provider of the plurality of services providers is allowed to use all of a plurality of available antennas and available wireless spectrum resources. The processing circuitry is further configured to execute an optimization procedure to periodically determine a multiple input multiple output, MIMO, precoding matrix that minimizes an accumulated precoding deviation from a virtualization demand subject to constraints, the virtualization demand and the MIMO precoding matrix being based at least in part on the virtual precoding matrices received from the plurality of service providers. The processing circuitry is further configured to apply the determined MIMO precoding matrix to signals applied to the plurality of available antennas to achieve a sum of throughputs for plurality of service providers that is greater than a sum of the throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

According to this aspect, in some embodiments, the optimization procedure includes periodically determining a gradient of a convex loss function and providing the determined gradient as feedback to the determination of the MIMO precoding matrix. In some embodiments, feedback in the optimization procedure is allowed to be delayed for multiple time slots, received out of order and/or partly missing within an update period. In some embodiments, the optimization procedure includes an online projected gradient ascent algorithm that provides O(√{square root over (T)}) regret and O(1) long term constraint violation, where T is a total time horizon over which multiple updates of the determined MIMO precoding matrix occur. In some embodiments, the accumulated precoding deviation is determined according to f_t(x_t) where T is a time horizon, x_t is a decision in a sequence of decisions made by an agent, f_t(x_t) is a convex loss function, f_t(x_t) is an accumulated loss and x^o arg min_x∈X0f_t(x) is the argument of f_t(x) that produces a minimum value of f_t(x). In some embodiments, the processing circuitry is further configured to divide a total time horizon T into update periods, each update period having a duration of T_o time slots, T_o being at least one timeslot, and updating the MIMO precoding matrix at a beginning or end of each update period. In some embodiments, at a beginning of each update period, a decision is taken from a known convex decision space and a loss is determined by an end of the duration of T_o time slots based at least in part on the decision, the loss being based at least in part on a convex loss function. In some embodiments, the virtualization demand is further based at least in part on past channel states. In some embodiments, the constraints include long term transmit power constraints and short term transmit power constraints. In some embodiments, the MIMO precoding matrix is determined to provide sub-linear T-slot regret with partial feedback on the accumulated precoding deviation from the virtualization demand, where T is a total time horizon.

According to yet another aspect, a network node configured for sharing of wireless network resources among a plurality of service providers is provided. The network node includes processing circuitry configured to receive from each of the plurality of service providers a virtual precoder matrix determined by the corresponding service provider. The processing circuitry is further configured to determine a precoder matrix by minimizing a precoding deviation from a virtualization demand of the network subject to at least one power constraint, the virtualization demand being based at least in part on a product of a channel state matrix and a virtual precoder matrix. The processing circuitry is further configured to apply the determined precoder matrix to signals applied to a plurality of antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

According to this aspect, in some embodiments, determining the precoder matrix includes solving a minimization problem that includes determining a gradient of a convex loss function. In some embodiments, determining the precoder matrix includes decomposing the minimization problem into a number of subproblems, each subproblem involving a gradient of a convex loss function of a local precoding matrix. In some embodiments, the minimization problem includes comparing a function of the determined precoder matrix to an offline fixed precoding strategy. In some embodiments, the virtual precoder matrices received from the plurality of service providers are each based at least in part on a condition that each service provider is allowed to use all of a plurality of available antennas and wireless spectrum resources. In some embodiments, the precoding deviation from virtualization demand is determined at least in part by:f_t(V)∥H_tV−D_t∥_F² where V is a past precoding matrix, H_t is the channel state and D_t is the virtualization demand.

According to another aspect, a method of online coordinated multi-cell precoding for a network node configurable at least in part by an infrastructure provider is provided, where the network node is configured to facilitate sharing of wireless network infrastructure resources by a plurality of service providers. The method includes receiving from each of the plurality of service providers a virtual precoder matrix determined by the corresponding service provider. The method includes determining a precoder matrix by minimizing a precoding deviation from a virtualization demand of the network subject to at least one power constraint, the virtualization demand being based at least in part on a product of a channel state matrix and a virtual precoder matrix. The method also includes applying the determined precoder matrix to signals applied to a plurality of antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

According to this aspect, in some embodiments, determining the precoder matrix includes solving a minimization problem that includes determining a gradient of a convex loss function. In some embodiments, determining the precoder matrix includes decomposing the minimization problem into a number of subproblems, each subproblem involving a gradient of a convex loss function of a local precoding matrix. In some embodiments, the minimization problem includes comparing a function of the determined precoder matrix to an offline fixed precoding strategy. In some embodiments, the virtual precoder matrices received from the plurality of service providers are each based at least in part on a condition that each service provider is allowed to use all of a plurality of available antennas and wireless spectrum resources. In some embodiments, the precoding deviation from virtualization demand is determined at least in part by:f_t(V)∥H_tV−D_t∥_F² where V is a past precoding matrix, H_t is the channel state and D_t is the virtualization demand.

BRIEF DESCRIPTION OF THE DRAWINGS

A more complete understanding of the present embodiments, and the attendant advantages and features thereof, will be more readily understood by reference to the following detailed description when considered in conjunction with the accompanying drawings wherein:

FIG. 1 shows periodic updates of an OCO algorithm;

FIG. 2 shows virtual networks;

FIG. 3 is a timeline illustrating online precoding design for downlink MIMO WNV;

FIG. 4A is a graph of ρ(T) versus a time horizon, T;

FIG. 4B is a graph of P(T) versus a time horizon T;

FIG. 4C is a graph of R(T) versus a time horizon T;

FIG. 5A is a graph of ρ(T) versus a time horizon, T;

FIG. 5B is a graph of P(T) versus a time horizon T;

FIG. 5C is a graph of R(T) versus a time horizon T;

FIG. 6A is a graph of ρ(T) versus a number of antennas;

FIG. 6B is a graph of R(T) versus a number of antennas;

FIG. 7A is a graph of ρ(T) of the disclosed algorithm versus an offline algorithm;

FIG. 7B is a graph of R(T) of the disclosed algorithm versus an offline algorithm;

FIG. 8A is a graph of ρ(T) versus the long-term transmit power limit P_c;

FIG. 8B is a graph of R(T) versus the long term transmit power limit P_c;

FIG. 9A is a graph of ρ(T) versus a number of cells;

FIG. 9B is a graph of R(T) versus a number of cells;

FIG. 10A is a graph of ρ(T) versus a time horizon, T;

FIG. 10B is a graph of P(T) versus a time horizon T;

FIG. 10C is a graph of R(T) versus a time horizon T;

FIG. 11 is a block diagram of a network node constructed in accordance with principles set forth herein;

FIG. 12 is a flowchart of an exemplary process for determining a MIMO precoding matrix; and

FIG. 13 is a flowchart of an alternative exemplary process for determining a MIMO precoding matrix.

DETAILED DESCRIPTION

Before describing in detail exemplary embodiments, it is noted that the embodiments reside primarily in combinations of apparatus components and processing steps related to online convex optimization with periodic updates for downlink multi-cell multiple input multiple output (MIMO) wireless network virtualization. Accordingly, components have been represented where appropriate by conventional symbols in the drawings, showing only those specific details that are pertinent to understanding the embodiments so as not to obscure the disclosure with details that will be readily apparent to those of ordinary skill in the art having the benefit of the description herein.

As used herein, relational terms, such as “first” and “second,” “top” and “bottom,” and the like, may be used solely to distinguish one entity or element from another entity or element without necessarily requiring or implying any physical or logical relationship or order between such entities or elements.

The term “network node” or “radio network node” used herein can be any kind of network node comprised in a radio network which may further comprise any of base station (BS), radio base station, base transceiver station (BTS), base station controller (BSC), radio network controller (RNC), g Node B (gNB), evolved Node B (eNB or eNodeB), Node B, multi-standard radio (MSR) radio node such as MSR BS, multi-cell/multicast coordination entity (MCE), relay node, donor node controlling relay, radio access point (AP), transmission points, transmission nodes, Remote Radio Unit (RRU) Remote Radio Head (RRH), a core network node (e.g., mobile management entity (MME), self-organizing network (SON) node, a coordinating node, positioning node, MDT node, etc.), an external node (e.g., 3rd party node, a node external to the current network), nodes in distributed antenna system (DAS) etc. The network node may also comprise test equipment. The term “radio node” used herein may be used to also denote a wireless device such as a WD or a radio network node.

A method and network node for online coordinated multi-cell precoding are provided. According to one aspect, a method includes receiving from each of the plurality of service providers a virtual precoder matrix determined by the corresponding service provider. The method also includes determining a precoder matrix by minimizing a precoding deviation from a virtualization demand of the network subject to at least one power constraint, the virtualization demand being based at least in part on a product of a channel state matrix and a virtual precoder matrix. The method further includes applying the determined precoder matrix to signals applied to a plurality of antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

Problem Formulation

In this subsection, a constrained OCO problem is formulated. Then, for a performance metric, a regret with partial feedback and long-term constraint violations is defined. Finally, an OCO algorithm with periodic updates is derived.

Under standard OCO settings, at the beginning of each time slot t, an agent makes a decision x_t∈X₀, where X₀∈ⁿ is a known compact convex set. At the end of each time slot t, the system reveals the convex loss function f_t(x): ⁿ to the agent, who then suffers a loss f_t(x_t). A goal of an OCO algorithm is to select a sequence {x_t} such that the accumulated loss f_t(x_t) is comparable with the one yielded by the optimal offline fixed strategy defined as x^o arg min_x∈X0f_t(x), with all convex loss functions known in hindsight, where ={0, . . . ,T−1}. A desired OCO algorithm should provide sub-linear regret defined as f_t(x_t)−f_t(x^o). With the assumption of bounded gradient or sub-gradient, it is known that the online projected gradient ascent algorithm provides O(√{square root over (T)}) regret. Note that the compact convex set X₀ can be expressed as a set of short-term constraints, _c(x_t)≤0, where _c(x_t): ⁿ, c∈ with ={1, . . . , C}.

The standard OCO formulation is useful in many applications. However, it does not accommodate long-term constraints. OCO with long-term constraints was formulated that does not require _c(x_t)≤0 to be satisfied at each time slot t, but only requires the long-term constraint violation, defined as _c(x_t), to grow sub-linearly. Recently, a virtual-queue-based OCO algorithm in provides O(√{square root over (T)}) regret and O(1) long-term constraint violation under the per-time-slot OCO settings.

Thus an OCO problem formulation with periodic updates as shown in FIG. 1 is provided. The total time horizon T is segmented into I update periods, indexed by the set ={0, . . . , I−1}, each having a duration of T₀≥1 time slots. At the beginning of each update period i∈, an agent is required to make a decision x_i for the next T₀ time slots, from a known compact convex set X₀ that incorporates the short-term constraints. The variable t_i denotes the first time slot of update period i. Let f_t(x): ⁿ be a convex loss function at time slot t, which is determined by the system (e.g., sum rate of a wireless network determined by the wireless channel conditions). Within each update period i, the agent is sent feedback multiple times of the gradient or sub-gradient information ∇f_t(x_i) for t_i≤t<t_i+1. Let _i={1, . . . , S_i} be the index set of such feedback. Let τ_i^s denote the time slot when feedback s, s∈S_i is sent. Then, at the end of time slot τ_i^s+d_i^s−1, ∇f_τis(x_i) is revealed to the agent, where d_i^s is the delay of the feedback at time slot τ_i^s.

While the set X₀ accommodates short-term constraints, long-term constraints in may also be considered in the problem formulation. Let g_c(x): ⁿ, c∈ be the known convex long-term constraint function. The agent tries to enforce the long-term constraint T₀(x_i)0 over I periods, where (x)[₁(x), . . . , _c(x)]^T and denotes element-by-element inequality.

In some embodiments, the following two conventional goals of an OCO algorithm, with slight modification for the system model are achieved. One goal of an OCO algorithm is to select a sequence {x_i}∈X₀, such that the accumulated loss is competitive with an optimal offline fixed strategy defined with partial feedback

$x_p^{\circ }\stackrel{△}{=}\arg {\min }_{x\in 𝒳}{\sum }_{i\in 𝒥}\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}{f}_{{\tau }_i^s}\left(x\right),$ where X{x∈X₀: g(x)0}. To capture this, define the regret with partial feedback as follows:

$\begin{array}{cc}{\mathrm{RE}}_p\left(T\right)\stackrel{△}{=}\frac{T_0}{S_i}{\sum }_{i\in 𝒥}{\sum }_{s\in {𝒮}_i}{f}_{{\tau }_i^s}\left(x_i\right)-f_{{\tau }_i^s}\left(x_p^{\circ }\right).& \left(1\right)\end{array}$ Another goal of an OCO algorithm is to provide sub-linear long-term constraint violation defined for any c as follows:V0_c(T)T_0c(x_i)  (2)Note that the long-term constraints introduce correlation on the decisions over time, and lead to a more complicated online problem when the decision is fixed for one update period that lasts for multiple time slots, while the underlying system evolves over time.

Assume that each feedback ∇f_τis(x_i), s∈S_i is delayed for at least one time slot and received by the agent before it makes the next decision x_i+1, i.e., 1≤d_i^s≤t_i+1−τ_i^s∀s∈S_i, ∀i∈. Also assume that there is at least one feedback received by the agent within each update period, i.e., 1≤S_i, ∀i∈. Note that the feedback within each update period is not assumed to be received in order, i.e., even if r<s, the following statement is not necessarily true: τ_i^r+d_i^r≤τ_i^s+d_i^s, ∀r, ∈_i.

When the feedback received for the gradient or sub-gradient of the convex loss functions is incomplete, i.e., S_i<T, it is unfair to compare the algorithm described herein with the standard optimal offline fixed strategy x^o that knows all convex loss functions in hindsight. Let ={τ_i¹, . . . , τ_i^si} be the set of time slots with feedback and _i={t: t_i≤t<t_i+1} be the set of time slots where decision x_i is made. For the offline setting, assume the convex loss function at time slot t∈_i\ is the average of the convex loss functions received within update period i, i.e.,

$f_t\left(x\right)=\frac{1}{S_i}{\sum }_{s\in {𝒮}_i}{f}_{{\tau }_i^s}\left(x\right),\forall t\in {ℱ}_i\backslash {𝒯}_i,\forall i\in 𝒥.$ As such, the OCO algorithm derived herein is compared with an optimal offline fixed strategy x_p^o that has the same partial feedback information. Note that when all gradient or sub-gradient feedback for all time slots in each period are received, i.e., S_i=T₀, ∀i∈, x_p^o becomes the standard optimal offline fixed strategy x⁰.

OCO Algorithm with Periodic Updates

The OCO algorithm described herein with periodic updates is presented as Algorithm 1, summarized below. It selects a decision sequence {x_i} and a virtual queue sequence {Q_i}, which can be viewed as the primal and dual variables in the saddle-point-typed OCO algorithms. The main difference is that Algorithm 1 uses the virtual queues as the dual variables to incorporate the violations of long-term constraints. The virtual queue sequence {Q_i} serves as a queue backlog of the violations of long-term constraints, and an upper bound on the virtual queue vector can be readily translated to an upper bound on the long-term constraint violations. It has been shown that a virtual-queue-based algorithm has some advantages over the saddle-point-typed algorithms regarding performance guarantees under the standard OCO settings.

In some embodiments, algorithm 1 may be summarized as follows:

• • A. Let γ, α>0 be constant algorithm parameters. At the beginning of period i=0, initialize Q₀ [Q₀¹, . . . , Q₀^C]^T=0 and x₀∈X₀. By the end of period 0, receive ∇f_τ0s(x₀), s∈₀.
  • B. If i+1≥I, then End.
  • C. If i+1≤I, then:
  • D. At the beginning of each i+1-st update period t_i+1, use ∇f_τis(x_i), s∈_i to do the following:
    • Update the virtual queue for all c∈Q_i+1^c=max{−T_0c(x_i),Q_i^c+T_0c(x_i)}  (3)where _c(x_i)γ_c(x_i).
  • E. Solve P1 for x_i+1 as

$P1:{\min }_{x\in 𝒳}\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}\left\{▽f_{{\tau }_i^s}^T\left(x_i\right)\left(x-x_i\right)\right\}+\alpha {x-x_i}_2^2+$ ${\left[Q_{i+1}+T_0\tilde{g}\left(x_i\right)\right]}^T\left[T_0\tilde{g}\left(x\right)\right]\mathrm{where}{Q}_{i+1}\stackrel{△}{=}{\left[Q_{i+1}^1,\dots ,Q_{i+1}^C\right]}^T,\mathrm{and}\tilde{g}\left(x\right)\stackrel{△}{=}{\left[{\tilde{g}}_1\left(x\right),\dots ,{\tilde{g}}_C\left(x\right)\right]}^T$

• • F. By the end of period i+1, receive ∇f_τi+1s(x_i+1), s∈_i+1.
  • G. i=i+1.
  • H. Return to B.The following assumptions, which are common in the OCO literature, are made to facilitate the regret and long-term constraint violation bound analysis of Algorithm 1.Assumption 1. There exists a constant D>0 such that for any x∈X₀ and any t∈T, and∥∇f_t(x)∥₂≤D  (4)Assumption 2. There exists a constant β>0 such that g(x) is Lipschitz continuous with modulus β for any x, y∈X₀ as∥g(x)−g(y)∥₂≤β∥x−y∥₂  (5)Assumption 3. There exists a constant G>0 such that for any x∈X₀, and∥g(x)∥₂≤G  (6)Assumption 4. There exists a constant R>0 such that for any x, y∈X₀, and∥x−y∥₂≤R  (7)Assumption 5. There exists a constant ∈>0 and a {circumflex over (x)}∈X₀ such for any c∈, andg_c({circumflex over (x)})≤−∈  (8)Assumptions 1 and 2 are the standard assumptions of bounded gradient or sub-gradient for the convex loss functions and the Lipschitz continuity assumption for long-term constraints. Assumptions 3 and 4 bound the outputs of the long-term constraint functions and the decision set. Assumption 5 states that there is at least one interior point for any of the long-term constraint functions.

Note that the disclosed OCO formulation and Algorithm 1 account for the impact of multiple-time-slot update period on the regret bound and long-term constraint violation bound, which has not been studied in the literature before. Compared with the virtual-queue-based OCO algorithm under standard OCO settings, Algorithm 1 only uses the gradient or sub-gradient information instead of the loss function information. Furthermore, when the long-term and short-term constraint functions are separable with respect to blocks of x, P1 in Algorithm 1 can be equivalently decomposed into several sub-problems, each corresponding to a block of x_i+1 and can be solved with lower computational complexity.

Performance Analysis

In this subsection, performance bounds are provided for the OCO algorithm described herein. Impacts of the update period on the regret and long-term constraint violation bounds are explicitly analyzed.

First, bounds are provided on the virtual queues in the following lemma.

Lemma 1. The virtual queues are bounded for any i∈ and any c∈ as followsQ_i^c≥0,  (9)Q_i+1^c+T₀{tilde over (g)}_c(x_i)≥0,  (10)∥Q_i+1∥₂²≥∥T₀{tilde over (g)}_c(x_i)∥₂²,  (11)∥Q_i+1∥₂≤∥Q_i∥₂+∥T₀{tilde over (g)}_c(x_i)∥₂  (12)Define

$L_i\stackrel{△}{=}\frac{1}{2}{Q_i}_2^2$ as the quadratic Lyapunov function and Δ_iL_i+1−L_i as the corresponding Lyapunov drift. Leveraging results in Lemma 1, an upper bound on the Lyapunov drift is provided in the following lemma.Lemma 2. The Lyapunov drift Δ_i at each update period i∈ is upper bounded byΔ_i≤Q_i^T[T₀{tilde over (g)}(x_i)]+∥T₀{tilde over (g)}(x_i)∥₂²  (13)From the definition of strong convexity, function f(x) is said to be strongly convex on a convex set X∈ⁿ with modulus α, if there exists a constant α>0 such that

$f\left(x\right)-\frac{\alpha }{2}{x}_2^2$ is convex on X. Based on the properties of strongly convex functions, the following lemma is commonly used in the literature of OCO for regret analysis.Lemma 3. Let X∈ⁿ be a convex set. Let function f(x) be strongly convex on X with modulus α and x_opt be a global minimum of f(x) on X. Then, for any x∈X:

$f\left(x_{\mathrm{opt}}\right)\le f\left(x\right)-\frac{\alpha }{2}{x_{\mathrm{opt}}-x}_2^2.$ Leveraging results in Lemmas 1-3 and OCO techniques, the following theorem provides regret bounds for Algorithm 1.Theorem 4. If

$\gamma =T^{\frac{1}{4}},\mathrm{and}\alpha =\frac{1}{2}\left(T_0+T_0^2{\beta }^2\right)\sqrt{T}$ in Algorithm 1, then for all I>0:

$\begin{array}{cc}{\mathrm{RE}}_p\left(T\right)\le \frac{D^2+\left(T_0+T_0^2{\beta }^2\right)R^2+T_0^2{G}^2}{2}\sqrt{T}=O\left(\sqrt{T}\right).& \left(14\right)\end{array}$ Next, an upper bound on the long-term constraint violation in (2) is provided for Algorithm 1. The virtual queue is related to the long-term constraint violation in the following lemma.Lemma 5. For the virtual queue Q_i produced by Algorithm 1, for any c∈

$\begin{array}{cc}{\mathrm{VO}}_c\left(T\right)\le \frac{1}{\gamma }{Q}_i^c.& \left(15\right)\end{array}$ From Lemma 5, the long-term constraint violation VO_c(T) can be bound through an upper bound on the virtual queue Q_i^c. The following theorem provides an upper bound on the long-term constraint violation over any given I>0 update periods.

Theorem 6. If

$\gamma =T^{\frac{1}{4}},\mathrm{and}\alpha =\frac{1}{2}\left(T_0+T_0^2{\beta }^2\right)\sqrt{T}$ in Algorithm 1, then for all I>0 and any c∈:

$\begin{array}{cc}{\mathrm{VO}}_c\left(T\right)\le 2T_{0}G+\frac{2\mathrm{DR}+\left(1+T_0{\beta }^2\right)R^2+4T_0{G}^2}{2ϵ}=O\left(1\right).& \left(16\right)\end{array}$ Theorems 4 and 6 show that Algorithm 1 provides O(√{square root over (T)}) regret defined with partial feedback in (1) and O(1) long-term constraint violation in (2). Performance analysis explicitly considers the impacts of periodic updates on the performance guarantees of OCO, which has not been studied in the literature before. The above results can be extended to a more general system model with time-varying update periods and unknown time horizon. This is explained in the following remarks.

Define the variation of update period as V_T=(T_i−T_i+1)², where T_i is the i-th update period. Under the assumptions that T_i≤T_max, ∀i∈ and V_T≤O(1), i.e., both the update period and the variation of update period are upper bounded, it can be shown that Algorithm 1 still yields O(√{square root over (T)}) regret and O(1) long-term constraint violation.

The parameters γ and α in Algorithm 1 depend on the time horizon T. When T is unknown, the standard doubling trick can be applied to preserve the O(√{square root over (T)}) regret bound of Algorithm 1 when T is known. It can be easily shown that applying the doubling trick to Algorithm 1 yields O(log₂ T) long-term constraint violation.

Online Multi-Cell MIMO WNV

In this subsection, an online coordinated multi-cell precoding problem for downlink MIMO WNV with unknown CDI and delayed CSI is disclosed. Both long-term and short-term transmit power constraints are accommodated in the formulation. Leveraging the OCO algorithm described herein, an online downlink coordinated multi-cell MIMO WNV algorithm is disclosed, with a distributed closed-form precoding solution and guaranteed performance bounds.

Idealized MIMO WNV Model

In the system under consideration, an InP performs WNV to a MIMO cellular network that consists of C cells, each cell c∈ has one BS equipped with N_c antennas, for a total of N=N_c antennas in the network. The InP serves M SPs, each SP m∈ has K_c^m users, for a total of K_c=K_c^m users in cell c and a total of K=K_c users in the network. Let ={1, . . . , C}, ={1, . . . , M} and ={1, . . . , K}. Consider a time-slotted system with time indexed by t, and denote Ĥ_t^lcm∈^KlmNc as the local channel state between the K_l^m users of SP m in cell l and the BS in cell c at time slot t.

For ease of exposition, first consider an idealized WNV framework, with channel state feedback in each time slot without channel state feedback delay, as shown in FIG. 2, where the double arrowed lines extending from SP1 to SP2 indicating sharing between these two SPs. At each time slot t, in each cell c, the InP shares the corresponding channel state Ĥ_t^ccm∈^Kcm×Nc with SP m, and allocates transmit power P_c^m to the SP, such that P_c^m≤P_max^c with P_max^c being the maximum transmit power limit of cell c. Each SP m designs its own precoding matrix Ŵ_t^ccm∈^Nc×Kcm based on the service needs of its subscribing users, subject to the transmit power limit ∥Ŵ_t^ccm∥_F²≤P_c^m, and then sends this matrix to the InP as a virtual precoding matrix. Note that each SP m designs Ŵ_t^ccm without the need to be aware of its own users in the other cells and the other SPs. With the virtual precoding matrices designed by the SPs, the desired received signal vector (noiseless) for all K_c users in cell c is{tilde over (y)}_t^c′={tilde over (D)}_t^cx_t^c where {tilde over (D)}_t^cblkdiag{D_t^c1, . . . , D_t^cM} is the virtualization demand from cell c with D_t^cmĤ_t^ccmŴ_t^ccm being the virtualization demand from SP m in cell c, and x_t^c=[x_t^c1H, . . . , x_t^cMH]^H is the transmitted signal vectors for the K_c users in cell c, with x_t^cm being the transmitted signal vector for the K_c^m users of SP m in cell c. The desired received signal vector at all K users in the network isy′_t=D_tx_t where D_tblkdiag{{tilde over (D)}_t¹, . . . , {tilde over (D)}_t^c} is the virtualization demand from the network, x_t=[x_t^1H, . . . , x_t^CH]^H, and y′_t=[{tilde over (y)}_t^c′^H]^H. Without loss of generality, assume the transmitted signal to each user is zero-mean with unit power and uncorrelated to each other at each time slot t, i.e., {x_t}=0 and {x_tx_t^H}=I, ∀t∈.

The InP coordinates the cells at the precoding level to serve all the users directly. In each cell c, with local channel state {tilde over (H)}_t^c=[H_t^1cH, . . . , H_t^CcH]^H∈^K×Nc, the InP designs the actual downlink precoding matrix {tilde over (V)}_t^c∈^Nc×Kc to serve the K_c users in cell c, where H_t^lc=[Ĥ_t^lc1H, . . . , Ĥ_t^lcMH]^H∈^Kl×Nc is the channel state between the K_l users in cell l and the BS in cell c. The actual received signal vector at all K users (for now without considering noise) isy_t=H_tV_tx_t where H_t=[{tilde over (H)}_t¹, . . . , {tilde over (H)}_t^C]∈^K×N is the global channel state, and V_t=blkdiag{{tilde over (V)}_t¹, . . . , {tilde over (V)}_t^C}∈^N×K is the global precoding matrix designed by the InP. The expected deviation of the received signal vector at all K users via the InP's actual global precoding matrix from that via the SPs' local virtual precoding matrices is given by{∥y_t−y′_t∥₂²}={∥H_tV_t−D_t∥_F²}where ∥·∥_F represents the Frobenius norm.

Problem Formulation

Define the convex set for InP's global precoding matrix V∈^N×K as ₀{V=blkdiag{{tilde over (V)}¹, . . . , {tilde over (V)}^c}: ∥{tilde over (V)}^c∥_F²≤P_max^c, ∀c ∈}, where {tilde over (V)}^c∈^Nc×Kc is the local precoding matrix for cell c. Define the precoding deviation from virtualization demand of the network as

$\begin{array}{cc}{f}_t\left(V\right)\stackrel{△}{=}{H_{t}V-D_t}_F^2=\sum _{c\in 𝒞}\left\{{{\stackrel{\_}{H}}_t^{\mathrm{cc}}{\tilde{V}}^c-{\tilde{D}}_t^c}_F^2+\sum _{l\in 𝒞,l\ne c}{{\stackrel{\_}{H}}_t^{\mathrm{lc}}{\tilde{V}}^c}_F^2\right\}& \left(17\right)\end{array}$ where the first term is the precoding deviation from virtualization demand of cell c, and the second term is the leakage from cell c to all the other cells. Define the long-term transmit power constraint function for cell c∈ asg_c({tilde over (V)}^c)∥{tilde over (V)}^c∥_F²−P_c  (18)where P_c≤P_max^c is the long-term transmit power limit set by the InP. The InP coordinates the cells at the precoding level to minimize the accumulated precoding deviation from the virtualization demand, subject to both long-term and short-term transmit power constraints.

Consider a periodic demand-response mechanism between the InP and the SPs. At the beginning of each time slot t_i=iT₀, i∈, the InP coordinates the C cells at the precoding level and designs a fixed global precoding matrix V_i∈^N×K for the next T₀ time slots. The InP then schedules channel state feedback to itself (with any channel estimation and feedback transmission method) at time slot τ_i^s, s∈_i, and receives the global channel state H_τis at the end of time slot τ_i^s+d_i^s−1, where d_i^s is the channel state feedback delay that satisfies 1≤d_i^s≤t_i+1−τ_i^s. The InP is allowed to schedule partial channel state feedbacks within each demand-response period, i.e., 1≤S_i≤T₀, ∀i∈. In each cell c, the InP shares Ĥ_τis^ccm with SP m, who then designs its own precoding matrix Ŵ_τis^ccm on the fly and sends it to the InP before it designs the next global precoding matrix V_i+1.

The InP designs {V_i}∈₀ to provide sub-linear T-slot regret defined with partial feedback on the precoding deviation from the virtualization demand

$\begin{array}{cc}{\mathrm{RE}}_p\left(T\right)\stackrel{△}{=}{\sum }_{i\in 𝒥}\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}{f}_{{\tau }_i^s}\left(V_i\right)-f_{{\tau }_i^s}\left(V_p^{\circ }\right)\mathrm{where}{V}_p^{\circ }\stackrel{△}{=}\arg {\min }_{V\in 𝒱}{\sum }_{i\in 𝒥}\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}{f}_{{\tau }_i^s}\left(V\right)& \left(19\right)\end{array}$ is the optimal offline fixed precoding strategy with partial feedbacks, and {V=diag{{tilde over (V)}¹, . . . , {tilde over (V)}^c}: ∥{tilde over (V)}^c∥_F²≤P_c, ∀c∈}. Note that when all the channel state feedbacks are received by the InP, (19) becomes the standard regret definition in the literature of OCO. Another goal of the InP is to provide sub-linear long-term constraint violation on the long-term transmit power defined for any c∈ asVO_c(T)T₀g_c({tilde over (V)}_i^c)  (20)

Online Downlink Coordinated Multi-Cell MIMO WNV Algorithm

The OCO algorithm described herein provides a solution to the afore-mentioned online coordinated multi-cell precoding problem for downlink MIMO WNV. From the first-order condition of a real-valued scalar convex function with respect to the complex-valued matrix variable, for any V, V_i∈₀ f_t(V)≥f_t(V_i)+2Re{tr{∇_Vi*f_t^H(V_i)(V−V_i)}}  (21)where ∇_Vi*f_t(V_i)=H_t^H(H_tV_i−D_t) is the partial derivative of f_t(V_i) with respect to V_i^* (the complex conjugate of V_i).

Leveraging the OCO algorithm described herein, at the beginning of each demand-response period t_i+1, the InP solves P2 for the global precoding matrix V_i+1 as follows:

$P2:{\min }_{V\in {𝒱}_0}\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}2Re\left\{\mathrm{tr}\left\{{\nabla }_{V_i^{*}}{f}_{{\tau }_i^s}^H\left(V_i\right)\left(V-V_i\right)\right\}\right\}+\alpha {V-V_i}_F^2+{\left[Q_{i+1}+T_0\widetilde{ℊ}\left(V_i\right)\right]}^T\left[T_0\widetilde{ℊ}\left(V\right)\right].$ Since both the long-term and short-term transmit power constraints are on each cell, it then follows that P2 can be equivalently decomposed into C subproblems, each corresponding to the local precoding matrix {tilde over (V)}_i+1^c for cell c∈ as follows:

$\begin{array}{cc}P3:{\min }_{{\tilde{V}}^c}\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}2\mathrm{Re}\left\{\mathrm{tr}\left\{{\nabla }_{{\tilde{V}}_i^{c*}}{f}_{{\tau }_i^s}^H\left(V_i\right)\left({\tilde{V}}^c-{\tilde{V}}_i^c\right)\right\}\right\}+\alpha {{\tilde{V}}^c-{\tilde{V}}_i^c}_F^2+\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]\left[T_0\widetilde{{ℊ}_c}\left({\tilde{V}}^c\right)\right]s.t.{{\tilde{V}}^c}_F^2-P_{\max }^c\le 0\mathrm{where}{\nabla }_{{\tilde{V}}_i^{c*}}{f}_{{\tau }_i^s}\left(V_i\right)={\tilde{H}}_{{\tau }_i^s}^{\mathrm{cH}}\left({\tilde{H}}_{{\tau }_i^s}^c{\tilde{V}}_i^c-{\tilde{D}}_{{\tau }_i^s}^{c^{\prime }}\right),\mathrm{and}{\tilde{D}}_{{\tau }_i^s}^{c^{\prime }}={\left[0,\dots ,D_{{\tau }_i^s}^{\mathrm{cH}},\dots ,0\right]}^H.& \left(22\right)\end{array}$

P3 is a convex optimization problem satisfying the Slater's condition, thus strong duality holds. P3 can be solved by studying the Karush-Kuhn-Tucker (KKT) conditions. The Lagrange function for P3 is

$L\left({\tilde{V}}^c,{\lambda }^c\right)=\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}2\mathrm{Re}\left\{\mathrm{tr}\left\{{\nabla }_{{\tilde{V}}_i^{c*}}{f}_{{\tau }_i^s}^H\left(V_i\right)\left({\tilde{V}}^c-{\tilde{V}}_i^c\right)\right\}\right\}+\alpha {{\tilde{V}}^c-{\tilde{V}}_i^c}_F^2+\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]\left[T_0\widetilde{{ℊ}_c}\left({\tilde{V}}^c\right)\right]+{\lambda }^c\left({{\tilde{V}}^c}_F^2-P_{\max }^c\right)$ where λ^c is the Lagrangian multiplier associated with constraint (22). Taking partial derivative of L({tilde over (V)}^c, λ^c) with respect to {tilde over (V)}^c*:

${\nabla }_{{\tilde{V}}^{c*}}L\left({\tilde{V}}^c,{\lambda }^c\right)=\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}{\nabla }_{{\tilde{V}}_i^{c*}}{f}_{{\tau }_i^s}\left(V_i\right)+\alpha \left({\tilde{V}}^c-{\tilde{V}}_i^c\right)+\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]T_0\gamma {\tilde{V}}^c+{\lambda }^c{\tilde{V}}^c.$ The KKT conditions for ({tilde over (V)}^co, λ^co) being globally optimal are given by

$\begin{array}{cc}{\tilde{V}}^{c\circ }=\frac{\alpha {\tilde{V}}_i^c-\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}{\nabla }_{{\tilde{V}}_i^{c*}}{f}_{{\tau }_i^s}\left(V_i\right)}{\alpha +\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]T_0\gamma +{\lambda }^{c\circ }},& \left(23\right)\end{array}$ $\begin{array}{cc}{{\tilde{V}}^{c\circ }}_F^2-P_{\max }^c\le 0,& \left(24\right)\end{array}$ $\begin{array}{cc}{\lambda }^{c\circ }\ge 0,& \left(25\right)\end{array}$ $\begin{array}{cc}{\lambda }^{c\circ }\left({{\tilde{V}}^{c\circ }}_F^2-P_{\max }^c\right)=0& \left(26\right)\end{array}$ where (23) follows by setting ∇_{{tilde over (V)}c*}L({tilde over (V)}^c, λ^c)=0. Consider the following cases.

• • 1) λ^co>0: From (23), there is a closed-form solution for {tilde over (V)}^co, by noting that λ^co>0 serves as a power regularization factor such that ∥{tilde over (V)}^co∥_F²=P_max^c.
  • 2) λ^co=0: From (23), a closed-form solution for {tilde over (V)}^co:

$\begin{array}{cc}{\tilde{V}}^{c\circ }=\frac{\alpha {\tilde{V}}_i^c-\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}{\nabla }_{{\tilde{V}}_i^{c*}}{f}_{{\tau }_i^s}\left(V_i\right)}{\alpha +\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]T_0\gamma },& \left(27\right)\end{array}$ Equation (27) can be categorized into two subcases: 2.1) If {tilde over (V)}^co in (27) satisfies (24), then it is the optimal solution. 2.2) If {tilde over (V)}^co in (27) cannot satisfy (24), which means the condition in Case 2) does not hold in optimality, i.e., λ^co>0, and the optimal solution is given by Case 1).

An online downlink coordinated multi-cell MIMO WNV algorithm in accordance with the present disclosure and referred to herein as algorithm 2, is summarized as follows:

• • A. For large T, set

$\gamma =T^{\frac{1}{4}}\mathrm{and}\alpha =\frac{1}{2}\left(T_0+T_0^2{\beta }^2\right)\sqrt{T};$ For intermediate value of T, solve P4 for γ and α.

• • B. At the beginning of period i=0, initialize Q₀=0 and V₀∈₀. By the end of period 0, receive Hτ_o^s and Dτ_o^s, s∈₀.
  • C. If, i+1≥I then end;
  • D. If, i+1≤I then:
  • E. At the beginning of each demand-response period i+1, use Hτ_i^s and D_τis, s∈_i to do the following:
    • Update virtual queue for all c∈Q_i+1^c=max{−T₀{tilde over (g)}_c({tilde over (V)}_i^c),Q_i^c+T₀{tilde over (g)}_c({tilde over (V)}_i^c)}  (28)
  • where {tilde over (g)}_c({tilde over (V)}_i^c)γg_c({tilde over (V)}_i^c).
    • Update {tilde over (V)}_i+1^c, ∀c∈ as

$\blacksquare {\tilde{V}}_{i+1}^c=\{\begin{array}{c}{\tilde{x}}_i^c,\mathrm{if}{{\tilde{x}}_i^c}_F^2\le P_{\max }^c\\ \frac{\sqrt{P_{\max }^c}}{{{\tilde{x}}_i^c}_F}{\tilde{x}}_i^c,\mathrm{else}\end{array}$ $\blacksquare \mathrm{where}{\tilde{x}}_i^c=\frac{\alpha {\tilde{V}}_i^c-\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}{\tilde{H}}_{{\tau }_i^s}^{\mathrm{cH}}\left({\tilde{H}}_{{\tau }_i^s}^c{\tilde{V}}_i^c-{\tilde{D}}_{{\tau }_i^s}^{c^{\prime }}\right)}{\alpha +\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]T_0\gamma }$

• • F. By the end of period i+1, receive H_τi+1s and D_τi+1s, s∈_i+1.
  • G. i=i+1
  • H. Return to C,

Performance Analysis

The following assumption is made that the channel gain is upper bounded for theoretical performance bound analysis.

Assumption 6. There exists a constant B>0 such that the channel gain is upper bounded for any t∈ as∥H_t∥_F≤B.  (29)The following lemma shows that Assumptions 1-5 hold for the formulated online downlink MIMO WNV problem.Lemma 7. Under Assumption 6, the formulated online downlink MIMO WNV problem with decision set ₀, loss function f_t(V), and long-term constraint function g_c({tilde over (V)}^j), c∈, satisfies Assumptions 1-5, with constants D, β, G, R, ∈ given as follows:

$\begin{array}{cc}D=4B^2\sqrt{{\sum }_{c\in 𝒞}{P}_{\max }^c},& \left(30\right)\end{array}$ $\begin{array}{cc}\beta =2\sqrt{\max \left\{P_{\max }^c,\forall c\in 𝒞\right\}},& \left(31\right)\end{array}$ $\begin{array}{cc}G=\sqrt{{\sum }_{c\in C}\max \left\{{\stackrel{\_}{P}}_c^2,{\left(P_{\max }^c-{\stackrel{\_}{P}}_c\right)}^2\right\}},& \left(32\right)\end{array}$ $\begin{array}{cc}R=2\sqrt{{\sum }_{c\in 𝒞}{P}_{\max }^c},& \left(33\right)\end{array}$ $\begin{array}{cc}ϵ=\min \left\{{\stackrel{\_}{P}}_c,\forall c\in 𝒞\right\}.& \left(34\right)\end{array}$ Leveraging the results in Lemma 7, Theorem 4, and Theorem 6, the following Theorem shows that Algorithm 2 provides O(√{square root over (T)}) regret on the precoding deviation from the virtualization demand and O(1) long-term transmit power constraint violation.Theorem 8. If

$\gamma =T^{\frac{1}{4}}\mathrm{and}\alpha =\frac{1}{2}\left(T_0+T_0^2{\beta }^2\right)\sqrt{T},$ Algorithm 2 yields an O(√{square root over (T)}) upper bound given in (15) on the regret defined in (19) and an O(1) upper bound given in (16) on the long-term constraint violation defined in (20), with constants D, β, G, R, ∈ given in Lemma 7.Furthermore, three remarks regarding the algorithm parameter selection, the optimal offline fixed precoding strategy, and the short-term per-antenna transmit power constraints are provided below.

Algorithm 2 provides O(√{square root over (T)}) regret and O(1) long-term constraint violation from Theorem 8 for a large T. However, the constants hidden behind the O(√{square root over (T)}) and O(1) upper bounds can be relatively large for an intermediate value of T. The following problem can be solved to obtain the algorithm parameters γ and α that yield the best regret bound, subject to a constraint on the long-term constraint violation upper bound:

$P4:{\min }_{\gamma >0,\alpha >0}\frac{D^2}{2}\sqrt{T}+\alpha R^2+\frac{1}{2}{T}_0^2{\gamma }^2{G}^2$ $s.t.2T_{0}G+\frac{T_0\mathrm{DR}+\alpha R^2+2T_0^2{\gamma }^2{G}^2}{T_0{\gamma }^2ϵ}$ $\le \min \left\{P_{\max }^c-{\stackrel{\_}{P}}_c,\forall c\in 𝒞\right\}\sqrt{T},$ $\frac{1}{2}\left(T_0\sqrt{T}+T_0^2{\beta }^2{\gamma }^2\right)\le \alpha$ where D, β, G, R, ∈ are given in Lemma 7. P4 is a geometric programming (GP) problem that can be easily solved by a standard convex program solver such as CVX.

Algorithm 2 provides an O(√{square root over (T)}) regret compared with the optimal offline fixed precoding solution V_p⁰ defined with the same partial feedbacks for fair comparison. However, for algorithm performance evaluation in Subsection 6.4, the performance of Algorithm 2 is compared with the one yielded by the standard optimal offline fixed strategy V_opt⁰, that solves the following optimization problem with complete feedbackP5: min_vopt∥H_tV_opt−D_t∥_F² s.t.∥{tilde over (V)}_opt^c∥_F²≤P_c,∀c∈  (35)where V_opt=blkdiag{{tilde over (V)}_opt¹, . . . , {tilde over (V)}_opt^c}. Note that V_opt^o is a more knowledgeable fixed off-line strategy than V_p^o. Since P5 is a convex optimization problem satisfying the Slater's condition, the strong duality holds. P5 can be solved by studying the KKT conditions. The Lagrange function for P5 isL(V_opt,λ)=∥H_tV_opt−D_t∥_F²+λ^c(∥{tilde over (V)}_opt^c∥_F²−P_c)where λ^c, c∈ is the Lagrangian multiplier associated with constraint (35). Taking partial derivative of L(V_opt, λ) with respect to {tilde over (V)}_opt^c*:∇_{{tilde over (V)}optc*}L(V_opt,λ)={tilde over (H)}_t^cH({tilde over (H)}_t^c{tilde over (V)}_opt^c−{tilde over (D)}_t^c′)+λ^c{tilde over (V)}_opt^c where {tilde over (D)}_t^c′=[0, . . . , {tilde over (D)}_t^cH, . . . , 0]^H. The KKT conditions for ({tilde over (V)}_opt^coλ^co), c∈ being globally optimal are given by(A_c+λ^coI){tilde over (V)}_opt^co=B_c  (36)∥{tilde over (V)}_opt^co∥_F²−P_c≤0  (37)λ^co≥0  (38)λ^co(∥{tilde over (V)}_opt^co∥_F²−P_c)=0  (39)where (36) follows by setting ∇_{{tilde over (V)}optco}L(V_opt, λ)=0, A_c=Σ_t∈τ{tilde over (H)}_t^cH{tilde over (H)}_t^c and B_c={tilde over (H)}_t^cH{tilde over (D)}_t^c′. Consider the following cases.

• • 1) λ^co>0: From (36), A_c+λ^coI>0, which implies that it is invertible, and:{tilde over (V)}_opt^co=(A_c+λ^coI)⁻¹B_c  (40)in which λ^co>0 can be found by bisection search such that ∥(A_c+λ^coI)⁻¹B_c∥_F²=P_c.
  • 2) λ^co=0: From (36), the optimal solution must satisfyA_c{tilde over (V)}_opt ^co=B_c  (41)categorize (41) into two subcases: 2.1) If A_c is a rank deficient matrix, there are infinitely many solutions for {tilde over (V)}_opt^co. {tilde over (V)}_opt^co is chosen to minimize ∥{tilde over (V)}_opt^co∥_F² subject to (41), which is an under-determined least-square problem with a closed-form solution:{tilde over (V)}_opt^co=A_c^H(A_cA_c^H)⁻¹B_c  (42)By (37), if ∥A_c^H(A_cA_c^H)⁻¹B_c∥_F²≤P_c, then {tilde over (V)}_opt^co in (42) is the optimal solution. 2.2) If A_c is of full rank, there is a unique solution:{tilde over (V)}_opt^co=A_c⁻¹B_c  (43)Again, if ∥A_c⁻¹B_c∥_F²≤P_c, then {tilde over (V)}_opt^co in (43) is the optimal solution. For both subcases 2.1) and 2.2), {tilde over (V)}_opt^co in (42) or (43) cannot satisfy (37), which means the condition in Case 2) does not hold in optimality, i.e., λ^co>0, and the optimal solution is given by Case 1).

Short-term per-antenna transmit power constraints can be incorporated into the problem formulation, by redefining the convex decision set as ₀ {V=blkdiag{{tilde over (V)}¹, . . . , {tilde over (V)}^c}: ∥v^cn∥_F²≤P_max^cn, ∀n∈}, where {tilde over (V)}^c=[v^c1, . . . , v^cNc]^H∈^Nc×Kc is the precoding matrix for cell c, v^cn∈ ^Kc×1 is the precoding vector and P_max^cn is the maximum transmit power limit for the n-th antenna in cell c. Leveraging the OCO algorithm described herein, the InP solves P6 to obtain {tilde over (V)}_i+1^c as follows:

$\begin{array}{cc}P6:{\min }_{{\tilde{V}}^c}\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}2\mathrm{Re}\left\{\mathrm{tr}\left\{{\nabla }_{{\tilde{V}}_i^{c*}}{f}_{{\tau }_i^s}^H\left(V_i\right)\left({\tilde{V}}^c-{\tilde{V}}_i^c\right)\right\}\right\}+\alpha {{\tilde{V}}^c-{\tilde{V}}_i^c}_F^2& \left(44\right)\end{array}$ $+\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]\left[T_0{\widetilde{ℊ}}_c\left({\tilde{V}}^c\right)\right]$ $s.t.{v^{\mathrm{cn}}}_F^2-P_{\max }^{\mathrm{cn}}\le 0,\forall n\in {𝒩}_c.$ P6 can be equivalently decomposed into N_c sub-problems, each corresponding to the precoding vector v^cn for the n-th antenna in cell c as

$\begin{array}{cc}P7:{\min }_{v^{\mathrm{cn}}}\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}2\mathrm{Re}\left\{\mathrm{tr}\left\{{\nabla }_{v_i^{\mathrm{cn}*}}{f}_{{\tau }_i^s}^H\left(V_i\right)\left(v^{\mathrm{cn}}-v_i^{\mathrm{cn}}\right)\right\}\right\}+\alpha {v^{\mathrm{cn}}-v_i^{\mathrm{cn}}}_F^2& \left(45\right)\end{array}$ $+\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]\left[T_0\gamma {v^{\mathrm{cn}}}_F^2\right]$ $s.t.{v^{\mathrm{cn}}}_F^2-P_{\max }^{\mathrm{cn}}\le 0$ where ∇_vicn*f_τis(V_i)=h_τis^cnH({tilde over (H)}_τis^c{tilde over (V)}_i^c−{tilde over (D)}_τis^c′) and h_τis^cn is the n-th column vector of {tilde over (H)}_τis^c. By studying the KKT conditions, a closed-form solution for v_i+1^cnº is given by

$\begin{array}{cc}{v}_{i+1}^{\mathrm{cn}\circ }=\{\begin{array}{cc}{x}_i^{\mathrm{cn}},& \mathrm{if}{x_i^{\mathrm{cn}}}_F^2\le P_{\max }^{\mathrm{cn}}\\ \frac{\sqrt{P_{\max }^{\mathrm{cn}}}}{{x_i^{\mathrm{cn}}}_F}{x}_i^{\mathrm{cn}},& \mathrm{else}\end{array}& \left(46\right)\end{array}$ $\mathrm{where}$ $x_i^{\mathrm{cn}}=\frac{\alpha v_i^{\mathrm{cn}}-\frac{T_0}{S_i}{\sum }_{s\in {𝒮}_i}{h}_{{\tau }_i^s}^{\mathrm{cnH}}\left({\tilde{H}}_{{\tau }_i^s}^c{\tilde{V}}_i^c-{\tilde{D}}_{{\tau }_i^s}^{c^{\prime }}\right)}{\alpha +\left[Q_{i+1}^c+T_0{\widetilde{ℊ}}_c\left({\tilde{V}}_i^c\right)\right]T_0\gamma }.$

Numerical Performance Evaluation

Consider an InP that owns a cellular network consisting of C=7 urban hexagon micro-cells, each cell c is of 500 m radius and has a BS at the center equipped with N_c=32 antennas by default. The InP serves M=4 SPs, each SP m serves K_c^m=2 users, for a total of K_c=8 users in each cell c. Set maximum transmit power limit P_max^c=33 dBm, long-term transmit power limit P_c=32 dBm, noise spectral density N₀=−174 dBm/Hz, noise figure N_F=10 dB, and transmission channel bandwidth B_W=15 kHz.

The fading channel is modeled as a first-order Gauss-Markov processh_t+1^ck=α_hh_t^ck+z_t^ck where h_t^ck˜(0, β_ckI) is the channel state between user k and BS antennas in cell c, β_ck[dB]=−31.54−33 log₁₀d_ck−ψ_ck captures path-loss and shadowing, d_ck is the distance in kilometers from the BS in cell c to user k, ψ_ck˜(0, σ_ز) accounts for shadowing with a σ_ø=8 dB, α_h is the channel correlation coefficient, and z_t^ck˜(0, (1−α_h²)β_ckI) is independent of h_t^ck. For a coherence time of

$T_c=\frac{9}{16\pi f_d}$ and transmission channel bandwidth B_w, α_h is determined by

${\alpha }_h^{T_c{B}_W}=\varnothing \mathrm{where}{f}_d=\frac{v}{v_c}{f}_c$ is the Doppler spread, v is the user speed assumed to be the same for all K users, f_c=2.5 GHz is the central frequency, v_c is the speed of light, and Ø=0.1 is the level of de-correlation. For α_h=0.995 as default in the simulation, v≈3 km/h. FIG. 3 shows a timeline illustrating online precoding design for downlink MIMO WNV. Under the standard LTE transmission frame structure, one symbol is transmitted with symbol duration

$\Delta t=\frac{1}{B_W}=66.67\mathrm{us}.$ Define Δt as one time slot and set the demand-response period T₀=8Δt (533.6 us) as default in the simulation, such that T₀ can be seen as one slot (500 us) that consists of 6 or 7 symbol durations plus cyclic prefix under the standard LTE transmission frame settings. An overview of the default simulation parameters is shown in Table 1.

	TABLE 1
	Parameter	Value
	Number of cells C	7
	Number of SPs M	4
	Number of users per cell Kc	8
	Number of users per cell Nc	32
	Transmission channel bandwidth BW	15	kHz
	Central frequency fc	2.5	GHz
	Noise spectral density N0	−174	dBm/Hz
	Noise figure NF	10	dB
	User speed v	3	km/h
	Maximum transmit power per cell Pmaxc	33	dBm
	Long-term transmit power limit per cell Pc	32	dBm
	Demand-response period T0	8
	Number of channel feedbacks per period Si	1
	Time horizon T	600

For performance study, consider the cases where each SP m uses MRT or ZF precoding, which are two commonly used precoding schemes in current MIMO systems, to design their own virtual precoding matrices given by

${\hat{W}}_t^{\mathrm{ccm}}=\{\begin{array}{c}{\alpha }_t^{\mathrm{ccm}}{\hat{H}}_t^{\mathrm{ccmH}},\mathrm{MRT}\\ {\alpha }_t^{\mathrm{ccm}}{{\hat{H}}_t^{\mathrm{ccmH}}\left({\hat{H}}_t^{\mathrm{ccm}}{\hat{H}}_t^{\mathrm{ccmH}}\right)}^{-1},\mathrm{ZF}\end{array}$ where a_t^ccm is a power normalizing factor such that

${{\hat{W}}_t^{\mathrm{ccm}}}_F^2=\frac{P_{\max }^c}{M},$ i.e., all SPs are allocated equal proportions of the maximum transmit power in each cell. Based on {V_i} produced by Algorithm 2, define the normalized time-averaged precoding deviation from the virtualization demand as

$\stackrel{\_}{\rho }\left(T\right)\stackrel{△}{=}\frac{1}{T}{\sum }_{i\in 𝒥}{\sum }_{t\in {ℱ}_i}\frac{f_t\left(V_i\right)}{{D_t}_F^2},$ the time-averaged transmit power as

$\stackrel{\_}{P}\left(T\right)\stackrel{△}{=}\frac{T_0}{\mathrm{TC}}{\sum }_{i\in 𝒥}{V_i}_F^2,$ and the time-averaged rate per user as

$\stackrel{\_}{R}\left(T\right)\stackrel{△}{=}\frac{1}{\mathrm{TK}}{\sum }_{i\in 𝒥}{\sum }_{t\in {ℱ}_i}{\sum }_{k\in 𝒦}{\log }_2\left(1+\mathrm{SIN}{R}_t^{\mathrm{ik}}\right)\mathrm{where}\mathrm{SIN}{R}_t^{\mathrm{ik}}=\frac{{❘h_t^{\mathrm{kT}}{v}_i^k❘}^2}{{\sigma }_n^2+{\sum }_{k^{\prime }\in 𝒦,k^{\prime }\ne k}{❘h_t^{\mathrm{kT}}{v}_i^{k^{\prime }}❘}^2},$ h_t^k and v_i^k are the channel state at time slot t and the precoding vector for the i-th demand-response period between user k and the N antennas, and σ_n²=N₀B_W+N_F is the noise power. For the upper bound B of the channel gain in Assumption 6, B is set as B=1.645√{square root over (N_cβ_ck)}, which gives a Chernoff upper bound of 9.07×10⁻¹⁴ for the probability of bound violation {∥H_t∥_F>B}.

Performance vs. Demand-Response Period T_o

First evaluate the performance of the disclosed algorithm described herein with different values of the demand-response period T₀. Consider the case where there is only one channel state feedback at the beginning of each demand-response period, i.e., ={iT₀}, ∀i∈, which is the same as the standard LTE transmission frame structure. FIGS. 4A, 4B, 4C, 5A, 5B and 5C show ρ(T), P(T), and R(T) versus time horizon T for different values of T₀, when all SPs adopt MRT and ZF precoding, respectively. Observe fast convergence of the disclosed algorithm (within 600 time slots). Then use the averaged value over the last 100 time slots as the steady-state value in the remaining simulation results. It can be seen that the system performance deteriorates as T₀ increases, which is because the underlying channel state keeps varying over time slots while V_i is fixed for each update period. As can be seen from FIGS. 5A, 5B and 5C, as T₀ increases, P(T) converges slower since V_i is fixed for a longer period and thus it is harder for the virtual queue vector Q_i to converge. When all SPs adopt ZF precoding, the steady-state value of ρ(T) is more sensitive to T₀ compared with the MRT precoding case, as ZF precoding is known to be sensitive to channel noise and inaccuracy.

Performance Comparison

A performance comparison between the disclosed online downlink coordinated MIMO WNV algorithm is provided herein with a different, but generally related arrangement that discloses standard per-time-slot OCO settings and does not provide performance guarantees under the periodic update scenario. However, the information of the update period T₀ may be ignored. The steady-state precoding deviation ρ and rate per user R versus the number of antennas N_c, between the two algorithms as shown in FIGS. 6A and 6B, are compared. It can be seen that the performance advantage of the disclosed algorithm over the known arrangement becomes more substantial as N_c, increases, indicating the advantage of applying the disclosed algorithm to massive MIMO.

Performance vs. Optimal Offline Fixed Precoding Strategy

FIGS. 7A and 7B show the performance of the disclosed algorithm compared with the one achieved with the offline fixed precoding strategy V_opt^o that solves P5 and the one yielded by Yu et al., with different values of the channel correlation coefficient α_h and all SPs adopting MRT precoding. It can be seen that the disclosed algorithm achieves better system performance compared with the other two schemes, which shows that the disclosed algorithm is able to better suppress the interference for coordinated multi-cell MIMO precoding.

Performance vs. Long-term Transmit Power Limit P_c

The impact of the long-term transmit power limit P_c on the disclosed algorithm has been studied. See FIGS. 8A, and 8B which show that there is a natural trade-off between the steady-state rate per user R and the long-term transmit power limit P_c, which allows the InP to balance the system performance and the energy consumption. The disclosed algorithm is applicable to massive MIMO, since the steady-state precoding deviations ρ vary within 10% as the number of antennas N_c, increasing from 32 to 256. The time-averaged rate per user R is 2.06 bpcu when N_c=32 compared with 4.48 bpcu when N_c=256, with the same long-term transmit power limit P_c=32 dBm, indicating substantial performance gain of WNV in massive MIMO.

Performance vs. Number of Cells C

The impact of the number of cells C on the performance of the algorithm provided herein is described with reference to FIGS. 9A and 9B. As C increasing, the inter-cell interference becomes more severe, and the steady-state precoding deviation ρ increases. However, the steady-state rate per user R is robust to the number of cells C, indicating the effectiveness of the coordinated multi-cell precoding design in handling the inter-cell interference. The impact of the number of cells C becomes smaller as the number of antennas N_c increases, indicating the advantages of coordinated multi-cell WNV in massive MIMO.

Per-Antenna Transmit Power Constraint

The impact of short-term per-antenna transmit power constraints on the system performance of the disclosed algorithm has been studied. As shown in FIGS. 10A, 10B and 10C, the number of antennas is set to N_c=256, the per-antenna maximum transmit power

$P_{\max }^{\mathrm{cn}}=\frac{P_{\max }^c}{N_c},$ and long-term transmit power limit P_c=32.5 dBm, such that the per-antenna constraint is much stricter than the per-cell transmit power constraint. It can be seen that the system performance deteriorates after adding the per-antenna transmit power constraint, since it further limits the convex decision set ₀ for the sequence of global precoding matrices {V_i}. The per-antenna constraint is observed to have negligible impact on the algorithm performance.

Thus, an OCO algorithm with periodic updates is provided herein. In existing standard OCO settings, decision making and convex loss function feedback are in a strict per-time-slot fashion. In contrast, the OCO algorithm described herein makes a decision at the beginning of each update period that can last for multiple time slots. The convex loss function can change arbitrarily at each time slot with unknown statistics. The gradient or sub-gradient feedbacks are allowed to be delayed for multiple time slots, received out of order, and missing for some time slots within each update period. Based only on the past gradient or sub-gradient information, the OCO algorithm described herein provides O(√{square root over (T)}) regret defined with partial feedbacks and O(1) long-term constraint violation.

An online downlink coordinated multi-cell MIMO WNV algorithm is disclosed, with unknown CDI and delayed CSI under standard LTE transmission frame settings. In the WNV framework, each SP is allowed to utilize all antennas and wireless spectrum resources simultaneously and design its own virtual precoding matrix in a cell without the need to be aware of the other SPs, or to consider the inter-cell interference created by its own users. The InP may coordinate the cells at the precoding level to minimize the current and future precoding deviations from the virtualization demands, based only on the past channel states and virtual precoding matrices. The sequence of global precoding matrices designed by the InP may subject to both long-term and short-term transmit power constraints. In such a case, the online coordinated multi-cell MIMO precoding problem has a distributed closed-form solution with low computational complexity. The disclosed algorithm provides theoretically guaranteed performance, as it yields O(√{square root over (T)}) regret on the precoding deviation from the virtualization demand, and O(1) long-term transmit power constraint violation.

The performance of the disclosed online downlink coordinated multi-cell MIMO WNV algorithm under typical LTE cellular network settings has been numerically validated. Simulation studies show fast convergence of the time-averaged performance. The disclosed algorithm shows substantial system performance advantage over an optimal offline fixed precoding strategy, and over the existing standard OCO algorithm with long-term constraints. Extensive simulation results are provided to demonstrate the impact of different values of demand-response period, precoding strategies adopted by the SPs, long-term transmit power limits, numbers of cells and antennas, and per-antenna transmit power constraint.

Referring to FIG. 11. one or more blocks described herein may be performed by one or more elements of network node 2 such as by one or more of processing circuitry 10 (including the OCO algorithm 12), processor 8, and radio interface 4. Network node 2 such as via processing circuitry 10 and/or processor 8 and radio interface 4, is configured to implement the OCO algorithm described herein. In some embodiments, the network node 2 is configurable by an infrastructure provider. A memory 6 stores data and computer code to cause the processor 8, which is part of processing circuitry 10, to execute an OCO algorithm as described above.

The radio interface 4 may be formed as or may include, for example, one or more RF transmitters, one or more RF receivers, and/or one or more RF transceivers. In particular, in addition to or instead of a processor, such as a central processing unit, and memory, the processing circuitry 10 may comprise integrated circuitry for processing and/or control, e.g., one or more processors and/or processor cores and/or FPGAs (Field Programmable Gate Array) and/or ASICs (Application Specific Integrated Circuitry) adapted to execute instructions. The processor 8 may be configured to access (e.g., write to and/or read from) the memory 6, which may comprise any kind of volatile and/or nonvolatile memory, e.g., cache and/or buffer memory and/or RAM (Random Access Memory) and/or ROM (Read-Only Memory) and/or optical memory and/or EPROM (Erasable Programmable Read-Only Memory). The processing circuitry 10 may be configured to control any of the methods and/or processes described herein and/or to cause such methods, and/or processes to be performed, e.g., by network node 2. Processor 8 corresponds to one or more processors for performing network node 2 functions described herein. The memory 6 is configured to store data, programmatic software code and/or other information described herein. In some embodiments, the OCO algorithm 12 may include instructions that, when executed by the processor 8 and/or processing circuitry 10, causes the processor 8 and/or processing circuitry 10 to perform the processes described herein with respect to network node 2. The network node 2, configured to execute the algorithm disclosed herein, may be operated by the infrastructure provider. The network node 2 may be in radio communication with one or more other network nodes operated or used by the service providers.

FIG. 12 is a flowchart of an exemplary process for determining a MIMO precoding matrix that minimizes an accumulated deviation from a virtualization demand subject to constraints. The process may be performed by the radio interface 4 and processing circuitry 10. The process includes obtaining a global channel state, the global channel state being based at least in part on a channel state in each cell of a plurality of cells (Block S100). The process also includes communicating a channel state to a corresponding service provider (Block S102). The process further includes receiving a virtual precoding matrix from each service provider of the plurality of services providers, each virtual precoding matrix being based at least in part on the channel state communicated to the corresponding service provider and being further based at least in part on a condition that each service provider of the plurality of services providers is allowed to use all of a plurality of available antennas and available wireless spectrum resources (Block S104). The process also includes executing an optimization procedure to periodically determine a multiple input multiple output, MIMO, precoding matrix that minimizes an accumulated precoding deviation from a virtualization demand subject to constraints, the virtualization demand and the MIMO precoding matrix being based at least in part on the virtual precoding matrices received from the plurality of service providers (Block S106). The process also includes applying the determined MIMO precoding matrix to signals applied to the plurality of available antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of the throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals (Block S108).

FIG. 13 is a flowchart of an alternative exemplary process for determining a MIMO precoding matrix that minimizes an accumulated deviation from a virtualization demand subject to constraints. The process may be performed by the radio interface 4 and processing circuitry 10. The process includes receiving from each of the plurality of service providers a virtual precoder matrix determined by the corresponding service provider (Block S110). The process also includes determining a precoder matrix by minimizing a precoding deviation from a virtualization demand of the network subject to at least one power constraint, the virtualization demand being based at least in part on a product of a channel state matrix and a virtual precoder matrix (Block S112). The process further includes applying the determined precoder matrix to signals applied to a plurality of antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals (Block S114).

According to one aspect, a method of online coordinated multi-cell precoding for a network node 2 configurable at least in part by an infrastructure provider is provided. The network node 2 is configured to facilitate sharing of wireless network infrastructure resources by a plurality of service providers. The method includes obtaining, via the processing circuitry 10 a global channel state, the global channel state being based at least in part on a channel state in each cell of a plurality of cells and communicating a channel state to a corresponding service provider. The method also includes receiving, via the radio interface 4, a virtual precoding matrix from each service provider of the plurality of services providers, each virtual precoding matrix being based at least in part on the channel state communicated to the corresponding service provider and being further based at least in part on a condition that each service provider of the plurality of services providers is allowed to use all of a plurality of available antennas and available wireless spectrum resources. The method further includes executing, via the processing circuitry 10, an optimization procedure to periodically determine a multiple input multiple output, MIMO, precoding matrix that minimizes an accumulated precoding deviation from a virtualization demand subject to constraints, the virtualization demand and the MIMO precoding matrix being based at least in part on the virtual precoding matrices received from the plurality of service providers. The method includes applying, via the processing circuitry 10, the determined MIMO precoding matrix to signals applied to the plurality of available antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of the throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

According to this aspect, in some embodiments, the optimization procedure includes periodically determining a gradient of a convex loss function and providing the determined gradient as feedback to the determination of the MIMO precoding matrix. In some embodiments, feedback in the optimization procedure is allowed to be delayed for multiple time slots, received out of order and/or partly missing within an update period. In some embodiments, the optimization procedure includes an online projected gradient ascent algorithm that provides O(√{square root over (T)}) regret and O(1) long term constraint violation, where T is a total time horizon over which multiple updates of the determined MIMO precoding matrix occur. In some embodiments, the accumulated precoding deviation is determined, via the processing circuitry 10, according to f_t(x_t) where T is a time horizon, x_t is a decision in a sequence of decisions made by the network node 2, f_t(x_t) is a convex loss function, f_t(x_t) is an accumulated loss and x⁰arg min_{x∈ X0}f_t(x) is the argument of f_t(x) that produces a minimum value of f_t(x). In some embodiments, the method further includes dividing, via the processing circuitry 10, a total time horizon T into update periods, each update period having a duration of T_o time slots, T_o being at least one time slot, and updating the MIMO precoding matrix at a beginning or end of each update period. In some embodiments, at a beginning of each update period, a decision is taken, via the processing circuitry 10, from a known convex decision space and a loss is determined by an end of the duration of T_o time slots based at least in part on the decision, the loss being based at least in part on a convex loss function. In some embodiments, the virtualization demand is further based at least in part on past channel states. In some embodiments, the constraints include long term transmit power constraints and short term transmit power constraints. In some embodiments, the MIMO precoding matrix is determined to provide sub-linear T-slot regret with partial feedback on the accumulated precoding deviation from the virtualization demand, where T is a total time horizon.

According to another aspect, a network node 2 configured for online coordinated multi-cell precoding, the network node 2 configurable at least in part by an infrastructure provider is provided. The network node 2 is configured to facilitate sharing of wireless network infrastructure resources by a plurality of service providers. The network node 2 includes processing circuitry 10 configured to: obtain a global channel state, the global channel state being based at least in part on a channel state in each cell of a plurality of cells and communicate a channel state to a corresponding service provider. The processing circuitry 10 is further configured to receive a virtual precoding matrix from each service provider of the plurality of services providers, each virtual precoding matrix being based at least in part on the channel state communicated to the corresponding service provider and being further based at least in part on a condition that each service provider of the plurality of services providers is allowed to use all of a plurality of available antennas and available wireless spectrum resources. The processing circuitry 10 is further configured to execute an optimization procedure to periodically determine a multiple input multiple output, MIMO, precoding matrix that minimizes an accumulated precoding deviation from a virtualization demand subject to constraints, the virtualization demand and the MIMO precoding matrix being based at least in part on the virtual precoding matrices received from the plurality of service providers. The processing circuitry 10 is further configured to apply the determined MIMO precoding matrix to signals applied to the plurality of available antennas to achieve a sum of throughputs for plurality of service providers that is greater than a sum of the throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

According to this aspect, in some embodiments, the optimization procedure includes periodically determining a gradient of a convex loss function and providing the determined gradient as feedback to the determination of the MIMO precoding matrix. In some embodiments, feedback in the optimization procedure is allowed to be delayed for multiple time slots, received out of order and/or partly missing within an update period. In some embodiments, the optimization procedure includes an online projected gradient ascent algorithm that provides O(√{square root over (T)}) regret and O(1) long term constraint violation, where T is a total time horizon over which multiple updates of the determined MIMO precoding matrix occur. In some embodiments, the accumulated precoding deviation is determined, via the processing circuitry 10, according to f_t(x_t) where T is a time horizon, x_t is a decision in a sequence of decisions made by an agent, f_t(x_t) is a convex loss function, f_t(x_t) is an accumulated loss and x⁰arg min_x∈X0f_t(x) is the argument of f_t(x) that produces a minimum value of f_t(x). In some embodiments, the processing circuitry 10 is further configured to divide a total time horizon T into update periods, each update period having a duration of T_o time slots, T_o being at least one timeslot, and updating the MIMO precoding matrix at a beginning or end of each update period. In some embodiments, at a beginning of each update period, a decision is taken, via the processing circuitry 10, from a known convex decision space and a loss is determined by an end of the duration of T_o time slots based at least in part on the decision, the loss being based at least in part on a convex loss function. In some embodiments, the virtualization demand is further based at least in part on past channel states. In some embodiments, the constraints include long term transmit power constraints and short term transmit power constraints. In some embodiments, the MIMO precoding matrix is determined to provide sub-linear T-slot regret with partial feedback on the accumulated precoding deviation from the virtualization demand, where T is a total time horizon.

According to yet another aspect, a network node 2 configured for sharing of wireless network resources among a plurality of service providers is provided. The network node 2 includes processing circuitry 10 configured to receive from each of the plurality of service providers a virtual precoder matrix determined by the corresponding service provider. The processing circuitry 10 is further configured to determine a precoder matrix by minimizing a precoding deviation from a virtualization demand of the network subject to at least one power constraint, the virtualization demand being based at least in part on a product of a channel state matrix and a virtual precoder matrix. The processing circuitry 10 is further configured to apply the determined precoder matrix to signals applied to a plurality of antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

According to this aspect, in some embodiments, determining the precoder matrix includes solving a minimization problem that includes determining a gradient of a convex loss function. In some embodiments, determining the precoder matrix includes decomposing the minimization problem into a number of subproblems, each subproblem involving a gradient of a convex loss function of a local precoding matrix. In some embodiments, the minimization problem includes comparing a function of the determined precoder matrix to an offline fixed precoding strategy. In some embodiments, the virtual precoder matrices received from the plurality of service providers are each based at least in part on a condition that each service provider is allowed to use all of a plurality of available antennas and wireless spectrum resources. In some embodiments, the precoding deviation from virtualization demand is determined, via the processing circuitry 10, at least in part by:f_t(V)∥H_tV−D_t∥_F² where V is a past precoding matrix, H_t is the channel state and D_t is the virtualization demand.

According to another aspect, a method of online coordinated multi-cell precoding for a network node 2 configurable at least in part by an infrastructure provider is provided, where the network node 2 is configured to facilitate sharing of wireless network infrastructure resources by a plurality of service providers. The method includes receiving, via the radio interface 4, from each of the plurality of service providers a virtual precoder matrix determined by the corresponding service provider. The method includes determining, via the processing circuitry 10, a precoder matrix by minimizing a precoding deviation from a virtualization demand of the network subject to at least one power constraint, the virtualization demand being based at least in part on a product of a channel state matrix and a virtual precoder matrix. The method also includes applying, via the processing circuitry 10, the determined precoder matrix to signals applied to a plurality of antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

According to this aspect, in some embodiments, determining the precoder matrix includes solving a minimization problem that includes determining a gradient of a convex loss function. In some embodiments, determining the precoder matrix includes decomposing the minimization problem into a number of subproblems, each subproblem involving a gradient of a convex loss function of a local precoding matrix. In some embodiments, the minimization problem includes comparing a function of the determined precoder matrix to an offline fixed precoding strategy. In some embodiments, the virtual precoder matrices received from the plurality of service providers are each based at least in part on a condition that each service provider is allowed to use all of a plurality of available antennas and wireless spectrum resources. In some embodiments, the precoding deviation from virtualization demand is determined, via the processing circuitry 10, at least in part by:f_t(V)∥H_tV−D_t∥_F² where V is a past precoding matrix, H_t is the channel state and D_t is the virtualization demand.

ABBREVIATIONS

5G: Fifth Generation

CDI: Channel Distribution Information

CSI: Channel State Information

C-RAN: Cloud Radio Networks

GP: Geometric Programming

InP: Infrastructure Provider

KKT: Karush-Kuhn-Tucker

LTE: Long-Term Evolution

MIMO: Multiple Input Multiple Output

NOMA: Non-orthogonal Multiple Access

OCO: Online Convex Optimization

OFDM: Orthogonal Frequency Division Multiplexing

SP: Service Provider

WNV: Wireless Network Virtualization

As will be appreciated by one of skill in the art, the concepts described herein may be embodied as a method, data processing system, and/or computer program product. Accordingly, the concepts described herein may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects all generally referred to herein as a “circuit” or “module.” Furthermore, the disclosure may take the form of a computer program product on a tangible computer usable storage medium having computer program code embodied in the medium that can be executed by a computer. Any suitable tangible computer readable medium may be utilized including hard disks, CD-ROMs, electronic storage devices, optical storage devices, or magnetic storage devices.

Some embodiments are described herein with reference to flowchart illustrations and/or block diagrams of methods, systems and computer program products. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer readable memory or storage medium that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instruction means which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. It is to be understood that the functions/acts noted in the blocks may occur out of the order noted in the operational illustrations. For example, two blocks shown in succession may in fact be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality/acts involved. Although some of the diagrams include arrows on communication paths to show a primary direction of communication, it is to be understood that communication may occur in the opposite direction to the depicted arrows. Computer program code for carrying out operations of the concepts described herein may be written in an object oriented programming language such as Java® or C++. However, the computer program code for carrying out operations of the disclosure may also be written in conventional procedural programming languages, such as the “C” programming language. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer. In the latter scenario, the remote computer may be connected to the user's computer through a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider).

Many different embodiments have been disclosed herein, in connection with the above description and the drawings. It will be understood that it would be unduly repetitious and obfuscating to literally describe and illustrate every combination and subcombination of these embodiments. Accordingly, all embodiments can be combined in any way and/or combination, and the present specification, including the drawings, shall be construed to constitute a complete written description of all combinations and subcombinations of the embodiments described herein, and of the manner and process of making and using them, and shall support claims to any such combination or subcombination.

It will be appreciated by persons skilled in the art that the embodiments described herein are not limited to what has been particularly shown and described herein above. In addition, unless mention was made above to the contrary, it should be noted that all of the accompanying drawings are not to scale. A variety of modifications and variations are possible in light of the above teachings without departing from the scope of the following claims.

Claims

1. A method of online coordinated multi-cell precoding for a network node configurable at least in part by an infrastructure provider, the network node being configured to facilitate sharing of wireless network infrastructure resources by a plurality of service providers, the method comprising: obtaining a global channel state, the global channel state being based at least in part on a channel state in each cell of a plurality of cells;communicating a channel state to a corresponding service provider;receiving a virtual precoding matrix from each service provider of the plurality of services providers, each virtual precoding matrix being based at least in part on the channel state communicated to the corresponding service provider and being further based at least in part on a condition that each service provider of the plurality of services providers is allowed to use all of a plurality of available antennas and available wireless spectrum resources;executing an optimization procedure to periodically determine a multiple input multiple output, MIMO, precoding matrix that minimizes an accumulated precoding deviation from a virtualization demand subject to constraints, the virtualization demand and the MIMO precoding matrix being based at least in part on the virtual precoding matrices received from the plurality of service providers; andapplying the determined MIMO precoding matrix to signals applied to the plurality of available antennas to achieve a sum of throughputs for the plurality of service providers that is greater than a sum of the throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

2. The method of claim 1, wherein the optimization procedure includes periodically determining a gradient of a convex loss function and providing the determined gradient as feedback to the determination of the MIMO precoding matrix.

3. The method of claim 2, wherein feedback in the optimization procedure is allowed to be delayed for multiple time slots, at least one of: received out of order; andpartly missing within an update period.

4. The method of claim 1, wherein the optimization procedure includes an online projected gradient ascent algorithm that provides O(√{square root over ())} regret and O(1) long term constraint violation, where T is a total time horizon over which multiple updates of the determined MIMO precoding matrix occur.

5. The method of claim 1, wherein the accumulated precoding deviation is determined according to ft(xt) where T is a time horizon, xt is a decision in a sequence of decisions made by the network node, ft(xt) is a convex loss function, ft(xt) is an accumulated loss and xºarg minx∈X0ft(x) is the argument of ft(x) that produces a minimum value of ft(x).

6. The method of claim 1, further comprising dividing a total time horizon T into update periods, each update period having a duration of To time slots, To being at least one time slot, and updating the MIMO precoding matrix at a beginning or end of each update period.

7. The method of claim 6, wherein at a beginning of each update period, a decision is taken from a known convex decision space and a loss is determined by an end of the duration of To time slots based at least in part on the decision, the loss being based at least in part on a convex loss function.

8. The method of claim 1, wherein the virtualization demand is further based at least in part on past channel states.

9. The method of claim 1, wherein the constraints include long term transmit power constraints and short term transmit power constraints.

10. The method of claim 1, wherein the MIMO precoding matrix is determined to provide sub-linear T-slot regret with partial feedback on the accumulated precoding deviation from the virtualization demand, where T is a total time horizon.

11. A network node configured for online coordinated multi-cell precoding, the network node configurable at least in part by an infrastructure provider, the network node being configured to facilitate sharing of wireless network infrastructure resources by a plurality of service providers, the network node including processing circuitry configured to: obtain a global channel state, the global channel state being based at least in part on a channel state in each cell of a plurality of cells;communicate a channel state to a corresponding service provider;receive a virtual precoding matrix from each service provider of the plurality of services providers, each virtual precoding matrix being based at least in part on the channel state communicated to the corresponding service provider and being further based at least in part on a condition that each service provider of the plurality of services providers is allowed to use all of a plurality of available antennas and available wireless spectrum resources;execute an optimization procedure to periodically determine a multiple input multiple output, MIMO, precoding matrix that minimizes an accumulated precoding deviation from a virtualization demand subject to constraints, the virtualization demand and the MIMO precoding matrix being based at least in part on the virtual precoding matrices received from the plurality of service providers; andapply the determined MIMO precoding matrix to signals applied to the plurality of available antennas to achieve a sum of throughputs for plurality of service providers that is greater than a sum of the throughputs for the plurality of service providers achievable when the virtual precoder matrices are applied to the signals.

12. The network node of claim 11, wherein the optimization procedure includes periodically determining a gradient of a convex loss function and providing the determined gradient as feedback to the determination of the MIMO precoding matrix.

13. The network node of claim 12, wherein feedback in the optimization procedure is allowed to be delayed for multiple time slots, at least one of: received out of order; andpartly missing within an update period.

14. The network node of claim 11, wherein the optimization procedure includes an online projected gradient ascent algorithm that provides O(√{square root over (T)}) regret and O(1) long term constraint violation, where T is a total time horizon over which multiple updates of the determined MIMO precoding matrix occur.

15. The network node of claim 11, wherein the accumulated precoding deviation is determined according to ft(xt) where T is a time horizon, xt is a decision in a sequence of decisions made by the network node, ft(xt) is a convex loss function, ft(xt) is an accumulated loss and xºarg minx∈X0ft(x) is the argument of ft(x) that produces a minimum value of ft(x).

16. The network node of claim 11, wherein the processing circuitry is further configured to divide a total time horizon T into update periods, each update period having a duration of To time slots, To being at least one timeslot, and updating the MIMO precoding matrix at a beginning or end of each update period.

17. The network node of claim 16, wherein at a beginning of each update period, a decision is taken from a known convex decision space and a loss is determined by an end of the duration of To time slots based at least in part on the decision, the loss being based at least in part on a convex loss function.

18. The network node of claim 11, wherein the virtualization demand is further based at least in part on past channel states.

19. The network node of claim 11, wherein the constraints include long term transmit power constraints and short term transmit power constraints.

20. The network node of claim 11, wherein the MIMO precoding matrix is determined to provide sub-linear T-slot regret with partial feedback on the accumulated precoding deviation from the virtualization demand, where T is a total time horizon.