WorksEngineering
Computationally Constrained First-Order Prediction-Correction Approach for Time-Varying Convex Optimization
Abstract
Time-varying optimization methods often predict the locations of time-varying optima, then sample the objective function being minimized, and then compute corrections to those predictions. However, existing methods typically do not account for settings in which the number of computations is limited. Therefore, in this work we formulate a prediction-correction algorithm that is designed to obey a budget constraint on the total number of iterations it can execute over a given time horizon. We first identify a threshold on the average number of computations between samples of the objective, above which a bound from prior work certifies contraction of iterates toward a minimizer. Below this threshold, we introduce a finite-horizon contraction certificate, which we use to give budget conditions under which the tracking error admits an explicit bound that decays at a prescribed rate. We also present an algorithm to determine how many prediction and correction steps should be executed for each sample of the objective, while obeying a given budget on the allowable number of computations. Empirically, we show in a simulation that our approach reduces cumulative tracking error by 45.5 % relative to a comparable baseline approach.
Introduction
Time-varying optimization problems arise in robotics [1], [2], power systems [3], transportation networks [4], communication systems [5], and distributed systems [6], [7]. A common way to solve such problems is with a prediction-correction method [8], which uses the objective sampled at a prior time to predict how the solution will evolve in the future. Then, after resampling the objective, corrections are made to this prediction to account for differences between the predicted and actual movements of the solution. Such methods often use gradient or Newton steps to correct for those differences [1], [2], [8]. After completing some number of correcting steps, the prediction-correction cycle is repeated by predicting the subsequent evolution of the solution, then sampling the objective, and then correcting those predictions. This class of method has been shown to track minimizers with error that is acceptable in a wide range of applications [6], [7], [9].
Prior works such as [1, 2, 3], [6]–[10] have used prediction-correction methods in applications including target tracking, model-predictive control, and various decentralized formulations. All of these works (and many others) use fixed numbers of prediction steps and correction steps between samples of the objective. However, in practice it might not be possible to execute the desired number of computations, e.g., due to constraints on resources such as battery power, which can limit the total number of computations that can be executed over a given time horizon. Such constraints have driven interest in various forms of resource awareness across different settings. For example, existing works in [11]–[18] have investigated resource awareness in multi-agent control, connected and autonomous electric vehicles, and event-triggered control, but to the best of our knowledge there has not been a systematic study of time-varying optimization under computational constraints.
Therefore, in this paper, we develop a computationally constrained prediction-correction algorithm for time-varying optimization. Computation constraints are encoded as an upper bound on the total number of prediction and correction steps that can be executed over a given time horizon, which we call the “budget” for a problem. We consider a first-order prediction-correction algorithm that uses gradient steps for both predictions and corrections.
We first identify a critical threshold on the number of computations that can be executed for each sample of the objective function. Then we show that if the budget permits at least that many computations to be executed for each sample of the objective, then we can certify that the iterates of the first-order algorithm contract toward a minimizer.
In contrast, if the threshold we derive cannot be met for some samples of the objective, the aforementioned contraction guarantee no longer applies. However, we show that one can still allocate prediction and correction steps so that the distance to the minimizer admits a different form of explicit bound that decreases at a prescribed rate. Then, we develop an algorithm for determining how many prediction and correction steps should be completed for each sample of the objective function, and this algorithm is designed to attain the error bounds we derive on the distance to a time-varying optimum.
Summary of Contributions
The contributions of this paper are:
We identify a critical threshold on the average number of computations per sample of the objective. Above this threshold, we guarantee that a bound on the distance to the minimizer converges geometrically to a neighbor-
Chuy and Hale were supported by the Office of Naval Research under grant N00014-26-1-2068, and Musavi and Ornik were supported by the Office of Naval Research under grant N00014-25-1-2519 and N00014-25-1-2369. hood whose size is proportional to the time that elapses between taking samples of the objective (Corollary 1).
Below this threshold, we introduce a finite-horizon contraction certificate that encodes a desired rate of decay in the distance from iterates to a minimum. Then we derive budget conditions under which this certificate can be satisfied and thus ensure that the distance to a minimizer decreases at a user-specified rate (Theorems 1 and 2).
We develop an algorithm that determines how many prediction and correction steps to execute for each sample of the objective while obeying a total computation budget constraint (Algorithm 1, Theorems 2 and 3).
We show in simulation that our method tracks time-varying optima with 45.5% less cumulative error than a comparable baseline approach § 5.
The rest of this paper is organized as follows. Section II provides background and a problem formulation. Section III develops the computationally constrained algorithm, and Section IV computes error bounds. Then, § 5 presents simulations, and Section VI concludes the paper.
Notation
We use to denote the reals and to denote the nonnegative reals. For an integer , we use to denote the set of integers greater than or equal to . We denote the 2-norm of a vector by . We use to denote the identity matrix. For a function that takes arguments , we denote its gradient with respect to by , and we denote its derivative with respect to by . We use to denote the Hessian of with respect to , and is the time derivative of .
Preliminaries and Problem Formulation
In this section, we first define the time-varying optimization problem in Section II-A. Then in Section II-B, we describe the prediction–correction method. Finally, in Section II-C, we formulate the problem that we address in this paper.
Time-Varying Optimization
Let be a time-varying function, and let be a finite time horizon. We make the following assumption on .
Assumption 1. The function is twice continuously differentiable in and there exist constants such that
as well as a constant such that
Assumption 1 ensures that is uniformly strongly convex and smooth, which guarantees that it has a unique minimizer for each . The inequalities in (1) provide uniform curvature bounds over time, and the constant bounds the temporal variation of the gradient. The bound in (2) means that the time derivative of is bounded.
Under Assumption 1, consider time-varying convex optimization problems of the form
where is a decision variable. The objective is only known at past and current times , and future objectives are not known .
Suppose that the objective can only be accessed at discrete time instances. Then, the problem (3) can be considered in discrete time. Following [19, 20], the objective is sampled at the time instances , where is a constant sampling interval and for a finite time horizon such that . This yields the sequence of problems
which is the problem considered in the remainder of this paper. At each , the current objective and its derivatives at , including and , are known.
Prediction-Correction Method
In this section we describe the prediction–correction method of [2], which proceeds as follows. Starting with an initial estimate , the method generates estimates of the optimal trajectory in (4). For , the method uses the estimate of to predict by executing prediction steps. Then, the objective is sampled, and correction steps are executed to correct the prediction. The final correction step produces , from which this process repeats. The prediction and correction steps are described in Sections II-B.1 and II-B.2. In [2], the numbers of prediction and correction steps are the same for every ; here, they may vary with .
Prediction
Prediction: Between two sampling times and , prediction steps are executed to predict from . A first-order expansion in space and time gives
where , , and . Since is not known before it is sampled, solving exactly may not be practical. Instead, as in [2], a relaxed optimality condition is imposed by seeking a point that satisfies with . This relaxation leads to the prediction step
for , where is the th prediction iterate computed for the th sample of the objective, , and is the prediction step size.
Correction
After all prediction steps are completed, the objective function is sampled, and correction steps are executed. Initialized at , the correction step is
for , where is the th correction iterate computed for the th sample of the objective, and is the correction step size. After correction steps, , and then this process of prediction and correction repeats.
Problem Formulation
Definition 1. An allocation is a collection , with , of the numbers of prediction and correction steps (5)–(6) executed between consecutive sampling times and .
The sum is the number of iterations used to compute from . For a given allocation , let denote the iterates generated by (5)–(6) from an initial estimate . We define the tracking error of this allocation as
We consider a total iteration budget over the sampling intervals. Each prediction iteration (5) and each correction iteration (6) incurs one unit of cost, while sampling the objective and evaluating , , and do not count toward the budget. Our goal is to solve the following problem.
Problem 1. Given an iteration budget , determine an allocation with bounded tracking errors for all such that
The budget constraint permits the per-sample totals to vary across sampling intervals. To address Problem 1, we first characterize how the allocation affects the tracking error. We then use this characterization to develop allocation rules that meets the total iteration budget and admit a finite-horizon tracking bounds.
Computationally Constrained Prediction-Correction Method
In this section, we develop a computationally constrained allocation strategy for prediction and correction iterations that 1) provides a finite-horizon tracking guarantee and 2) satisfies the budget constraint in (8). We first present tracking error bounds for a general allocation . Then, given a budget , we derive allocations under two different regimes.
Tracking-Error Bound and Budget Regimes
Let
where and are the curvature bounds from Assumption 1, and and are the step sizes of the prediction and correction steps in (5) and (6), respectively. The terms and are the contraction factors of the prediction and correction steps. Since , it follows from Assumption 1 that [2]. For a general allocation, the following proposition describes how the tracking error evolves from one sampling time to the next.
Proposition 1 (Prediction-Correction Tracking Error Bound [2]). Suppose Assumption 1 holds, let be any allocation, and let the iterates be generated by the prediction and correction steps in (5)–(6). Then, the following inequality holds:
where
Proof. This result follows from Appendices A and C for case a) in [2]. There, the bound is stated for a fixed number of prediction and correction steps. Because (9) relates only and , the same argument applies at each with in place of the fixed counts.
For ease of notation, let and for . Recursively unrolling (9) gives
Moreover, since , it follows from (10) that
where . The bound in (11) holds for any allocation, and shows that both the initial error and each term in the sum are multiplied by products of the factors . If for all , then these products decay geometrically in . The factor in (10) is nonincreasing in both and , and therefore ensuring requires a sufficient number of iterations per sample. Let us define the minimum such number as
The minimum in (13) is attained, since for sufficiently large and . Since is nonincreasing in each argument, any also admits a pair with and .
For a given horizon and budget , we define two regimes based on :
Regime I: and Regime II: .
Since the budget is shared across all samples, the regimes depend on the average budget per sample . In Regime I, the budget allows for all . In Regime II, it does not, and therefore any allocation satisfying (8) has at least one with . The following subsections develop allocations for each regime.
Regime I: Sufficient Average Budget
Suppose that , and define
where is the largest number of iterations that makes all ’s equal while obeying the budget constraint in (8). Since is an integer and , we have . Define . We choose a pair
and we set for all . Since , the set is nonempty. This uniform allocation uses total iterations over the full time horizon, and therefore it satisfies (8).
Regime II: Insufficient Average Budget
If , then cannot be guaranteed for all . If the budget is sufficiently large, an allocation can still ensure that the products of the factors in (11) decay at a prescribed rate, which yields an explicit tracking bound over . We formalize this notion with the following finite-horizon contraction condition.
Finite-horizon contraction certificate
Finite-horizon contraction certificate: For user-specified choices of and , we consider the condition
for all with . We call (15) a certificate because, when it holds, the tracking error can be shown to be bounded. This bound is derived in Section IV.
To enforce (15) while choosing one iteration at a time, we quantify the extent to which exceeds at each iteration. Since and for all choices of and , we define
Intuitively, measures the extent to which a given pair ensures at iteration . For any , if , then , and if , then . We also define
which we use to measure the accumulated positive values of and hence the cumulative failure of pairs to ensure over time. Starting with , we set
The term increases if at iteration , decreases or remains at zero if at iteration , and is reset to zero whenever the accumulated sum becomes negative at iteration . The following lemma shows that one can enforce (15) by bounding from (17).
Lemma 1 (Necessary and sufficient condition for (15)). Let be as defined in (17). Then the certificate (15) holds for all satisfying if and only if
Proof. See Appendix A.
Lemma 1 provides a necessary and sufficient condition for the certificate (15) to hold. Since the budget is shared across iterations, maintaining this condition requires reserving enough iterations for the remaining samples, which we quantify next through the reserve function.
Reserve function
Reserve function: Let denote the number of remaining samples of the objective function, and consider . We define the reserve function as
where the constraint keeps , and counts the minimum total number of prediction and correction iterations needed over the remaining samples. We set , since no iterations are needed when no samples remain, and if no pair satisfies the constraint in (18). The value of also depends on the certificate parameters and from (15) and the function from (10), and it is solved offline for all , before the horizon begins. In practice, we compute on a grid of possible values of , rounding each transition up to the next grid point, which can only increase and therefore preserves the guarantees.
Computing Allocations
Computing Allocations: The reserve function determines how many iterations must be saved for future samples so that (15) can still be satisfied. It remains to choose and at each . The remaining budget at step for is
Then the feasible set for choosing and at iteration is
where the first condition ensures , so that, by Lemma 1, the certificate (15) continues to hold after iteration , and the second condition ensures that the budget constraint in (8) is satisfied.
To choose among feasible pairs, we estimate the two terms in (9) using information available at . Since cannot be evaluated online, we estimate it with
By strong convexity of , using the fact that the solution satisfies , we find that . Combined with the Cauchy–Schwarz inequality, this gives . To estimate in (10), we define
which differs from in (10) by using the value of at instead of the worst-case bound from Assumption 1.
At iteration , we then choose and using
The objective in (22) is a heuristic score that mimics the right-hand side of (9) using information available at . The procedure for allocating the budget in Regime II is summarized in Algorithm 1.
Tracking Analysis
In this section, we analyze the allocation strategies derived for Regime I in (14) and for Regime II in (22).
Regime I Tracking Guarantees
With the allocation strategy from (14), the tracking error of the prediction-correction algorithm that executes prediction steps using (5) and correction steps using (6) follows directly from Proposition 1.
Corollary 1 (Tracking in Regime I). Suppose Assumption 1 holds and , where is from (8) and is from (13). Consider the allocation of prediction and correction steps from (14), and define . Then the budget constraint in (8) is satisfied, , and we have
for , where is from (12).
| Algorithm 1 Computationally Constrained Allocation |
| Require: Initial iterate , horizon , budget from (8), certificate parameters and from (15), and from (18), computed offline |
| 1: Verify |
| 2: , |
| 3: for do |
| 4: Compute and from (20) |
| 5: Compute using (21) |
| 6: Form using (19) |
| 7: Select according to (22) |
| 8: Perform prediction iterations from (5) using the objective |
| 9: Sample the objective |
| 10: Perform correction iterations from (6) using the objective to obtain |
| 11: |
| 12: |
| 13: end for |
Table 1.
Proof. The budget constraint holds by Section III-B, and because . Since the allocation is constant, for all , and from (12), for all . Thus, the result follows from (11).
Corollary 1 shows that under Regime I a uniform allocation suffices: the tracking error bound converges geometrically to a neighborhood of size . This neighborhood size depends on the sampling interval from Section II-A, which can be changed by a user. We now bound the tracking error in Regime II.
Regime II Feasibility and Tracking Guarantees
We first provide a sufficient condition under which (15) can be satisfied.
Theorem 1 (Certificate feasibility). Suppose Assumption 1 holds, and let and be given. If , where is from (18), then there exists an allocation that satisfies both the budget constraint (8) and the certificate (15).
Proof. See Appendix B.
Under the condition of Theorem 1, Algorithm 1 generates such an allocation, and its tracking error is bounded as follows.
Theorem 2 (Tracking in Regime II). Suppose Assumption 1 holds, we define
and holds. Then, is nonempty for all . Moreover, the allocation generated by Algorithm 1 satisfies the budget constraint in (8), and
where and are from (15) and is from (12).
Proof. See Appendix C.
Theorem 2 provides a bound on the tracking error incurred by allocations produced by Algorithm 1. As in Regime I, the bound converges geometrically to a neighborhood whose size is proportional to the sampling interval .
A simpler condition using uniform allocation
Checking the condition in Theorem 1 requires solving the dynamic program (18). A simpler sufficient condition is obtained by restricting to uniform allocations for all , which yields a budget condition, and an allocation that satisfies it, without computing .
Theorem 3 (Uniform-allocation budget condition). Suppose Assumption 1 holds. Define
where and are from (15). Also define
and let be any pair such that . If , then the uniform allocation for satisfies the budget constraint in (8), the certificate in (15), and the bound in (24). Moreover, , and therefore Algorithm 1 can be used.
Proof. See Appendix § 7.4.
Since , the condition in Theorem 3 is more conservative than that in Theorem 1, but it does not require computing . The results in Theorem 3 show that we can still attain the bound in (24) even while using a uniform allocation of total iterations.
Moreover, since the allocation is the same for every , computing reduces to finding, for each number of prediction steps , the minimum number of correction steps that satisfies . Since , solving this inequality for gives
where is from (10), which implies that
Then the uniform allocation calculation is reduced to an integer search over prediction counts, without using a dynamic program.
Numerical Experiments
In this section, we compare our adaptive allocation against a uniform allocation under the same total iteration budget .
Problem and Experimental Setup
Consider the following unconstrained problem, adapted from [2], in which the minimizer follows a moving reference whose speed changes over the horizon:
where , is a unit vector, , and the reference traces a closed curve with angular speed
which produces a slow–fast–slow pattern of temporal variation. The objective in (25) satisfies Assumption 1 since is uniformly bounded between and with and , and since gives . The optimizer is computed numerically and is used only to evaluate tracking performance.
We use , , , , , , , and , which gives , , and . We also use , , , , , , , and . The step size minimizes , and the remaining parameters are chosen so that the fast phase of (26) occurs within the horizon and the experiment lies in Regime II. Here while , and therefore the experiment lies in Regime II. For the adaptive allocation, we compute the reserve function in (18) on a uniform grid of values with spacing . The uniform baseline uses for all . Both methods use , which gives , and we set . To compare the two methods, we use the cumulative tracking errors
where is from (7), and we refer to as the terminal error. The initialization error is excluded from and because the prediction-correction method cannot change it.
Adaptive and Uniform Allocation
Under the above conditions, adaptive allocation reduces the cumulative tracking error by about and the maximum error by about $64.7% relative to the uniform baseline, at the cost of a larger terminal error ( for the adaptive allocation vs. for the uniform baseline). This behavior follows from the online allocation rule in (22). Since and are nondecreasing in and , the rule spends as many iterations as allows. As a result, the adaptive allocation front-loads the budget: it uses 10 iterations per interval during the first seven intervals, while reserving enough budget to satisfy the certificate, and it then tapers to two prediction steps per interval over the final five intervals. Spending more iterations early reduces the large initial error quickly, which lowers and , while the small allocation at the end leads to the larger terminal error. The uniform allocation uses throughout.
Sensitivity to Initialization and Budget
This subsection examines whether the advantage of the adaptive allocation persists across initial errors and budgets. We first repeat the preceding experiment with for ten initial-error magnitudes , , holding the direction of fixed. The reduction in grows with , from about 7.4% at to about 67.7% at . The terminal error of each method is the same across this range (0.013 for the adaptive allocation and 0.002 for the uniform baseline), and therefore terminal error alone does not reflect how well either method tracked the trajectory.
Next, we examine the cumulative error between the uniform baseline and the adaptive allocation with the budget ranging over . These values all lie above the value and below the Regime I budget . We consider the initial error values , which are shown in Figure 1. The adaptive allocation achieves a lower cumulative tracking error for all budgets and initial errors considered.

Figure 1. Cumulative error vs. budget , adaptive (solid) against uniform (dashed), at four values of . The drop in the uniform method between and is due to the increase in from 6 to 7. The adaptive allocation outperforms the uniform allocation for all budgets and initial errors considered.
Finally, we examine the error reduction between the adaptive allocation and the uniform baseline with a budget ranging over and values of the initial error ranging over . We find that the error reduction ranges from 7.4% (, ) to 76.1% (, ), with an average of 35.4%, which is shown in Figure 2. The decrease in this reduction that occurs between a budget of 135 and 140 corresponds to the improvement in the cumulative tracking error of the uniform baseline due to the increase in from 6 (with an allocation of to 7 (with an allocation of ) iterations. The advantage of the adaptive method is largest for large initial errors and for budgets below where the uniform baseline is limited to six iterations per sample, and the adaptive method produces a decrease in error across all values of the budget and initial error.

Figure 2. Reduction in cumulative tracking error over the joint grid. The adaptive method provides the largest decreases in error for large and for budgets below , and it decreases the error for all values of these parameters considered.
Conclusion
This paper presented a method of adaptive allocation for the number of prediction and correction steps in a prediction-correction algorithm that is subject to computational constraints. We showed that when the average computational budget per sample exceeds a certain threshold, a uniform allocation suffices for a decaying tracking error bound. When it does not, we introduced a finite-horizon contraction certificate that yields budget conditions and an allocation algorithm that enforce explicit tracking error bounds. Simulations show that the adaptive allocation reduces cumulative tracking error compared with a uniform allocation across all problems considered. Future work includes extending this approach to other prediction schemes.
Appendix
Proof of Lemma 1
The recursion in (17) gives
Consequently, using (16), we see that having for every in is equivalent to having
for every satisfying . Exponentiating proves the equivalence.
Proof of Theorem 1
Suppose . We construct a feasible allocation using the minimizers in (18). Since and the values of from (18) are nonnegative integers or , the minimum defining is attained. Let be a minimizer and , so that
which implies . Repeating the same argument at , where is from (17) for , yields
This allocation satisfies the budget constraint in (8) since . Moreover, each minimizer satisfies the constraint in (18), so the certificate (15) holds by Lemma 1.
Proof of Theorem 2
We show by induction that and for all . For , we have . Suppose . Then , and the minimum in (18) defining is attained by some pair . This pair satisfies and
Then . For the selected , the second condition in (19) gives , which completes the induction.
Since , we have , and (8) holds. The first condition in (19) gives for all , and the certificate in (15) holds by Lemma 1. Using the tracking error recursion in (11) and the certificate in (15), where an empty product equals , we find
Since from (23), we find
The result follows from bounding using (12).
Proof of Theorem 3
Consider the uniform allocation for . Since , the total budget satisfies . Moreover, by the definition of , we have . Therefore, for all integers and such that , we have
where the last inequality follows from and . Hence, the certificate in (15) is satisfied. Following the same steps from (27) to (28) in the proof of Theorem 2, we find
Finally, by Lemma 1, the uniform allocation keeps for all , and therefore it is feasible in (18) at every stage. Hence $
References
- [1]A. Simonetto, A. Mokhtari, A. Koppel, G. Leus, and A. Ribeiro, “A class of prediction-correction methods for time-varying convex optimization,” IEEE Transactions on Signal Processing, vol. 64, no. 17, pp. 4576–4591, 2016.DOI
- [2]A. Simonetto and E. Dall’Anese, “Prediction-correction algorithms for time-varying constrained optimization,” IEEE Transactions on Signal Processing, vol. 65, no. 20, pp. 5481–5494, 2017.DOI
- [3]Z. Lin, J. Hou, and X. Zeng, “Optimal prediction-correction algorithm using sparse linear extrapolation for time-varying optimization,” IEEE Transactions on Signal Processing, vol. 74, pp. 2564–2578, 2026.DOI
- [4]A. Simonetto, E. Dall’Anese, S. Paterlain, G. Leus, and G. B. Giannakis, “Time-varying convex optimization: Time-structured algorithms and applications,” Proceedings of the IEEE, vol. 108, no. 11, pp. 2032–2048, 2020.arxiv.org/abs/2006.08500
- [5]G. Bianchin, J. Cortés, J. I. Poveda, and E. Dall’Anese, “Time-varying optimization of LTI systems via projected primal-dual gradient flows,” IEEE Transactions on Control of Network Systems, vol. 9, no. 1, pp. 474–486, 2022.arxiv.org/abs/2101.01799
- [6]A. Simonetto, “Dual prediction-correction methods for linearly constrained time-varying convex programs,” IEEE Transactions on Automatic Control, vol. 64, no. 8, pp. 3355–3361, 2019.arxiv.org/abs/1709.05850
- [7]A. Simonetto, A. Koppel, A. Mokhtari, G. Leus, and A. Ribeiro, “Decentralized prediction-correction methods for networked time-varying convex optimization,” IEEE Transactions on Automatic Control, vol. 62, no. 11, pp. 5724–5738, 2017.arxiv.org/abs/1602.01716
- [8]——, “Prediction-correction methods for time-varying convex optimization,” in 49th Asilomar Conference on Signals, Systems and Computers, 2015, pp. 666–670.DOI
- [9]S. Paternain, M. Morari, and A. Ribeiro, “A prediction-correction method for model predictive control,” in 2018 American Control Conference (ACC), 2018, pp. 4189–4194.DOI
- [10]N. Bastianello, R. Carli, and A. Simonetto, “Extrapolation-based prediction-correction methods for time-varying convex optimization,” Signal Processing, vol. 210, 2023, art. no. 109089.DOI
- [11]P. O. M. Skaf, D. Q. Mayne, and J. B. Rawlings, “Suboptimal model predictive control (feasibility implies stability),” IEEE Transactions on Automatic Control, vol. 44, no. 3, pp. 648–654, 1999.
- [18]H. Zhang, L. Yuan, and X. Cao, “Event-triggered nonlinear model predictive control for energy-aware UAV obstacle avoidance,” IEEE Internet of Things Journal, 2026, early access, doi: 10.1109/JIOT.2026.3728027.DOI
- [19]Y. Nesterov, Lectures on Convex Optimization, 2nd ed. Cham, Switzerland: Springer, 2018.DOI
- [20]S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, UK: Cambridge University Press, 2004.