Abstract
We consider the generalized trust region subproblem (GTRS) of minimizing a nonconvex quadratic objective over a nonconvex quadratic constraint. A lifting of this problem recasts the GTRS as minimizing a linear objective subject to two nonconvex quadratic constraints. Our first main contribution is structural: we give an explicit description of the convex hull of this nonconvex set in terms of the generalized eigenvalues of an associated matrix pencil. This result may be of interest in building relaxations for nonconvex quadratic programs. Moreover, this result allows us to reformulate the GTRS as the minimization of two convex quadratic functions in the original space. Our next set of contributions is algorithmic: we present an algorithm for solving the GTRS up to an \(\epsilon \) additive error based on this reformulation. We carefully handle numerical issues that arise from inexact generalized eigenvalue and eigenvector computations and establish explicit running time guarantees for these algorithms. Notably, our algorithms run in linear (in the size of the input) time. Furthermore, our algorithm for computing an \(\epsilon \)-optimal solution has a slightly-improved running time dependence on \(\epsilon \) over the state-of-the-art algorithm. Our analysis shows that the dominant cost in solving the GTRS lies in solving a generalized eigenvalue problem—establishing a natural connection between these problems. Finally, generalizations of our convex hull results allow us to apply our algorithms and their theoretical guarantees directly to equality-, interval-, and hollow-constrained variants of the GTRS. This gives the first linear-time algorithm in the literature for these variants of the GTRS.
Similar content being viewed by others
Notes
In fact, this assumption can be made without loss of generality; see Remark 1.
Our definition of accuracy is presented in (9).
Recall that a convex quadratic function \(x^\top A x+ 2b^\top x + c\) is 2L-smooth if and only if \(A\preceq LI\).
References
Adachi, S., Nakatsukasa, Y.: Eigenvalue-based algorithm and analysis for nonconvex QCQP with one constraint. Math. Program. 173(1), 79–116 (2019)
Ben-Tal, A., den Hertog, D.: Hidden conic quadratic representation of some nonconvex quadratic optimization problems. Math. Program. 143(1), 1–29 (2014)
Ben-Tal, A., Teboulle, M.: Hidden convexity in some nonconvex quadratically constrained quadratic programming. Math. Program. 72(1), 51–63 (1996)
Buchheim, C., De Santis, M., Palagi, L., Piacentini, M.: An exact algorithm for nonconvex quadratic integer minimization using ellipsoidal relaxations. SIAM J. Optim. 23(3), 1867–1889 (2013)
Burer, S., Kılınç-Karzan, F.: How to convexify the intersection of a second order cone and a nonconvex quadratic. Math. Program. 162(1), 393–429 (2017)
Conn, A.R., Gould, N.I.M., Toint, P.L.: Trust Region Methods. MOS-SIAM Series on Optimization. SIAM, Philadelphia (2000)
Fallahi, S., Salahi, M., Terlaky, T.: Minimizing an indefinite quadratic function subject to a single indefinite quadratic constraint. Optimization 67(1), 55–65 (2018)
Feng, J.-M., Lin, G.-X., Sheu, R.-L., Xia, Y.: Duality and solutions for quadratic programming over single non-homogeneous quadratic constraint. J. Glob. Optim. 54(2), 275–293 (2012)
Fortin, C., Wolkowicz, H.: The trust region subproblem and semidefinite programming. Optim. Methods Softw. 19(1), 41–67 (2004)
Fradkov, A.L., Yakubovich, V.A.: The S-procedure and duality relations in nonconvex problems of quadratic programming. Vestn. LGU Ser. Mat. Mekh. Astron 6(1), 101–109 (1979)
Gould, N.I.M., Lucidi, S., Roma, M., Toint, P.L.: Solving the trust-region subproblem using the Lanczos method. SIAM J. Optim. 9(2), 504–525 (1999)
Guo, C.-H., Higham, N.J., Tisseur, F.: An improved arc algorithm for detecting definite Hermitian pairs. SIAM J. Matrix Anal. Appl. 31(3), 1131–1151 (2009)
Hazan, E., Koren, T.: A linear-time algorithm for trust region problems. Math. Program. 158(1), 363–381 (2016)
Ho-Nguyen, N., Kılınç-Karzan, F.: Online first-order framework for robust convex optimization. Oper. Res. 66(6), 1670–1692 (2018)
Ho-Nguyen, N., Kılınç-Karzan, F.: A second-order cone based approach for solving the trust region subproblem and its variants. SIAM J. Optim. 27(3), 1485–1512 (2017)
Jiang, R., Li, D.: Simultaneous diagonalization of matrices and its applications in quadratically constrained quadratic programming. SIAM J. Optim. 26(3), 1649–1668 (2016)
Jiang, R., Li, D.: Novel reformulations and efficient algorithms for the generalized trust region subproblem. SIAM J. Optim. 29(2), 1603–1633 (2019)
Jiang, R., Li, D.: A linear-time algorithm for generalized trust region subproblems. SIAM J. Optim. 30(1), 915–932 (2020)
Jiang, R., Li, D., Baiyi, W.: SOCP reformulation for the generalized trust region subproblem via a canonical form of two symmetric matrices. Math. Program. 169(2), 531–563 (2018)
Kılınç-Karzan, F., Yıldız, S.: Two-term disjunctions on the second-order cone. Math. Program. 154(1), 463–491 (2015)
Kuczynski, J., Wozniakowski, H.: Estimating the largest eigenvalue by the power and Lanczos algorithms with a random start. SIAM J. Matrix Anal. Appl. 13(4), 1094–1122 (1992)
Locatelli, M.: Some results for quadratic problems with one or two quadratic constraints. Oper. Res. Lett. 43(2), 126–131 (2015)
Megiddo, N.: Linear-time algorithms for linear programming in \({\mathbb{R}}^3\) and related problems. SIAM J. Comput. 12(4), 759–776 (1983)
Modaresi, S., Vielma, J.P.: Convex hull of two quadratic or a conic quadratic and a quadratic inequality. Math. Program. 164(1–2), 383–409 (2017)
Moré, J.J.: Generalizations of the trust region problem. Optim. Methods Softw. 2(3–4), 189–209 (1993)
Moré, J.J., Sorensen, D.C.: Computing a trust region step. SIAM J. Sci. Stat. Comput. 4(3), 553–572 (1983)
Nesterov, Y.: Lectures on Convex Optimization. Springer Optimization and Its Applications, 2nd edn. Springer, Basel (2018)
Nocedal, J., Wright, S.J.: Numerical Optimization. Springer Series in Operations Research and Financial Engineering. Springer, New York (2006)
Pólik, I., Terlaky, T.: A survey of the S-lemma. SIAM Rev. 49(3), 371–418 (2007)
Pong, T.K., Wolkowicz, H.: The generalized trust region subproblem. Comput. Optim. Appl. 58(2), 273–322 (2014)
Rendl, F., Wolkowicz, H.: A semidefinite framework for trust region subproblems with applications to large scale minimization. Math. Program. 77(2), 273–299 (1997)
Salahi, M., Taati, A.: An efficient algorithm for solving the generalized trust region subproblem. Comput. Appl. Math. 37(1), 395–413 (2018)
Stern, R.J., Wolkowicz, H.: Indefinite trust region subproblems and nonsymmetric eigenvalue perturbations. SIAM J. Optim. 5(2), 286–313 (1995)
Wang, A.L., Kılınç-Karzan, F.: On the tightness of SDP relaxations of QCQPs. Technical Report, ArXiV (2019). arXiv:1911.09195
Wang, J., Xia, Y.: A linear-time algorithm for the trust region subproblem based on hidden convexity. Optim. Lett. 11(8), 1639–1646 (2017)
Yang, B., Anstreicher, K., Burer, S.: Quadratic programs with hollows. Math. Program. 170(2), 541–553 (2018)
Yıldıran, U.: Convex hull of two quadratic constraints is an LMI set. IMA J. Math. Control Inf. 26(4), 417–450 (2009)
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This research is supported in part by National Science Foundation Grant CMMI 1454548.
Appendices
Proofs of Theorems 3 and 4
In this appendix, we outline how to modify the proofs of Theorems 1 and 2 to prove Theorems 3 and 4.
Theorem 3 Suppose there exists \(\gamma ^*\ge 0\) such that \(A(\gamma ^*)\succ 0\). Consider the closed nonempty interval \(\varGamma {:}{=}\left\{ \gamma \in {{\mathbb {R}}}_+:\, A(\gamma )\succeq 0\right\} \). Let \(\gamma _-\) denote its leftmost endpoint.
-
If \(\varGamma \) is bounded above, let \(\gamma _+\) denote its rightmost endpoint. Then,
$$\begin{aligned} {{\,\mathrm{conv}\,}}(\mathcal{S}) = \mathcal{S}(\gamma _-)\cap \mathcal{S}(\gamma _+). \end{aligned}$$In particular, we have \(\min _{x\in {{\mathbb {R}}}^n}\left\{ q_0(x):\, q_1(x)\le 0\right\} = \min _{x\in {{\mathbb {R}}}^n} \max \left\{ q(\gamma _-, x),\, q(\gamma _+,x)\right\} \).
-
If \(\varGamma \) is not bounded above, then \(q_1(x)\) is convex and
$$\begin{aligned} {{\,\mathrm{conv}\,}}(\mathcal{S}) = \mathcal{S}(\gamma _-)\cap \left\{ (x,t)\in {{\mathbb {R}}}^{n+1}:\, q_1(x)\le 0\right\} . \end{aligned}$$In particular, we have \(\min _{x\in {{\mathbb {R}}}^n}\left\{ q_0(x):\, q_1(x)\le 0\right\} = \min _{x\in {{\mathbb {R}}}^n} \left\{ q(\gamma _-,x):\, q_1(x)\le 0\right\} \).
Proof
The “\(\subseteq \)” inclusions follow from a trivial modification of Lemma 2. It suffices to prove the “\(\supseteq \)” inclusions. The case where \(A_0\) and \(A_1\) are both nonconvex is covered by Theorem 1. We consider the four remaining cases:
-
Suppose \(A_0\) and \(A_1\) are both convex. In this case, \(\varGamma = [0,\infty )\) and it suffices to show that \({{\,\mathrm{conv}\,}}(\mathcal{S}) = \left\{ (x,t):\, q_0(x)\le t,\, q_1(x) \le 0\right\} = \mathcal{S}\). This holds as \(\mathcal{S}\) is convex.
-
Suppose \(A_0\) is nonconvex and \(A_1\) is convex. In this case, \(\varGamma = [\gamma _-, \infty )\) is unbounded above. Furthermore, \(\gamma _-\) is positive and \(A(\gamma _-)\) has a zero eigenvalue. Suppose \(({\hat{x}},{\hat{t}})\) satisfies \(q(\gamma _-, {\hat{x}})\le {\hat{t}}\) and \(q_1({\hat{x}})\le 0\). If \(q_1({\hat{x}}) = 0\), then we also have \(q_0({\hat{x}}) = q(\gamma _-,{\hat{x}}) \le {\hat{t}}\), whence \(({\hat{x}},{\hat{t}})\in \mathcal{S}\). On the other hand, if \(q_1({\hat{x}}) < 0\), we may apply the argument in case (iii) in the proof of Lemma 3 verbatim (after replacing all occurrences of \(\gamma _+\) by \(\gamma ^*\)) to conclude that \(({\hat{x}},{\hat{t}})\in {{\,\mathrm{conv}\,}}(\mathcal{S})\).
-
Suppose \(A_0\) is convex and \(A_1\) is nonconvex. In this case, \(\varGamma = [0,\gamma _+]\) is bounded above and \(\gamma _-\) is defined to be \(\gamma _-=0\). Furthermore, \(A(\gamma _+)\) has a zero eigenvalue. Suppose \(({\hat{x}},{\hat{t}})\in \mathcal{S}(\gamma _-)\cap \mathcal{S}(\gamma _+)\). If \(q_1({\hat{x}}) \le 0\), then we also have \(q_0({\hat{x}}) = q(\gamma _-,{\hat{x}}) \le {\hat{t}}\), whence \(({\hat{x}},{\hat{t}})\in \mathcal{S}\). On the other hand, if \(q_1({\hat{x}}) > 0\), we may apply the argument in case (ii) in the proof of Lemma 3 verbatim to conclude that \(({\hat{x}},{\hat{t}})\in {{\,\mathrm{conv}\,}}(\mathcal{S})\).
\(\square \)
We will prove Theorem 4 using a limiting argument and reducing it to Theorem 3. The proof follows that of Lemma 5 almost verbatim.
Theorem 4 Suppose there exists \(\gamma ^*\ge 0\) such that \(A(\gamma ^*)\succeq 0\). Consider the closed nonempty interval \(\varGamma {:}{=}\left\{ \gamma \in {{\mathbb {R}}}_+:\, A(\gamma )\succeq 0\right\} \). Let \(\gamma _-\) denote its leftmost endpoint.
-
If \(\varGamma \) is bounded above, let \(\gamma _+\) denote its rightmost endpoint. Then,
$$\begin{aligned} \overline{{{\,\mathrm{conv}\,}}}(\mathcal{S}) = \mathcal{S}(\gamma _-)\cap \mathcal{S}(\gamma _+). \end{aligned}$$In particular, \(\inf _{x\in {{\mathbb {R}}}^n}\left\{ q_0(x):\, q_1(x)\le 0\right\} = \inf _{x\in {{\mathbb {R}}}^n} \max \left\{ q(\gamma _-, x),\, q(\gamma _+,x)\right\} \).
-
If \(\varGamma \) is not bounded above, then \(q_1(x)\) is convex and
$$\begin{aligned} \overline{{{\,\mathrm{conv}\,}}}(\mathcal{S}) = \mathcal{S}(\gamma _-)\cap \left\{ (x,t)\in {{\mathbb {R}}}^{n+1}:\, q_1(x)\le 0\right\} . \end{aligned}$$In particular, \(\inf _{x\in {{\mathbb {R}}}^n}\left\{ q_0(x):\, q_1(x)\le 0\right\} = \inf _{x\in {{\mathbb {R}}}^n} \left\{ q(\gamma _-,x):\, q_1(x)\le 0\right\} \).
Proof
The “\(\subseteq \)” inclusions follow from a trivial modification of Lemma 4. It suffices to prove the “\(\supseteq \)” inclusions.
Denote the set on the right hand side by \(\mathcal{R}\), i.e., \(\mathcal{R}{:}{=}\mathcal{S}(\gamma _-) \cap \mathcal{S}(\gamma _+)\) when \(\varGamma \) is bounded and \(\mathcal{R}{:}{=}\mathcal{S}(\gamma _-)\cap \left\{ (x,t):\, q_1(x)\le 0\right\} \) when \(\varGamma \) is unbounded.
Let \(({\hat{x}},{\hat{t}})\in \mathcal{R}\). It suffices to show that \(({\hat{x}},\hat{t}+\epsilon )\in {{\,\mathrm{conv}\,}}(\mathcal{S})\) for all \(\epsilon >0\).
We will perturb \(A_0\) slightly to create a new instance of the problem. Let \(\delta >0\) to be picked later. Define \(A_0' = A_0 +\delta I_n\) and let all remaining data be unchanged, i.e.,
We will denote all quantities related to the perturbed system with an apostrophe.
We claim it suffices to show that there exists \(\delta >0\) small enough such that \(({\hat{x}},{\hat{t}}+\epsilon )\in \mathcal{R}'\). Indeed, suppose this is the case. Note that for any \(x\in {{\mathbb {R}}}^n\), we have \(q_1(x) = q_1'(x)\) and \(q_0(x) \le q_0'(x)\). Hence, \({{\,\mathrm{conv}\,}}(\mathcal{S}')\subseteq {{\,\mathrm{conv}\,}}(\mathcal{S})\). Then, noting that \(A'(\gamma ^*) = A(\gamma ^*) + \delta I_n\succ 0\), we may apply Theorem 3 to the perturbed system to get \(({\hat{x}},{\hat{t}}+\epsilon )\in \mathcal{R}' = {{\,\mathrm{conv}\,}}(\mathcal{S}')\subseteq {{\,\mathrm{conv}\,}}(\mathcal{S})\) as desired.
First note that \(A_1 = A_1'\) so that \(\varGamma \) is bounded if and only if \(\varGamma '\) is bounded. We will then pick \(\delta >0\) small enough such that
where the last condition is only required when \(\gamma _+\) and \(\gamma _+'\) both exist. This is possible as the expression on the left of each inequality is continuous in \(\delta \) and is strictly satisfied when \(\delta =0\).
The following computation shows that \(q'(\gamma _-', {\hat{x}}) \le {\hat{t}} + \epsilon \).
The first inequality follows by noting \(q'(\gamma ,x) = q(\gamma ,x) + \delta \left\Vert x \right\Vert ^2\), the second inequality follows from our assumptions on \(\delta \), and the third inequality follows from the assumption that \(({\hat{x}}, {\hat{t}})\in \mathcal{S}(\gamma _-)\). Thus \((\hat{x},{\hat{t}}+\epsilon )\in \mathcal{S}'(\gamma _-')\). When \(\varGamma \) is bounded (or equivalently, when \(\gamma _+'\) and \(\gamma _+\) exist), a similar calculation shows that \(q'(\gamma _+',{\hat{x}}) - ({\hat{t}} + \epsilon )\le 0\) so that \(({\hat{x}},\hat{t}+\epsilon )\in \mathcal{S}'(\gamma '_+)\). Finally, when \(\varGamma \) is unbounded we have \(q_1'({\hat{x}}) = q_1({\hat{x}})\le 0\) so that \(({\hat{x}},\hat{t}+\epsilon )\in \left\{ (x,t):\, q_1'(x)\le 0\right\} \). Thus, \(({\hat{x}},\hat{t}+\epsilon )\) is in \(\mathcal{R}'\), concluding the proof. \(\square \)
Estimation of the regularity parameters
In Sect. 4 we gave algorithms to solve the GTRS assuming that we had access to \((\xi ,\zeta )\) and \({\hat{\gamma }}\) satisfying Assumption 5. In this appendix, we show how to compute these quantities.
Let \(q_0,q_1\) satisfy Assumption 4. Recall the definitions
We will find \((\xi ,\zeta )\) satisfying
and a \({\hat{\gamma }}\) such that \(\lambda _{\min }(A({\hat{\gamma }}))\ge \xi \).
We will accomplish this in two stages. We begin by estimating \(\xi ^*\) using only an upper bound \({\bar{\zeta }}\) of \(\zeta ^*\). Then using our estimate \(\xi \) we will compute \(\zeta \).
1.1 Computing \(\xi \) and \({\hat{\gamma }}\)
We start with the following guarantee for the algorithm TestXi (Algorithm 5).
Lemma 17
Given \(q_0,q_1\) satisfying Assumption 4, an arbitrary \(0<\xi \le 1\), an upper bound \({\bar{\zeta }}\ge \zeta ^*\), and a failure probability \(p_{\xi }>0\), TestXi (Algorithm 5) will output
with probability \(1-p_{\xi }\). This algorithm runs in time
Proof
We condition on the event that ApproxEig succeeds every time it is called. By the union bound, this happens with probability at least \(1-p_\xi \).
As we have conditioned on ApproxEig succeeding, any \({\hat{\gamma }}\) which is output by TestXi will satisfy
It is clear that TestXi will output “Fail” if \(\xi >2\xi ^*\) as there does not exist any \({\hat{\gamma }}\) such that \(\lambda _{\min }(A({\hat{\gamma }}))\ge \xi ^*\). It remains to show that, given \(\xi \le \xi ^*\), TestXi will output some \({\hat{\gamma }}\).
For the sake of contradiction, assume that the algorithm fails to output in each of the T rounds. Let \(P {:}{=}\left\{ \gamma :\, \lambda _{\min }(A(\gamma ))\ge 3\xi ^*/4\right\} \). Recall that \(\lambda _{\min }(A(\gamma ))\) is 1-Lipschitz in \(\gamma \). As there exists some \(\gamma \) such that \(\lambda _{\min }(A(\gamma ))\ge \xi ^*\) (see Definition 1), we conclude that P is an interval of length at least \(\xi ^*/2\).
Note that \(P\subseteq [s_0,t_0]\). We will inductively show that \(P\subseteq [s_k,t_k]\) for each \(k\in \left\{ 1,\dots ,T\right\} \). Let \(k\in \left\{ 0,\dots ,T-1\right\} \) and let \(s_k,{\bar{\gamma }},t_k\) be defined as in the algorithm and let x be the unit vector found in step 3.(d). We claim that \(x^\top A_1x\ne 0\). Indeed suppose \(x^\top A_1x = 0\), then \(x^\top A(\gamma )x = x^\top A({\bar{\gamma }})x \le 3\xi /4\) for all \(\gamma \). This contradicts the assumption that there exists some \(\gamma \) such that \(\lambda _{\min }(A(\gamma ))\ge \xi \). Now suppose \(\gamma \in P\), then
where the first inequality follows from \(\gamma \in P\), and the last one from the fact that the algorithm did not output in iteration k (and thus the if statement in step 3.(d) did not hold). Thus, if \(x^\top A_1x > 0\), then we have the implication \(\gamma \in P \implies \gamma \ge {\bar{\gamma }}\). Similarly, if \(x^\top A_1x<0\), then we have the implication \(\gamma \in P\implies \gamma \le {\bar{\gamma }}\). Then by induction, we have \(P\subseteq [s_{k+1},t_{k+1}]\).
We conclude that P, an interval of length at least \(\xi ^*/2\), is contained in \([s_T,t_T]\) an interval of length
Noting that \(\xi \le \xi ^*\) gives us the desired contradiction.
The running time of this algorithm follows from Lemma 11. \(\square \)
Given a lower bound \(\xi \le \xi ^*\), Lemma 17 guarantees that TestXi will find a \({\hat{\gamma }}\) satisfying \(\lambda _{\min }(A({\hat{\gamma }}))\ge \xi /2\) with high probability. In order to make use of this lemma without a lower bound on \(\xi ^*\), we will simply repeatedly call TestXi with decreasing guesses for \(\xi \). Consider Algorithm 6.
Theorem 8
Given \(q_0,q_1\) satisfying Assumption 4, an upper bound \({\bar{\zeta }}\ge \zeta ^*\), and a failure probability \(p>0\), ApproxXi (Algorithm 6) will output \(\xi \) and \({\hat{\gamma }}\) such that
and run in time
with probability \(1-p\).
Proof
We condition on the event that TestXi succeeds every time it is called. By the union bound, this happens with probability at least \(1-p\).
Let \(k^*\in \left\{ 1,2,\dots \right\} \) be such that \(\xi ^*/2\le 2^{-k^*}<\xi ^*\). Then, as we have conditioned on TestXi succeeding, Lemma 17 guarantees that TestXi\((q_0, q_1, 2^{-k},{\bar{\zeta }}, 2^{-(k+1)}p)\) outputs
Thus, TestXi will output “Fail” for every \(k<k^*\) and will output \({\hat{\gamma }}\) either on round \(k^*\) or \(k^*+1\). We can then bound
We bound the run time of the algorithm as follows.
\(\square \)
1.2 Computing \(\zeta \)
Recall the guarantee of the algorithm ApproxGammaPlus.
Lemma 12 Given \(q_0\), \(q_1\) satisfying Assumption 4, \((\xi ,\zeta )\) and \(\hat{\gamma }\) satisfying Assumption 5, \(\delta >0\), and \(p_{\tilde{\gamma }_+}\), ApproxGammaPlus (Algorithm 2) outputs \(\tilde{\gamma }_+\) satisfying
with probability \(1-p_{\tilde{\gamma }_+}\). This algorithm runs in time
We will repeatedly call ApproxGammaPlus with different choices of \(\delta \). Consider the algorithm ApproxZeta.
Theorem 9
Given \(q_0,q_1\) satisfying Assumption 4, \((\xi ,{\bar{\zeta }})\) and \({\hat{\gamma }}\) satisfying Assumption 5, and failure probability \(p>0\), ApproxZeta (Algorithm 7) will output \(\zeta \) such that
and run in time
with probability \(1-p\).
Proof
We condition on the event that ApproxGammaPlus succeeds every time it is called. By the union bound, this happens with probability at least \(1-p\).
We first check that the assumptions of Lemma 12 hold. For \(k = 1\), we have \(2^{-(k-1)}{\bar{\zeta }} = {\bar{\zeta }}\ge \zeta ^*\). Then by induction, and conditioning on ApproxGammaPlus succeeding, Lemma 12 guarantees
If ApproxZeta fails to terminate in round k, then 1.(b) ensures \({\hat{\zeta _k}}\le 2^{-(k+1)}{\bar{\zeta }}\). This in turn implies that \(\zeta ^* \le 2^{-((k+1)-1)}{\bar{\zeta }}\) and, by induction, the assumptions of Lemma 12 hold in every round that ApproxGammaPlus is called.
Let k be the round in which the algorithm terminates. If \(k = 1\), then the guarantee of Lemma 12 implies \(\zeta ^*\ge {\hat{\zeta }}_{1}\), whence
Thus, we may assume \(k\ge 2\). The condition of step 1.(b) then guarantees the two inequalities
Then, we have
where the first and fifth relations follow from Lemma 12 and the second and fourth relations follow from (14) above.
It remains to bound the run time of ApproxZeta. Let \(k^*\in \left\{ 1,2,\dots \right\} \) be such that \(\zeta ^* \le 2^{-(k^*-1)}{\bar{\zeta }} < 2\zeta ^*\). We show that ApproxZeta terminates within \(k^*\) rounds. Suppose ApproxZeta reaches the \(k^*\)th round. Then, we have
where we used Lemma 12 in the first relation, and the definition of \(k^*\) in the second relation. Therefore, ApproxZeta terminates in round \(k^*\) at the latest and we can bound the run time of this algorithm as
\(\square \)
Rights and permissions
About this article
Cite this article
Wang, A.L., Kılınç-Karzan, F. The generalized trust region subproblem: solution complexity and convex hull results. Math. Program. 191, 445–486 (2022). https://doi.org/10.1007/s10107-020-01560-8
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10107-020-01560-8