Abstract
In a first contribution, we revisit two certificates of positivity on (possibly non-compact) basic semialgebraic sets due to Putinar and Vasilescu (C R Acad Sci Ser I Math 328(6):495–499, 1999). We use Jacobi’s technique from (Math Z 237(2):259–273, 2001) to provide an alternative proof with an effective degree bound on the sums of squares weights in such certificates. As a consequence, it allows one to define a hierarchy of semidefinite relaxations for a general polynomial optimization problem. Convergence of this hierarchy to a neighborhood of the optimal value as well as strong duality and analysis are guaranteed. In a second contribution, we introduce a new numerical method for solving systems of polynomial inequalities and equalities with possibly uncountably many solutions. As a bonus, one can apply this method to obtain approximate global optimizers in polynomial optimization.
Similar content being viewed by others
References
Artin, E.: Über die zerlegung definiter funktionen in quadrate. Abhandlungen aus dem mathematischen Seminar der Universität Hamburg 5(1), 100–115 (1927)
Bajbar, T., Stein, O.: Coercive polynomials and their Newton polytopes. SIAM J. Optim. 25(3), 1542–1570 (2015)
Berg, C.: The multidimensional moment problem and semigroups. Proc. Symp. Appl. Math. 37, 110–124 (1987)
Blekherman, G.: There are significantly more nonegative polynomials than sums of squares. Isr. J. Math. 153(1), 355–380 (2006)
Blekherman, G., Gouveia, J., Pfeiffer, J.: Sums of squares on the hypercube. Math. Z. 284(1–2), 41–54 (2016)
Blekherman, G., Smith, G.G., Velasco, M.: Sharp degree bounds for sum-of-squares certificates on projective curves. J. Math. Pures Appl. 129, 61–86 (2019)
Boyd, S., Vandenberghe, L.: Convex Optimization. Cambridge University Press, Cambridge (2004)
Curto, R.E., Fialkow, L.A.: Truncated K-moment problems in several variables. J. Oper. Theory 54, 189–226 (2005)
Delzell, C.N.: Continous, piecewise-polynomial functions which solve Hilbert’s 17th problem. J. Reine Angew. Math. 1993(440), 157–174 (1993)
Demmel, J., Nie, J., Powers, V.: Representations of positive polynomials on noncompact semialgebraic sets via KKT ideals. J. Pure Appl. Algebra 209(1), 189–200 (2007)
Dickinson, P.J., Povh, J.: On an extension of Pólya’s Positivstellensatz. J. Global Optim. 61(4), 615–625 (2015)
Djukić, D., Janković, V., Matić, I., Petrović, N.: The IMO Compendium: A Collection of Problems Suggested for the International Mathematical Olympiads: 1959–2009, 2nd edn. Springer, Berlin (2011)
Gershwin, S.B.: KKT examples. min J. 10, 2 (2010)
Greuet, A., Safey El Din, M.: Probabilistic algorithm for polynomial optimization over a real algebraic set. SIAM J. Optim. 24(3), 1313–1343 (2014)
Henrion, D., Lasserre, J.B.: Detecting global optimality and extracting solutions in GloptiPoly. In: Henrion, D., Garulli, A. (eds.) Positive Polynomials in Control, pp. 293–310. Springer, Berlin (2005)
Hilbert, D.: Über die Darstellung Definiter Formen als Summe von Formenquadraten. Math. Ann. 32(3), 342–350 (1888)
Hu, S., Li, G., Qi, L.: A tensor analogy of Yuan’s theorem of the alternative and polynomial optimization with sign structure. J. Optim. Theory Appl. 168(2), 446–474 (2016)
Jacobi, T.: A representation theorem for certain partially ordered commutative rings. Math. Z. 237(2), 259–273 (2001)
Jeyakumar, V., Kim, S., Lee, G.M., Li, G.: Solving global optimization problems with sparse polynomials and unbounded semialgebraic feasible sets. J. Glob. Optim 65(2), 175–190 (2016)
Jeyakumar, V., Lasserre, J.B., Li, G.: On polynomial optimization over non-compact semi-algebraic sets. J. Optim. Theory Appl. 163(3), 707–718 (2014)
Jibetean, D., Laurent, M.: Semidefinite approximations for global unconstrained polynomial optimization. SIAM J. Optim. 16(2), 490–514 (2005)
Josz, C., Henrion, D.: Strong duality in Lasserre’s hierarchy for polynomial optimization. Optim. Lett. 10(1), 3–10 (2016)
Krivine, J.L.: Anneaux préordonnés. J. Anal. Mathé. 12(1), 307–326 (1964)
Lasserre, J.B.: Global optimization with polynomials and the problem of moments. SIAM J. Optim. 11(3), 796–817 (2001)
Lasserre, J.B.: A sum of squares approximation of nonnegative polynomials. SIAM Rev. 49(4), 651–669 (2007)
Lasserre, J.B.: An Introduction to Polynomial and Semi-algebraic Optimization, vol. 52. Cambridge University Press, Cambridge (2015)
Lasserre, J.B., Laurent, M., Rostalski, P.: Semidefinite characterization and computation of zero-dimensional real radical ideals. Found. Comput. Math. 8(5), 607–647 (2008)
Lasserre, J.B., Magron, V.: In SDP relaxations, inaccurate solvers do robust optimization. SIAM J. Optim. 29(3), 2128–2145 (2019)
Lasserre, J.B., Netzer, T.: SOS approximations of nonnegative polynomials via simple high degree perturbations. Math. Z. 256(1), 99–112 (2007)
Laurent, M.: Revisiting two theorems of Curto and Fialkow on moment matrices. Proc. Am. Math. Soc. 133(10), 2965–2976 (2005)
Laurent, M.: Sums of squares, moment matrices and optimization over polynomials. In: Putinar, M., Sullivant, S. (eds.) Emerging Applications of Algebraic Geometry, pp. 157–270. Springer, Berlin (2009)
Leep, D., Starr, C.: Polynomials in R[X, Y] that are sums of squares in R(X, Y). Proc Am Math Soc 129(11), 3133–3141 (2001)
Lombardi, H., Perrucci, D., Roy, M.F.: An elementary recursive bound for effective Positivstellensatz and Hilbert 17-th problem. arXiv preprint arXiv:1404.2338 (2014)
Magron, V., Safey El Din, M.: On exact polya and Putinar’s representations. In: Proceedings of the 2018 ACM International Symposium on Symbolic and Algebraic Computation, pp. 279–286 (2018)
Mai, N.H.A., Magron, V., Lasserre, J.B.: A sparse version of Reznick’s Positivstellensatz. arXiv preprint arXiv:2002.05101 (2020)
Marshall, M.: Approximating positive polynomials using sums of squares. Can. Math. Bull. 46(3), 400–418 (2003)
Navascués, M., García-Sáez, A., Acín, A., Pironio, S., Plenio, M.B.: A paradox in bosonic energy computations via semidefinite programming relaxations. New J. Phys. 15(2), 023026 (2013)
Nie, J.: Discriminants and nonnegative polynomials. J. Symb. Comput. 47(2), 167–191 (2012)
Nie, J.: Optimality conditions and finite convergence of Lasserre’s hierarchy. Math. Program. 146(1–2), 97–121 (2014)
Nie, J.: Tight relaxations for polynomial optimization and Lagrange multiplier expressions. Math. Program. 178(1–2), 1–37 (2019)
Nie, J., Demmel, J., Sturmfels, B.: Minimizing polynomials via sum of squares over the gradient ideal. Math. Program. 106(3), 587–606 (2006)
Nie, J., Schweighofer, M.: On the complexity of Putinar’s positivstellensatz. J. Complex. 23(1), 135–150 (2007)
Parrilo, P.A.: Structured semidefinite programs and semialgebraic geometry methods in robustness and optimization. Ph.D. thesis, California Institute of Technology (2000)
Pólya, G.: Über Positive Darstellung von Polynomen. Vierteljschr. Naturforsch. Ges. Zürich 73, 141–145 (1928)
Putinar, M.: Positive polynomials on compact semi-algebraic sets. Indiana Univ. Math. J. 42(3), 969–984 (1993)
Putinar, M., Vasilescu, F.H.: Positive polynomials on semi-algebraic sets. C. R. Acad. Sci. Ser. I Math. 328(7), 585–589 (1999)
Putinar, M., Vasilescu, F.H.: Solving moment problems by dimensional extension. C. R. Acad. Sci. Ser. I Math. 328(6), 495–499 (1999)
Quijorna, M.L.: An experimental approach for global polynomial optimization based on moments and semidefinite programming. arXiv preprint arXiv:1809.09043 (2018)
Reznick, B.: Uniform denominators in Hilbert’s seventeenth problem. Math. Z. 220(1), 75–97 (1995)
Reznick, B.: Some concrete aspects of Hilbert’s 17th problem. Contemp. Math. 253, 251–272 (2000)
Reznick, B.: On the absence of uniform denominators in Hilbert’s 17th problem. Proc. Am. Math. Soc. 133(10), 2829–2834 (2005)
Schabert, R.: Uniform denominators in Hilbert’s 17th problem theorems by Pólya and Reznick. Manuscript (2019). http://www.math.uni-konstanz.de/~infusino/FachseminarMA-SS19/Schabert-Handout.pdf
Scheiderer, C.: A Positivstellensatz for projective real varieties. Manuscr. Math. 138(1–2), 73–88 (2012)
Schmüdgen, K.: The K-moment problem for compact semi-algebraic sets. Math. Ann. 289(1), 203–206 (1991)
Schweighofer, M.: Iterated rings of bounded elements and generalizations of Schmüdgen’s Positivstellensatz. J. Reine Angew. Math. 554, 19–45 (2003)
Schweighofer, M.: Global optimization of polynomials using gradient tentacles and sums of squares. SIAM J. Optim. 17(3), 920–942 (2006)
Slinko, A.M.: USSR Mathematical Olympiads, 1989–1992, vol. 11. Australian Mathematics Trust, Bruce (1997)
Stengle, G.: A Nullstellensatz and a Positivstellensatz in semialgebraic geometry. Math. Ann. 207(2), 87–97 (1974)
Trnovska, M.: Strong duality conditions in semidefinite programming. J. Electr. Eng. 56(12), 1–5 (2005)
Vui, H.H., Son, P.T.: Representations of positive polynomials and optimization on noncompact semialgebraic sets. SIAM J. Optim. 20(6), 3082–3103 (2010)
Waki, H., Kim, S., Kojima, M., Muramatsu, M., Sugimoto, H.: Algorithm 883: SparsePOP—a sparse semidefinite programming relaxation of polynomial optimization problems. ACM Trans. Math. Softw. 35(2), 15 (2008)
Acknowledgements
The second author was supported by the Tremplin ERC Stg Grant ANR-18-ERC2-0004-01 (T-COPS project) and by the FMJH Program PGMO (EPICS project) (Grant number P-2019-0038) and EDF, Thales, Orange et Criteo. This work has benefited from the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie Actions, grant agreement 813211 (POEMA) as well as from the AI Interdisciplinary Institute ANITI funding, through the French “Investing for the Future PIA3” program under the Grant agreement \(\hbox {n}^{\circ }\)ANR-19-PI3A-0004.
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.
Appendix
Appendix
1.1 Proof of Theorem 5
Jacobi [18, Theorem 7] proves another result of Putinar and Vasilescu [47, Theorem 4.2], which states that if \(f,g_1,\dots ,g_m\) are polynomials of even degree and \(\tilde{f}>0\) on \(S(\{\tilde{g}_1,\dots ,\tilde{g}_m\})\backslash \{0\}\), then \(\theta ^kf\in Q(g)\) for k large enough, where \(\tilde{h}\) is the homogenization of given polynomial h. The idea of Jacobi is to apply Putinar’s Positivstellensatz for the fact that \(\tilde{f}>0\) on the intersection of \(S(\{\tilde{g}_1,\dots ,\tilde{g}_m\})\) with the unit sphere. Our proof for Theorem 5 replaces \(\tilde{f}\) here by the perturbation of \(\tilde{f}\) and replaces the unit sphere by a sphere with an arbitrary radius \(L^{1/2}\). This will have a direct implication on the complexity analysis derived in Proposition 1.
Proof
1. Let us prove the conclusion under condition (i). For every \(h \in {\mathbb {R}}[ x ]\), we define by \(\hat{h}\) the degree \(2d_1 ( h )\) homogenization of h, i.e.,
where \({d_1}( h ): = {1 + \lfloor {{{\deg ( h )}}/{2}} \rfloor }\). Let \(\varepsilon > 0\) be fixed. Let \(\tilde{f}\) be the degree 2d homogenization of f. Then \(\tilde{f} + \varepsilon {\Vert {( {x,{x_{n + 1}}} )} \Vert _2^{2d}}\) is a homogeneous polynomial of degree 2d. Let \(\hat{g} := \{ {{\hat{g}_1},\dots ,{\hat{g}_m}} \}\) and \(L>0\). We will show that
Let \(( {y,{y_{n + 1}}}) \in S( \hat{g}, \{L-\Vert ( x,x_{n + 1} \Vert _2^2\})\) be fixed. By (1.34), one has
We consider the following two cases:
-
Case 1: \({y_{n + 1}} \ne 0\). For \(j=1,\dots ,m\), by (1.36) and since \(y_{n + 1}^{2d_1 \left( g_j \right) }>0\), \({g_j}( {\displaystyle {y}/{{{y_{n + 1}}}}} ) \ge 0\). It implies that \(\displaystyle {y}/{{{y_{n + 1}}}} \in S\left( g\right) \). By assumption, \(f( \displaystyle {y}/{{{y_{n + 1}}}}) \ge 0\). From this, \(\tilde{f}( {y,{y_{n + 1}}}) = y_{n + 1}^{2d}f( {{y}/{{{y_{n + 1}}}}} ) \ge 0\). From this and since \(\Vert ( {y,{y_{n + 1}}} )\Vert _2^2=L\), \(\tilde{f}( {y,{y_{n + 1}}}) + \varepsilon {\Vert {( {y,{y_{n + 1}}} )} \Vert _2^{2d}} \ge \varepsilon L^{d}\).
-
Case 2: \({y_{n + 1}} = 0\). By definition of \(d_1(f)\), \(x_{n+1}\) divides \(\tilde{f}\) so \(\tilde{f}( {y,{y_{n + 1}}}) = 0\). From this and since \(\Vert ( {y,{y_{n + 1}}})\Vert _2^2=L\),
$$\begin{aligned} \tilde{f}\left( {y,{y_{n + 1}}}\right) + \varepsilon {\Vert {\left( {y,{y_{n + 1}}} \right) } \Vert _2^{2d}} = \varepsilon L^{d}. \end{aligned}$$
Thus, (1.35) holds. It is not hard to show that \(S( \hat{g} \cup \{L-\Vert {( {x,{x_{n + 1}}} )}\Vert _2^2\})\) satisfies the Archimedean condition. From this and by applying Theorem 2 (i),
Then there exist \(\psi _{j} \in \varSigma [ {x,{x_{n + 1}}} ],\ j = 0,1,\dots ,m\) and \(\varphi \in {\mathbb {R}}[ {x,{x_{n + 1}}} ]\) such that
Let \(( {z,{z_{n + 1}}}) \in {{\mathbb {R}}^{n + 1}}\backslash \{ 0\}\). By replacing \({( {x,{x_{n + 1}}} )}\) in the last equality by \(L^{1/2}\frac{{( {z,{z_{n + 1}}} )}}{{\Vert {( {z,{z_{n + 1}}} )} \Vert _2}}\) and the fact that \({\tilde{f} + \varepsilon {\Vert {( {x,{x_{n + 1}}})} \Vert _2^{2d}}}, \ \hat{g}_1,\dots ,\hat{g}_m\) are homogeneous polynomials of degree \(2d,\ 2d_1 \left( g_1 \right) ,\dots ,2d_1 \left( g_m\right) \) respectively, one has
Set
Obviously, K is even. After multiplying the two sides of the last equality with \(\Vert ( z,z_{n + 1}) \Vert ^K\), one has
where
Since \(( {z,{z_{n + 1}}}) \in {{\mathbb {R}}^{n + 1}}\backslash \{ 0\}\) is arbitrary,
Let \(j\in \{0,\dots ,m\}\) be fixed and set \(g_0:=1\) and \(d_1(g_0):=0\). We will show that
for some \(s_j\in \varSigma [x,x_{n+1}]_{K/2-d_1(g_j)}\) and \(\xi _j\in {\mathbb {R}}[x,x_{n+1}]\). By definition of K (\(K- 2d_1 ( g_j )\ge \deg (\psi _j)\), \(j=1,\dots ,m\)) and since \(\psi _j\) is SOS, \(\bar{\psi }_j=\sum \nolimits _{t = 1}^{l} {\phi _{t}^2}\) with \(\phi _{t}:=h_t+p_t\Vert (x,x_{n+1})\Vert _2\) for some \(h_t\in {\mathbb {R}}[x,x_n]_{K/2-d_1(g_j)}\) and \(p_t\in {\mathbb {R}}[x,x_n]_{K/2-d_1(g_j)-1}\). Then
By setting \(s_j:=\sum \nolimits _{t = 1}^{l} {(h_t^2+p_t^2\Vert (x,x_{n+1})\Vert _2^2)}\) and \(\xi _j:=2\sum \nolimits _{t = 1}^{l} {h_tp_t}\), (1.38) follows. By (1.37) and (1.38),
Note that \(\Vert {( {x,{x_{n + 1}}} )} \Vert _2\) is not a polynomial. From the last equality and since the right hand side is polynomial, the left hand side must be a polynomial, so \(\sum \nolimits _{j = 1}^m {L^{d_1(g_j)}\xi _j \hat{g}_j}=0\). Thus,
By setting \(k_\varepsilon :=K/2-d\), \({s_0} \in \varSigma {[ x,x_{n+1} ]_{{k_\varepsilon } + d}}\), \({s _j} \in \varSigma {[ x,x_{n+1} ]_{{k_\varepsilon } + d - d_1( {{g_j}})}},\,j = 1,\dots ,m\) and \(L^d{\Vert {( {x,{x_{n + 1}}})} \Vert _2^{2{k_\varepsilon }}}( {\tilde{f} + \varepsilon {{\Vert {( {x,{x_{n + 1}}} )} \Vert }_2^{2d}}} ) = {s _0} + \sum \nolimits _{j = 1}^m {L^{d_1(g_j)}{s_j}{{\hat{g}}_j}} \). By replacing \(x_{n+1}\) by 1, \( L^d{\theta ^{{k_\varepsilon }}}( {f + \varepsilon {\theta ^{d}}} ) = {s _0}( {x,1} ) + \sum \nolimits _{j = 1}^m {L^{d_1(g_j)}{s_j}( {x,1} ){g_j}} \). Since \({s_0} \in \varSigma {[ x,x_{n+1}]_{{k_\varepsilon } + d}}\), \({s_j} \in \varSigma {[ x,x_{n+1} ]_{{k_\varepsilon } + d - d_1( {{g_j}} )}},\,j = 1,\dots ,m\), \({s _0}( {x,1} ) \in \varSigma {[ x ]_{{k_\varepsilon } + d}}\), \({s _j}( {x,1}) \in \varSigma {[ x]_{{k_\varepsilon } + d - d_1( {{g_j}})}},\,j = 1,\dots ,m\). Note that \(d_1(g_j)\ge d_2(g_j)=u_j\), \(j=1,\dots ,m\). Hence, \({\theta ^{{k_\varepsilon }}}( {f + \varepsilon {\theta ^{d}}} ) \in {Q_{{k_\varepsilon } + d}}( g )\) by the definition of truncated quadratic module.
2. Let us show the conclusion under condition (ii). We do a similar process as part 1 (under condition (i)). The difference is that \(\hat{h}\) here is defined as the degree \(2d_2 ( h )\) (instead of \(2d_1 ( h )\)) homogenization of \(h \in {\mathbb {R}}[ x ]\) and the proof for (1.35). To show (1.35) in Case 2: \(y_{n+1}=0\) here, we rely on the constraint \(g_m=f+\lambda \) for some \(\lambda \ge 0\) (instead of \(d \ge d_1(f)\) and that \(x_{n+1}\) divides \(\tilde{f}\)). More explicitly, we assume by contradiction that
From this and since \(0 \le {\hat{g}_m}(y,{y_{n + 1}}) = \hat{f}(y,0)\),
It follows that \(\Vert ({y,{y_{n + 1}}})\Vert _2^2<L=\Vert ({y,{y_{n + 1}}})\Vert _2^2\). It is impossible. Thus, \(\tilde{f}({y,{y_{n + 1}}}) + \varepsilon {\Vert {( {y,{y_{n + 1}}} )}\Vert _2^{2d}} \ge \varepsilon L^{d}\). \(\square \)
1.2 Proof of Proposition 1
Proof
We keep all the notation from the proof of Theorem 5. Without loss of generality, let us assume that \(L=1/4\). By (1.35), \(S( \hat{g},{\{ { L - {{\Vert {( {x,{x_{n + 1}}} )} \Vert }^2_2}} \}} )\subset (-1,1)^{n+1}\) and
By definition of \(\hat{g}\) and since \(0_{{\mathbb {R}}^n}\in S(g)\), \((0_{{\mathbb {R}}^n},L^{1/2})\in S( \hat{g},{\{ { L - {{\Vert {( {x,{x_{n + 1}}} )} \Vert }^2_2}} \}} )\). Thus, \(S( \hat{g},{\{ { L - {{\Vert {( {x,{x_{n + 1}}} )} \Vert }^2_2}} \}} )\) is nonempty. By definition of K,
Since \(\Vert {(x,{x_{n + 1}})} \Vert _2^{2d} = \sum \limits _{\bar{\alpha }\in {\mathbb {N}}_d^{n + 1}} {c_{n+1}(\bar{\alpha })(x,{x_{n + 1}})^{2\bar{\alpha }}} \), one has
Note that \(\tilde{f} = \sum \nolimits _\alpha {f_\alpha x^\alpha x_{n+1}^{2d-|\alpha |}}\), so \(\Vert \tilde{f} \Vert _{\max }=\Vert f \Vert _{\max ,d}\). Thus,
From these and using Theorem 3, one can choose
for some \(C>0\). The right hand side of this inequality comes from (1.39), (1.40) and the fact that the function \(t \mapsto c\exp \left( {b{t^c}} \right) \) with positive constants b and c is increasing on \([0,\infty )\). By setting \(k_\varepsilon =K/2-d\), the conclusion follows. \(\square \)
1.3 Proof of Proposition 3
Proof
We denote by P and \(P^\star \) feasible set and optimal solutions set for the moment SDP (4.22), respectively. We claim that P is nonempty. Indeed, with z being the moment sequence of the 1-atomic measure \(\theta (\bar{x})^{-1}\delta _{\bar{x}}\) for some \(\bar{x}\in S(g)\), it is not hard to check that the truncation of z is a feasible solution of (4.22). By setting \(C: = {L_z}({\theta ^k}(f + \varepsilon {\theta ^d}))\), \(P^\star \) is also the set of optimal solutions for
with \(g_0=1\). Denote by \(\bar{P}\) the feasible set of (1.41). Note that \(\bar{P}\) is nonempty since the truncation of z is also a feasible solution of (1.41). We will prove that \(\bar{P}\) is bounded. By definition of \(\theta \), \({\theta ^r} = \sum \nolimits _{\alpha \in {\mathbb {N}}_r^n} {{C_\alpha ^r }{x^{2\alpha }}}\) for all \(r\in {\mathbb {N}}\), where \(C_\alpha ^r\ge 1\) for all \(\alpha \in {\mathbb {N}}_r^n\). Since \(f-f^\star \ge 0\) on S(g), by Theorem 5, there exists \(K\in {\mathbb {N}}\) such that
Note that K depends only on f, g and \(\varepsilon \). Assume that \(k\ge K\). Since \(\theta ^{k-K} \in \varSigma {[ x ]_{k-K}}\), one has
Then there exist \(G_j\succeq 0\), \(j=1,\dots ,m\) such that
Let \(y\in \bar{P}\). From these and since \({M_{k + d - {u_j}}}( {{g_j}y})\succeq 0\),
Thus, for every \(\beta \in {\mathbb {N}}_{k + d}^n\),
since every element \(y_{2\alpha }\) on the diagonal of the positive semidefinite matrix \({M_{k + d}}( {y})\) is nonnegative. Thus, \({y_{2\beta }}\le 2(C - f^\star )\varepsilon ^{-1}\) for every \(\beta \in {\mathbb {N}}_{k + d}^n\). Since \({M_{k + d }}( {y})\succeq 0\), \(| {{y_{\alpha + \beta }}} | \le {y_{2\beta }} \le 2(C - \underline{f})\varepsilon ^{-1}\) for all \(\alpha ,\beta \in {\mathbb {N}}_{k + d}^n\). This implies that \(\Vert y\Vert _2\) is bounded by \(2(C - \underline{f})\varepsilon ^{-1}\sqrt{s(2(d + k))}\). Since the objective function of (1.41) is linear and the feasible set of (1.41) is closed and bounded, the set \(P^\star \) of optimal solutions of (1.41) is nonempty and bounded. By using Trnovska’s result [59, Corrollary 1], \(\rho _k^2(\varepsilon )=\tau _k^2(\varepsilon )\), yielding the desired conclusion. \(\square \)
1.4 Proof of Proposition 4
Proof
We do a similar process as the proof of Proposition 3. The difference is the bound of y here obtained by the inequality constraint \(g_m=f-\underline{f}\ge 0\) in stead of Positivstellensatz. Thus, we do not need k sufficient large here. More explicitly, one has \(\Vert y\Vert _2 \le (C - \underline{f})\varepsilon ^{-1}\sqrt{s(2(d + k))}\), since for every \(\beta \in {\mathbb {N}}_{k + d}^n\),
1.5 Proof of Lemma 4
Proof
Let us show that \((\xi _t)_{t=0,\dots ,n}\) is real sequence. Obviously, \(\xi _t\ge 0\), \(t=0,\dots ,n\). Set \(r_t=\xi _t^{1/2}\), \(t=0,\dots ,n\). Then
where \(d(a,A):=\inf \{\Vert x-a\Vert _2:\, x\in A\}\) for \(a\in {\mathbb {R}}^n\) and \(A\subset {\mathbb {R}}^n\). It is sufficient to prove that \(r_t\) is real, \(t=0,\dots ,n\). It is easy to see that S(g, h) is closed and S(g, h) is nonempty by assumption. From this, \({r_0}\) is nonnegative real and \(S( g,h ) \cap \partial {B( {{a_0},{r_0}} )} \) is also closed and nonempty. It implies that \(r_1\) is nonnegative real and
is also closed and nonempty. By induction, for \(t \in \{ {0,\dots ,n} \}\), \(r_t\) is nonnegative real and
is closed and nonempty. Thus,
Let \({x^\star } \in S( {g,h \cup \{ {{\xi _t} - {{ \Vert {x - {a_t}} \Vert }_2^2}:\, t = 0,\dots ,n} \}} )\). Then \({x^\star } \in \partial B\left( {{a_0},{r_0}} \right) \cap \dots \cap \partial B( {{a_n},{r_n}} )\). It follows that \({\Vert {{x^\star } - {a_t}} \Vert _2^2} = r_t^2=\xi _t,\, t = 0,\dots ,n\). For \(t = 1,\dots ,n\),
It implies (4.31). Denote
The system (4.31) can be rewritten as \(Ax^\star =b\). Since \({a_j} - {a_0},\, j = 1,\dots ,n\) are linearly independent in \({\mathbb {R}}^n\), A is invertible. Hence, \(x^\star \) is determined uniquely by \(x^\star =A^{-1}b\). \(\square \)
1.6 Proof of Theorem 8
Proof
We will prove by induction that for \(t \in \left\{ {0,\dots ,n} \right\} \),
For \(t=0\), (4.33) is the SDP relaxation of order \(k+w\) of
By assumption, \({( {\eta _k^0} )_{k \in {\mathbb {N}}}}\) finitely converges to \(\xi _0\), i.e. there exist \(K_0\in {\mathbb {N}}\) such that \(\eta _k^0 = {\xi _0}\) for all \(k\ge K_0\). It follows that (1.43) is true for \(t=0\). Assume that (1.43) is true for \(t=T\), i.e.
We will show that (1.43) is true for \(t=T+1\). By (4.33) and (1.44), for all \(k\ge K_T\),
is the SDP relaxation of order \(k+w\) of problem
By assumption, \({( {\eta _k^{T+1}})_{k \ge K_T}}\) finitely converges to \(\xi _{T+1}\), i.e. there exist \(K_{T+1}\ge K_T\) such that \(\eta _k^{T+1} = {\xi _{T+1}}\) for all \(k\ge K_{T+1}\). It follows that (1.44) is true for \(t=T+1\). Thus, (1.44) is true for \(t=0,\dots ,n\). For \(t=n\), there exists \(K=K_n\in {\mathbb {N}}\) such that for all \(k\ge K\), \(\eta _k^t = {\xi _t}\), \(t = 0,\dots ,n\). Let \(k\ge K\) be fixed. Let y be the solution of problem
which is the SDP hierarchy relaxation of order \(k+w\) of problem
We will prove that this latter problem has a unique minimizer \(x^\star \). Set \(\hat{h}:=h\cup \{ \xi _0-\Vert x-a_0\Vert _2^2\}\). Then
By Lemma 4 (with \(S(g,h):=S(g,\hat{h})\)), there exists \(x^\star \) such that
Let a be a minimizer of the above POP with value \(\xi _n\). Then
and \({\Vert {a - {a_n}} \Vert }_2^2=\xi _n\). It follows that
From this and by (1.45), \(a=x^\star \). Since the above POP with optimal value \(\xi _n\) has a unique minimizer \(x^\star \) and its Lasserre’s hierarchy has finite convergence, the solution y of the SDP with optimal value \(\eta _k^n\) must have a representing 1-atomic measure \(\mu \) supported on \(x^\star \). Then y satisfies the flat extension condition when k is large enough and \({\mathop \mathrm{supp}\nolimits }( \mu ) = \{ {{x^\star }}\}\subset S(g,h)\). The conclusion follows. \(\square \)
Rights and permissions
About this article
Cite this article
Mai, N.H.A., Lasserre, JB. & Magron, V. Positivity certificates and polynomial optimization on non-compact semialgebraic sets. Math. Program. 194, 443–485 (2022). https://doi.org/10.1007/s10107-021-01634-1
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10107-021-01634-1
Keywords
- Nonnegativity certificate
- Putinar’s Positivstellensatz
- Basic semialgebraic set
- Sums of squares
- Polynomial optimization
- Semidefinite programming
- Moment-SOS hierarchy
- Uniform denominators
- Polynomial systems