On the Sample Complexity and Optimization Landscape for Quadratic Feasibility Problems

Finding a solution to a system of quadratic equations is an important problem with a wide range of applications. It arises in areas such as power system state estimation [1], phase retrieval [2, 3, 4, 5], x-ray crystallography [6], the turnpike problem [7], and unlabeled distance geometry problems[8, 9] among others. Such problems are often reduced to a quadratic feasibility problem, where one is concerned with finding a feasible vector $\mathbf{x}$ that conforms to a set of quadratic observations of the form $\{\langle A_{i}\mathbf{x},\mathbf{x}\rangle\}_{i=1}^{m}$ with respect to a set $\left\{A_{i}\right\}_{i=1}^{m}$ of measurement matrices. Formally, it can be cast as:

The quadratic feasibility problem is an instance of quadratically constrained quadratic programs (QCQPs)[10], which has enjoyed a long and rich research history dating back to 1941[11]. Given their broad applicability to critical problems, research in QCQPs continues to be of active interest [10, 12, 13, 14]. Unfortunately, it is known that solving QCQPs is an NP-hard problem [15]. This combined with the lack of tractable duality properties [16] has made it hard to establish a sound theoretical framework for understanding the solutions and computing them. However, an extremely productive line of research has instead considered subclasses of QCQPs that are both practically relevant and can be analyzed. In this paper, we take a similar approach and identify an important subclass of QCQPs that have a broad range of applications. In particular, we analyze the quadratic feasibility problem, and establish conditions under which such problems are identifiable, and then show that these conditions are in-fact sufficient for the efficient computation of their solutions.

We start by considering quadratic functions $\{\langle A_{i}\mathbf{x},\mathbf{x}\rangle\}_{i=1}^{m}$, where $\mathbf{x}\in\mathbb{C}^{n}$ and $A_{i}\in\mathbb{C}^{n\times n}$ are Hermitian matrices. We focus on their ability to generate injective maps up-to a phase factor (note that the quadratic functions $\{\langle A_{i}\mathbf{x},\mathbf{x}\rangle\}_{i=1}^{m}$ are invariant to phase shifts). We establish a relationship between injectivity and isometry and show that, in real world scenarios, it is not difficult for a set of quadratic measurements $\{\langle A_{i}\mathbf{x},\mathbf{x}\rangle\}_{i=1}^{m}$ to possess such an isometry property by establishing that this holds with very high probability when the matrices $\{A_{i}\}_{i=1}^{m}$ are complex Gaussian random matrices.

After establishing injectivity (and hence the identifiability) of the problem, we consider the question of computationally tractable approaches to actually find a feasible solution. Toward this end, a natural approach is to optimize the appropriate $\ell_{2}$-loss. Unfortunately, this turns out to be a nonconvex problem, that is NP-hard to solve in general. However, we show that under the same conditions required to establish injectivity, the landscape of this optimization problem is well-behaved. This allows us to establish that any gradient based algorithm with an arbitrary initialization can recover a globally optimal solution almost surely.

The rest of the paper is organized as follows. Section II highlights the main results of this work. We discuss some related work in Section III. In Section IV we establish and analyze isometry/identifiability properties of the mapping $\{\langle A_{i}\mathbf{x},\mathbf{x}\rangle\}$ when the measurement matrices are complex Gaussian. Finally, Section V casts the problem as a quadratic loss minimization problem (suitable for efficient algorithms) and establishes favorable properties of the loss landscape that allows one to find a solution using gradient-based methods with arbitrary initial points.

Before we state our main results of the paper, we introduce some notation that will be used throughout the paper.

QCQPs have enjoyed a lot of attention over the last century. However, due to the limitation of the duality properties of QCQPs [19], a significant fraction of research has focused predominantly on heuristic approaches to their solution [20, 21, 22]. Recently, an ADMM-based method has been proposed in[23] with an asymptotic convergence result based on the duality properties QCQPs. Our results in this paper bring new insights to this area by analyzing a subset of QCQPs, namely, the quadratic feasibility problems.

The quadratic feasibility problem (P1) arises in many applications, including phase retrieval [2] and power system state estimation [1]. Phase retrieval in and of itself finds applications in a wide variety of fields such as imaging, optics, quantum tomography, audio signal processing with a wide literature, including [4, 2, 24, 5]. In [3], an approximate $\ell_{1}$ isometry property was established for the phase retrieval problem, but the bounds therein are not strong enough to provide RIP-like guarantees. In this paper, we improve these bounds to establish isometry results for a large class of problems and provide RIP-type bounds for the case when $\{A_{i}\}_{i=1}^{m}$ are complex Gaussian random matrices.

A feasibility problem is often cast as a minimization problem with a suitably chosen loss function. Even with a nonconvex objective, gradient based methods have proven to work for phase retrieval [24, 5, 2], matrix factorization [25, 26] and robust linear regression [27]. The work in [28] has established landscape properties for the phase retrieval problem, which sheds light on the success of gradient based methods in solving the problem. In this work, we extend these results to a wider class of problems along with additional insights into the problem properties. In [29], it was shown that many nonconvex loss functions have specific landscape properties, which allows gradient based algorithm to recover a globally optimal solution without any additional information. One unfortunately cannot readily transport those results to our setting, mainly due to the significant differences between the real and complex vector spaces. For instance, a quadratic feasibility problem in $\mathbb{R}^{n}$ has only two isolated local minima, while it has a continuum of minima in $\mathbb{C}^{n}$.

The authors in [30] provide lower bounds on the minimum number of independent measurements required for a successful recovery for the quadratic feasibility problem. More recently, [31] showed that the quadratic feasibility problem can be solved, with high probability, by gradient descent provided a good initialization is used. In contrast, the current work takes a parallel track by analyzing the landscape of the associated $\ell_{2}$-loss function. In particular, for the $\ell_{2}$-loss function, we prove that all local minima are global and all saddle points are strict. Thus, our results enable gradient based algorithms with arbitrary initialization to recover the solution for the quadratic feasibility problem.

As a first step towards establishing our main results, we start by characterizing when the quadratic mapping $\mathcal{M}_{\mathcal{A}}$ defined in (1) is in fact an injective mapping. Notice that the injectivity of the mapping is equivalent to the problem being identifiable (and hence solvable). It is also worth noting that in the context of the phase retrieval problem, when $\mathcal{M}_{\mathcal{A}}$ is injective, it is said to possess the phase retrievability property [30].

We refer the reader to Appendix VI-A for a detailed proof. Lemma 1 characterizes the injectivity property of $\mathcal{M}_{\mathcal{A}}$ in terms of the action of the $A_{i}$ matrices on the outer product $[[\mathbf{u},\mathbf{v}]]$. In particular, it says that the mapping $\mathcal{M}_{\mathcal{A}}$ is injective if and only if the matrices $A_{i}$ do not distort the outer-product $[[\mathbf{u},\mathbf{v}]]$ too much. We use this characterization to obtain a more tractable condition that allows us to establish the injectivity of $\mathcal{M}_{\mathcal{A}}$ in the case of Gaussian measurement matrices (forthcoming in Lemma 2).

Our tractable condition is based on what we call the stability of the mapping. Informally speaking, the mapping $\mathcal{M}_{\mathcal{A}}$ is stable if for any given $\epsilon>0$, there exists $\delta>0$ such that $\|\mathcal{M}_{\mathcal{A}}(\mathbf{x})-\mathcal{M}_{\mathcal{A}}(\mathbf{y})\|\geq\delta$ whenever the vectors $\mathbf{x},\mathbf{y}\in\widehat{\mathbb{C}}^{n}$ satisfy $d(\mathbf{x},\mathbf{y})\geq\epsilon$. Such stability properties have been considered in the specific context of phase retrieval; see e.g., [4, 3, 32, 33].

The constants $\alpha,\beta$ depend on the choice of the metric $d(\cdot,\cdot)$. Throughout the paper, we will work with the metric $d$ as defined in (2). Note that our definition of $(\alpha,\beta)$-stability sandwiches the norm of the measurement differences, which is in contrast to the stability concept considered for phase retrieval in [4, 32]. The constants $\alpha,\beta$ can also be thought of as a condition number, thereby allowing one to quantify the quality of the map; the higher the ratio between $\alpha$ and $\beta$, the better the ability of the mapping to distinguish between two distinct inputs.

With this definition of stability in place, the next lemma states that the map $\mathcal{M}_{\mathcal{A}}$ is injective if and only if it is stable.

Please refer to Appendix VI-B for a proof of the lemma. This result demonstrates the usefulness of both our choice of the metric $d(\cdot,\cdot)$ from (2) and of our definition of stability (Definition 1). As we will see in what follows, Lemma 2 allows one to assess the conditions under which the measurement model implied by the mapping $\mathcal{M}_{\mathcal{A}}$ is identifiable.

Next, we turn our attention to the question of when one can establish the above stability condition, and thereby establish the identifiability of the underlying measurement model. To do so, we take our cues from the compressive sensing and phase retrieval literature, and we assume that Hermitian matrices in the set $\mathcal{A}=\{A_{i}\}_{i=1}^{m}$ are sampled from a complex Gaussian distribution. More precisely, we assume that each entry in the upper triangle (including the diagonal) of $A_{i}$ is drawn independently from $\mathcal{N}(0,1)$. The remaining entries are determined by the fact that the matrices $A_{i}$ are Hermitian.

We first observe that, when $\mathcal{A}$ is chosen as described above, then $\mathbb{E}\left[\left|\langle A_{i},\mathbf{x}\mathbf{x}^{\ast}-\mathbf{y}\mathbf{y}^{\ast}\rangle\right|^{2}\right]=d(\mathbf{x},\mathbf{y})^{2}$ for all $\mathbf{x},\mathbf{y}\in\mathbb{C}^{n}$, and for all $i\in[m]$ (see Appendix VII-A for more details). Our next result shows that these quantities are actually concentrated around their expected values.

Please see Appendix VII-A for a proof of Lemma 3. We would like to draw attention to the fact that proving such a concentration result requires careful attention to the fact that we are operating in the complex domain; our proof techniques may be of independent interest. Notice that this concentration result only holds for a fixed pair of vectors. We however need the concentration result to be uniform over all pairs of vectors in $\mathbb{C}^{n}$. For this, we will next adapt a standard covering argument as follows; we use $S^{n-1}$ to denote the $n-$dimensional unit sphere in $\mathbb{C}^{n}$.

We refer the reader to Appendix VII-B for a proof of this lemma. This argument is not new and has found application in several results; see e.g., [31, 34, 35].

We are finally ready to state and prove our first main result.

In the preceding section, we established conditions under which the mapping $\mathcal{M}_{\mathcal{A}}$ represents an identifiable measurement model. We next consider determining a feasible solution by applying efficient methods to the nonconvex optimization-based reformulation of the feasibility problem, as given in (P2). As noted earlier, in general, gradient-based aproaches need not converge to a global minimum of the $\ell_{2}$-loss minimization problem (P2). Lately, nonconvex optimization has received considerable attention. In particular, methods like SGD and other gradient based methods have been shows to be astonishingly successful in converging to global minima in many nonconvex problems [36, 37, 38]. Arguably, the reason for this is that the optimization landscape of these (somewhat well-behaved) nonconvex problems enjoy some advantageous properties which are crucial in the empirical success of the gradient based methods [39, 40, 28]. The work in [28] proves that the $\ell_{2}$-loss function for the phase retrieval problem enjoys properties such as all local minima being global, and each saddle point having a strictly negative curvature. This work adds to this rich body of work by demonstrating similarly advantageous properties of the optimization landscape for the quadratic feasibility problem. Before we proceed we observe that, since the $\ell_{2}$-loss function is not differentiable in the complex space $\mathbb{C}^{n}$, it is challenging to address the problem (P2) in a standard way. In what follows, we instead use techniques from Wirtinger calculus [2]. Our first step is to define the notion of a strict saddle function.

Intuitively, this implies that every $\mathbf{x}\in\mathbb{C}^{n}$ either violates optimality (condition 1 and 2) or is close to a local optimum. A line of recent work [41, 42, 27] has explored the efficacy of gradient based methods in finding a local optimum of functions satisfying Definition 2.

We analyze the optimization landscape of (P2) when our measurement matrices are Hermitian and complex Gaussian, and show that with a high probability every local minimum is in fact global (upto the equivalence relation $\sim$). Our next main result states that the function $f$ in (P2) is strict saddle.

Recall that the distance metric $d$ is defined on the quotient space $\widehat{\mathbb{C}}^{n}$, therefore, the second statement above says that
$\mathbf{w}\sim\mathbf{z}$. We refer the reader to Appendix VIII-C for the details, but we will give here the brief idea of the proof.

Finally, we remark that the properties of the optimization landscape that we have established allows one to use any gradient based iterative method to find a global optimum of problem (P2) – hence, find a solution to the quadratic feasibility problem. Furthermore, our results above also imply that a gradient method, with an arbitrary initial point, would work, which is in a sharp contrast with the existing works,such as [31]. Formally, we have the following result.

Given the landscape properties we have derived in Theorem 2, this result follows in a straightforward manner, for instance, from Theorem 4.1 in [41]. We would like to remark here that the broad flow of ideas in our proof of Theorem 2 bears similarities to those in papers like [29, 43]. However, to the best of our knowledge, the present paper is the first to derive such results in the complex domain.