1 Introduction
We thank all the discussants for their careful reading of our paper, and for their insightful critiques. We would also like to thank the editors for organizing this discussion. Our paper contributes to the area of highdimensional statistics which has received much attention over the past several years across the statistics, machine learning and signal processing communities. In this rejoinder we clarify and comment on some of the points raised in the discussions. Finally, we also remark on some interesting challenges that lie ahead in latent variable modeling.
Briefly, we considered the problem of latent variable graphical model selection in the Gaussian setting. Specifically, let
be a zeromean Gaussian random vector in
with and representing disjoint subsets of indices in with and . Here the subvector represents the observed variables and the subvector represents the latent variables. Given samples of only the variables , is it possible to consistently perform model selection? We noted that if the number of latent variables is small relative to and if the conditional statistics of the observed variables conditioned on the latent variables are given by a sparse graphical model, then the marginal concentration matrix of the observed variablesis given as the sum of a sparse matrix and a lowrank matrix. As a first step we investigated the identifiability of latent variable Gaussian graphical models—effectively, this question boils down to one of uniquely decomposing the sum of a sparse matrix and a lowrank matrix into the individual components. By studying the geometric properties of the algebraic varieties of sparse and lowrank matrices, we provided natural sufficient conditions for identifiability and gave statistical interpretations of these conditions. Second, we proposed the following regularized maximumlikelihood estimator to decompose the concentration matrix into sparse and lowrank components:
(1)  
Here represents the sample covariance formed from samples of the observed variables, is the Gaussian loglikelihood function, represents the estimate of the conditional graphical model of the observed variables conditioned on the latent variables, and represents the extra correlations induced due to marginalization over the latent variables. The norm penalty induces sparsity in and the trace norm penalty induces lowrank structure in . An important feature of this estimator is that it is given by a convex program that can be solved efficiently. Our final contribution was to establish the highdimensional consistency of this estimator under suitable assumptions on the Fisher information underlying the true model (in the same spirit as irrepresentability conditions for sparse model selection RavWRY2008 , ZhaY2006 ).
2 Alternative estimators
A number of the commentaries described alternative formulations for estimators in the latent variable setting.
2.1 EMbased methods
The discussions by Yuan and by Lauritzen and Meinshausen describe an EMbased alternative in which the rank of the matrix is explicitly constrained:
(2)  
(3) 
The experimental results based on this approach seem quite promising, and certainly deserve further investigation. On the one hand, we should reiterate that the principal motivation for our convex optimization based formulation was to develop a method for latent variable modeling with provable statistical and computational guarantees. One of the main drawbacks of EMbased methods is the existence of local optima in the associated variational formulations, thus leading to potentially different solutions depending on the initial point. On the other hand, one of the reasons for the positive empirical behavior observed by Yuan and by Lauritzen and Meinshausen may be that all the local optima in the experimental settings considered by the authors may be “good” models. Such behavior has in fact been rigorously characterized recently for certain nonconvex estimators in some missing data problems LohW2011 .
One of the motivations for the EM proposal of Yuan and of Lauritzen and Meinshausen seems to be that there are fairly mature and efficient solvers for the graphical lasso. As our estimator is relatively newer and as its properties are better understood going forward, we expect that more efficient solvers will be developed for (1) as well. Indeed, the LogdetPPA solver WanST2010 that we cite in our paper already scales to instances involving several hundred variables, while more recent efforts MaXZ2012 have resulted in algorithms that scale to instances with several thousand variables.
2.2 Thresholding estimators
Ren and Zhou propose and analyze an interesting thresholding based estimator for decomposing a concentration matrix into sparse and lowrank components. They apply a twostep procedure— norm thresholding followed by trace norm thresholding—to obtain the sparse component followed by the lowrank component. Roughly speaking, this twostep estimator can be viewed as the application of the first cycle of a block coordinate descent procedure to compute our estimator that alternately updates the sparse and lowrank pieces (we also refer the reader to the remarks in AgaNW2011 ).
However, in Theorem in the discussion by Ren and Zhou, a quite stringent assumption requires that in some scaling regimes the true lowrank component must vanish, that is, . The reason for this condition is effectively to ensure sign consistency in recovering the sparse component. In a pure sparse model selection problem (with no lowrank component in the population), the deviation away from the sparse component is given only by noise due to finite samples and this deviation is on the order of in the Gaussian setting—consequently, sparse model selection via norm thresholding is signconsistent when the minimum magnitude nonzero entry in the true model is larger than . In contrast, if the true model consists of both a sparse component and a lowrank component, the total deviation away from the sparse component in the finite sample regime is given by both sample noise as well as the lowrank component. This seems to be the reason for the stringent assumption on the vanishing of the lowrank component in Theorem of Ren and Zhou.
More broadly, one of the motivations of Ren and Zhou in proposing and analyzing their estimator is that it may be possible to weaken the assumptions on the minimum magnitude nonzero entry of the true sparse component
and the minimum nonzero singular value
of the true lowrank component —whether this is possible under less stringent assumptions on is an interesting question, and we comment on this point in Section 3 in the more general context of potentially improving the rates in our paper.2.3 Other proposals
Giraud and Tsybakov propose two alternative estimators for decomposing a concentration matrix into sparse and lowrank components. While our approach (1) builds on the graphical lasso, their proposed approaches build on the Dantzig selector of Candès and Tao CanT2007 and the neighborhood selection approach of Meinshausen and Bühlmann MeiB2006 . Several comments are in order here.
First, we note that the extension of neighborhood selection proposed by Giraud and Tsybakov to deal with the lowrank component begins by reformulating the neighborhood selection procedure to obtain a “global” estimator that simultaneously estimates all the neighborhoods. This reformulation touches upon a fundamental aspect of latent variable modeling. In many applications marginalization over the latent variables typically induces correlations between most pairs of observed variables—consequently, local procedures that learn model structure one node at a time are illsuited for latent variable modeling. Stated differently, requiring that a matrix be sparse with few nonzeros per row or column (e.g., expressing preference for a graphical model with bounded degree) can be done by imposing columnwise constraints. On the other hand, the constraint that a matrix be lowrank is really a global constraint expressed by requiring all minors of a certain size to vanish. Thus, any estimator for latent variable modeling (in the absence of additional conditions on the latent structure) must necessarily be global in nature.
Second, we believe that the reformulation based on the Dantzig selector perhaps ought to have an additional constraint. Recall that the Dantzig selector CanT2007 constrains the norm (the dual norm of the norm) of the correlated residuals rather than the norm of the residuals as in the lasso. As the dual norm of our combined /trace norm regularizer involves both an norm and a spectral norm, the following constraint set may be more appropriate in the Dantzig selector based reformulation of Giraud and Tsybakov:
Finally, we note that the Dantzig selector of CanT2007 has the property that its constraint set contains the lasso solution (with the same choice of regularization/relaxation parameters). In contrast, this property is not shared in general by the Dantzig selector reformulation of Giraud and Tsybakov in relation to our regularized maximumlikelihood estimator (1). It is unclear how one might achieve this property via suitable convex constraints in a Dantzig selector type reformulation of our estimator.
In sum, both of these alternative estimators deserve further study.
3 Comments on rates
Several of the commentaries (Wainwright, Giraud and Tsybakov, Ren and Zhou and Candès and Soltanolkotabi) bring up the possibility of improving the rates given in our paper. At the outset we believe that samples is inherent to the latent variable modeling problem if spectral norm consistency is desired in the lowrank component. This is to be expected since the spectral norm of the deviation of a sample covariance from the underlying population covariance is on the order of . However, some more subtle issues remain.
Giraud and Tsybakov point out that one may be concerned purely with estimation of the sparse component, and that the lowrank component may be a “nuisance” parameter. While this is not appropriate in every application, in problem domains where the conditional graphical model structure of the observed variables is the main quantity of interest one can imagine quantifying deviations in the lowrank component via “weaker” norms than the spectral norm—this may lead to consistent estimates for the sparse component with samples. The analysis in our paper does not rule out this possibility, and a more careful investigation is needed to establish such results.
Ren and Zhou suggest that while may be required for consistent estimation, one may be able to weaken the assumptions on and (the minimum magnitude nonzero entry of the sparse component and the minimum nonzero singular value of the lowrank component, respectively). From the literature on sparse model selection, a natural lower bound on the minimum magnitude nonzero entry for consistent model selection is typically given by the size of the noise measured in the norm (the dual of the regularizer). Building on this intuition, a natural lower bound that one can expect in our setting on is , while a natural bound on would be . The reason for this suggestion is that is the dual norm of the regularizer used in our paper. Therefore, it may be possible to only require and . However, one issue here is that the norm bound kicks in when
with probability approaching one polynomially fast, while the spectral norm bound only kicks in when
but holds with probability approaching one exponentially fast. Thus (as also noted by Giraud and Tsybakov), it may be possible that is required for overall consistent estimation, but that the assumption on could be weakened by only requiring that the probability of consistent estimation approach one polynomially fast.Candès and Soltanolkotabi comment that it would be of interest to establish an “adaptivity” property whereby if no lowrank component were present, the number of samples required for consistent estimation would boil down to just the rate for sparse graphical model selection, that is, . While such a feature would clearly be desirable to establish for our estimator, one potential roadblock may be that our estimator (1) “searches” over a larger classes of models than just those given by sparse graphical models; consequently, rejecting the hypothesis that the observed variables are affected by any latent variables may require that . This question deserves further investigation and, as suggested by Candès and Soltanolkotabi, recent results on adaptivity could inform a more refined analysis of our estimator.
Finally, Wainwright suggests the intriguing possibility that faster rates may be possible if the lowrank component has additional structure. For example, there may exist a sparse factorization of the lowrank component due to special structure between the latent and observed variables. In such settings the trace norm regularizer applied to the lowrank component is not necessarily the tightest convex penalty. In recent joint work by the authors and Recht ChaRPW2010 , a general framework for constructing convex penalty functions based on some desired structure is presented. The trace norm penalty for inducing lowrank structure is motivated from the viewpoint that a lowrank matrix is the sum of a small number of rankone matrices and, therefore, the norm induced by the convex hull of the rankone matrices (suitably scaled) is a natural convex regularizer as this convex hull (the trace norm ball) carries precisely the kind of facial structure required for inducing lowrank structure in matrices. In this spirit, one can imagine constructing convex penalty functions by taking the convex hull of sparse rankone matrices. While this convex hull is in general intractable to represent, relaxations of this set that are tighter than the trace norm ball could provide faster rates than can be obtained by using the trace norm.
4 Weakening of irrepresentability conditions
Wainwright asks a number of insightful questions regarding the potential for weakening our Fisher information based conditions. Giraud and Tsybakov also bring up connections between our conditions and irrepresentability conditions in previous papers on sparse model selection RavWRY2008 , ZhaY2006 .
In order to better understand if the Fisher information based conditions stated in our paper are necessary, Wainwright raises the question of obtaining a converse result by comparing to an oracle method that directly minimizes the rank and the cardinality of the support of the components. A difficulty with this approach is that we don’t have a good handle on the set of matrices that are expressible as the sum of a sparse matrix and a lowrank matrix. The properties of this set remain poorly understood, and developing a better picture has been the focus of research efforts in algebraic geometry DrtSS2007 and in complexity theory Val1977 . Nonetheless, a comparison to oracle estimators that have side information about the support of the sparse component and the row/column spaces of the lowrank component (in effect, side information about the tangent spaces at the two components) appears to be more tractable. This is closer to the viewpoint we have taken in our paper in which we consider the question of identifiability of the components given information about the underlying tangent spaces. Essentially, our Fisher information conditions state that these tangent spaces must be sufficiently transverse with respect to certain natural norms and in a space in which the Fisher information is the underlying innerproduct. More generally, as also pointed out by Giraud and Tsybakov, the necessity of Fisher information based conditions is an open question even in the sparse graphical model selection setting considered in RavWRY2008 . The experimental studies in RavWRY2008 describing comparisons to neighborhood selection in some simple cases provide a good starting point.
Wainwright raises the broader question of consistent model selection when transversality of the underlying tangent spaces does not hold. One approach AgaNW2011 is to quantify the level of identifiability based on a “spikiness” condition. A more geometric viewpoint may be that only those pieces of the sparse and lowrank components that do not lie in the intersection of their underlying tangent spaces are fundamentally identifiable and, therefore, consistency should be quantified with respect to these identifiable pieces.
Giraud and Tsybakov ask about the interpretability of our conditions and . These quantities are geometric in nature and relate to the tangent space conditions for identifiability. In particular, they are closely related to (and bounded by) the incoherence of the row/column spaces of the lowrank component and the maximum number of nonzeros per row/column ChaSPW2011 . These latter quantities have appeared in many papers on sparse graphical model selection (e.g., MeiB2006 , RavWRY2008 ) as well as on lowrank matrix completion CanR2009 , and computing them is straightforward. In our previous work on matrix decomposition ChaSPW2011 , we note that these quantities are bounded for natural random families of sparse and lowrank matrices based on results in CanR2009 .
5 Experimental issues and applications
Lauritzen and Meinshausen as well as Giraud and Tsybakov raise several points about the choice of the regularization parameters. Choosing these parameters in a datadriven manner (e.g., using the methods described in MeiB2010 ) is clearly desirable. We do wish to emphasize that the sensitivities of the solution with respect to the parameters and are qualitatively different. As described in our main theorem and in our experimental section, the solution of our estimator (1) is stable for a range of values of (see also ChaSPW2011 )—this point is observed by Yuan as well in his experiments. Further, the choice of ideally should not depend on , while the choice of clearly should.
On a different point regarding experimental results, Giraud and Tsybakov suggest at the end of their discussion that latent variable models don’t seem to provide significantly more expressive power than a sparse graphical model. In contrast, Yuan’s synthetic experiment seems to provide compelling evidence that our approach (1) provides better performance relative to models learned by the graphical lasso. The reason for these different observations may be tied to the manner in which their synthetic models were generated. Specifically, latent variable model selection using (1) is likely to be most useful when the latent variables affect many observed variables upon marginalization (e.g., latent variables are connected to many observed variables), while the conditional graphical model among the observed variables conditioned on the latent variables is sparse and has bounded degree. This intuition is based on the theoretical analysis in our paper and is also the setting considered in the experiment in Yuan’s discussion (as well as in the synthetic experiments in our paper). On the other hand, the experimental setup followed by Giraud and Tsybakov seems to generate a graphical model with large maximum degree and low average degree, and randomly selects a subset of the variables as latent variables. It is not clear if these latent variables are the ones with large degree, which may explain their remarks.
Finally, we note that sparse and lowrank matrix decomposition is relevant in applications beyond the one described in our paper. As observed by Lauritzen and Meinshausen, a natural matrix decomposition problem involving covariance matrices may arise if one considers directed latent variable models in the spirit of factor analysis. In such a context the covariance matrix may be expressed as the sum of a lowrank matrix and a sparse (rather than just diagonal) matrix, corresponding to the setting in which the distribution of the observed variables conditioned on the latent variables is given by a sparse covariance matrix. More broadly, similar matrix decomposition problems arise in domains beyond statistical estimation such as optical system decomposition, matrix rigidity and system identification in control ChaSPW2011 , as well as others as noted by Candès and Soltanolkotabi.
6 Future questions
Our paper and the subsequent discussions raise a number of research and computational challenges in latent variable modeling that we wish to highlight briefly.
6.1 Convex optimization in R
As mentioned by Lauritzen and Meinshausen, R remains the software of choice for practitioners in statistics. However, some of the recent advances in highdimensional statistical estimation have been driven by sophisticated convex optimization based procedures that are typically prototyped using packages such as SDPT3 TohTT and others in Matlab and Python. It would be of general interest to develop packages to invoke SDPT3 routines directly from R.
6.2 Sparse/lowrank decomposition as infimal convolution
Given a matrix , consider the following function:
(4) 
It is clear that is a norm, and it can be viewed as the infimal convolution Roc1996 of the (scaled) norm and the trace norm. In essence, it is a norm whose minimization induces matrices expressible as the sum of sparse and lowrank components (see also the atomic norm viewpoint of ChaRPW2010 ). We could then effectively restate (1) as
and then decompose by solving (4). This twostep approach suggests the possibility of decoupling the decomposition problem from the conditions fundamentally required for consistency via regularized maximumlikelihood, as the latter only ought to depend on the composite norm . This decoupling also highlights the different roles played by the parameters and (as discussed in Section 5). More broadly, such an approach may be useful as one analyzes general regularizers, for example, convex penalties other than the trace norm as described in Section 3.
6.3 NonGaussian latent variable modeling
As described in our paper and as raised by Wainwright, latent variable modeling with nonGaussian variables is of interest in many applications. Both the computational and algebraic aspects present major challenges in this setting. Specifically, the secant varieties arising due to marginalization in nonGaussian models (e.g., in models with categorical variables) are poorly understood, and computing the likelihood is also intractable. An approach based on matrix decomposition as described in our paper may be appropriate, although one would have to quantify the effects of the Gaussianity assumption.
References
 (1) [auto:STB—2012/05/30—10:51:56] Agarwal, A.A., Negahban, S.S. Wainwright, M. J.M. J. (2011). Noisy matrix decomposition via convex relaxation: Optimal rates in high dimensions. Preprint. imsref
 (2) [mr] Candes, EmmanuelE. Tao, TerenceT. (2007). The Dantzig selector: Statistical estimation when is much larger than . Ann. Statist. 35 2313–2351. doi=10.1214/009053606000001523, issn=00905364, mr=2382644 imsref
 (3) [mr] Candès, Emmanuel J.E. J. Recht, BenjaminB. (2009). Exact matrix completion via convex optimization. Found. Comput. Math. 9 717–772. doi=10.1007/s1020800990455, issn=16153375, mr=2565240 imsref
 (4) [auto:STB—2012/05/30—10:51:56] Chandrasekaran, V.V., Recht, B.B., Parrilo, P. A.P. A. Willsky, A. S.A. S. (2010). The convex geometry of linear inverse problems. Preprint. imsref
 (5) [mr] Chandrasekaran, VenkatV., Sanghavi, SujayS., Parrilo, Pablo A.P. A. Willsky, Alan S.A. S. (2011). Ranksparsity incoherence for matrix decomposition. SIAM J. Optim. 21 572–596. doi=10.1137/090761793, issn=10526234, mr=2817479 imsref
 (6) [mr] Drton, MathiasM., Sturmfels, BerndB. Sullivant, SethS. (2007). Algebraic factor analysis: Tetrads, pentads and beyond. Probab. Theory Related Fields 138 463–493. doi=10.1007/s0044000600332, issn=01788051, mr=2299716 imsref
 (7) [auto:STB—2012/05/30—10:51:56] Loh, P.P. Wainwright, M. J.M. J. (2011). Highdimensional regression with noisy and missing data: Provable guarantees with nonconvexity. Preprint. imsref
 (8) [auto:STB—2012/05/30—10:51:56] Ma, S.S., Xue, L.L. Zou, H.H. (2012). Alternating direction methods for latent variable Gaussian graphical model selection. Technical report, IMA and School of Statistics, Univ. Minnesota. imsref
 (9) [mr] Meinshausen, NicolaiN. Bühlmann, PeterP. (2006). Highdimensional graphs and variable selection with the lasso. Ann. Statist. 34 1436–1462. doi=10.1214/009053606000000281, issn=00905364, mr=2278363 imsref
 (10) [mr] Meinshausen, NicolaiN. Bühlmann, PeterP. (2010). Stability selection. J. R. Stat. Soc. Ser. B Stat. Methodol. 72 417–473. doi=10.1111/j.14679868.2010.00740.x, issn=13697412, mr=2758523 imsref
 (11) [mr] Ravikumar, PradeepP., Wainwright, Martin J.M. J., Raskutti, GarveshG. Yu, BinB. (2011). Highdimensional covariance estimation by minimizing penalized logdeterminant divergence. Electron. J. Stat. 5 935–980. doi=10.1214/11EJS631, issn=19357524, mr=2836766 imsref
 (12) [mr] Rockafellar, R. TyrrellR. T. (1997). Convex Analysis. Princeton Univ. Press, Princeton, NJ. mr=1451876 check yearimsref

(13)
[auto:STB—2012/05/30—10:51:56] Toh, K. C.K. C., Todd, M. J.M. J. Tutuncu, R. H.R. H. SDPT3—A MATLAB software package for semidefinitequadraticlinear programming. Available at
http:// www.math.nus.edu.sg/~mattohkc/sdpt3.html. imsref  (14) [mr] Valiant, Leslie G.L. G. (1977). Graphtheoretic arguments in lowlevel complexity. In Mathematical Foundations of Computer Science (Proc. Sixth Sympos., Tatranská Lomnica, 1977). Lecture Notes in Comput. Sci. 53 162–176. Springer, Berlin. mr=0660702 imsref
 (15) [mr] Wang, ChengjingC., Sun, DefengD. Toh, KimChuanK.C. (2010). Solving logdeterminant optimization problems by a NewtonCG primal proximal point algorithm. SIAM J. Optim. 20 2994–3013. doi=10.1137/090772514, issn=10526234, mr=2735941 imsref
 (16) [mr] Zhao, PengP. Yu, BinB. (2006). On model selection consistency of Lasso. J. Mach. Learn. Res. 7 2541–2563. issn=15324435, mr=2274449 imsref
Comments
There are no comments yet.