Mihailo Stojnic

Agostino Capponi, Mihailo Stojnic

In this paper, we revisit and further explore a mathematically rigorous connection between Causal inference (C-inf) and the Low-rank recovery (LRR) established in [10]. Leveraging the Random duality - Free probability theory (RDT-FPT) connection, we obtain the exact explicit typical C-inf asymmetric phase transitions (PT). We uncover a doubling low-rankness phenomenon, which means that exactly two times larger low rankness is allowed in asymmetric scenarios compared to the symmetric worst case ones considered in [10]. Consequently, the final PT mathematical expressions are as elegant as those obtained in [10], and highlight direct relations between the targeted C-inf matrix low rankness and the time of treatment. Our results have strong implications for applications, where C-inf matrices are not necessarily symmetric.

Agostino Capponi, Mihailo Stojnic

In this paper we establish a mathematically rigorous connection between Causal inference (C-inf) and the low-rank recovery (LRR). Using Random Duality Theory (RDT) concepts developed in [46,48,50] and novel mathematical strategies related to free probability theory, we obtain the exact explicit typical (and achievable) worst case phase transitions (PT). These PT precisely separate scenarios where causal inference via LRR is possible from those where it is not. We supplement our mathematical analysis with numerical experiments that confirm the theoretical predictions of PT phenomena, and further show that the two closely match for fairly small sample sizes. We obtain simple closed form representations for the resulting PTs, which highlight direct relations between the low rankness of the target C-inf matrix and the time of the treatment. Hence, our results can be used to determine the range of C-inf's typical applicability.

Mihailo Stojnic

Recent studies suggest that asymmetric binary perceptron (ABP) likely exhibits the so-called statistical-computational gap characterized with the appearance of two phase transitioning constraint density thresholds: \textbf{\emph{(i)}} the \emph{satisfiability threshold} $α_c$, below/above which ABP succeeds/fails to operate as a storage memory; and \textbf{\emph{(ii)}} \emph{algorithmic threshold} $α_a$, below/above which one can/cannot efficiently determine ABP's weight so that it operates as a storage memory. We consider a particular parametric utilization of \emph{fully lifted random duality theory} (fl RDT) [85] and study its potential ABP's algorithmic implications. A remarkable structural parametric change is uncovered as one progresses through fl RDT lifting levels. On the first two levels, the so-called $\c$ sequence -- a key parametric fl RDT component -- is of the (natural) decreasing type. A change of such phenomenology on higher levels is then connected to the $α_c$ -- $α_a$ threshold change. Namely, on the second level concrete numerical values give for the critical constraint density $α=α_c\approx 0.8331$. While progressing through higher levels decreases this estimate, already on the fifth level we observe a satisfactory level of convergence and obtain $α\approx 0.7764$. This allows to draw two striking parallels: \textbf{\emph{(i)}} the obtained constraint density estimate is in a remarkable agrement with range $α\in (0.77,0.78)$ of clustering defragmentation (believed to be responsible for failure of locally improving algorithms) [17,88]; and \textbf{\emph{(ii)}} the observed change of $\c$ sequence phenomenology closely matches the one of the negative Hopfield model for which the existence of efficient algorithms that closely approach similar type of threshold has been demonstrated recently [87].

Mihailo Stojnic

We consider the classical \emph{spherical} perceptrons and study their capacities. The famous zero-threshold case was solved in the sixties of the last century (see, \cite{Wendel62,Winder,Cover65}) through the high-dimensional combinatorial considerations. The general threshold, $κ$, case though turned out to be much harder and stayed out of reach for the following several decades. A substantial progress was then made in \cite{SchTir02} and \cite{StojnicGardGen13} where the \emph{positive} threshold ($κ\geq 0$) scenario was finally fully settled. While the negative counterpart ($κ\leq 0$) remained out of reach, \cite{StojnicGardGen13} did show that the random duality theory (RDT) is still powerful enough to provide excellent upper bounds. Moreover, in \cite{StojnicGardSphNeg13}, a \emph{partially lifted} RDT variant was considered and it was shown that the upper bounds of \cite{StojnicGardGen13} can be lowered. After recent breakthroughs in studying bilinearly indexed (bli) random processes in \cite{Stojnicsflgscompyx23,Stojnicnflgscompyx23}, \emph{fully lifted} random duality theory (fl RDT) was developed in \cite{Stojnicflrdt23}. We here first show that the \emph{negative spherical perceptrons} can be fitted into the frame of the fl RDT and then employ the whole fl RDT machinery to characterize the capacity. To be fully practically operational, the fl RDT requires a substantial numerical work. We, however, uncover remarkable closed form analytical relations among key lifting parameters. Such a discovery enables performing the needed numerical calculations to obtain concrete capacity values. We also observe that an excellent convergence (with the relative improvement $\sim 0.1\%$) is achieved already on the third (second non-trivial) level of the \emph{stationarized} full lifting.

Mihailo Stojnic

Recent progress in studying \emph{treelike committee machines} (TCM) neural networks (NN) in \cite{Stojnictcmspnncaprdt23,Stojnictcmspnncapliftedrdt23,Stojnictcmspnncapdiffactrdt23} showed that the Random Duality Theory (RDT) and its a \emph{partially lifted}(pl RDT) variant are powerful tools that can be used for very precise networks capacity analysis. Here, we consider \emph{wide} hidden layer networks and uncover that certain aspects of numerical difficulties faced in \cite{Stojnictcmspnncapdiffactrdt23} miraculously disappear. In particular, we employ recently developed \emph{fully lifted} (fl) RDT to characterize the \emph{wide} ($d\rightarrow \infty$) TCM nets capacity. We obtain explicit, closed form, capacity characterizations for a very generic class of the hidden layer activations. While the utilized approach significantly lowers the amount of the needed numerical evaluations, the ultimate fl RDT usefulness and success still require a solid portion of the residual numerical work. To get the concrete capacity values, we take four very famous activations examples: \emph{\textbf{ReLU}}, \textbf{\emph{quadratic}}, \textbf{\emph{erf}}, and \textbf{\emph{tanh}}. After successfully conducting all the residual numerical work for all of them, we uncover that the whole lifting mechanism exhibits a remarkably rapid convergence with the relative improvements no better than $\sim 0.1\%$ happening already on the 3-rd level of lifting. As a convenient bonus, we also uncover that the capacity characterizations obtained on the first and second level of lifting precisely match those obtained through the statistical physics replica theory methods in \cite{ZavPeh21} for the generic and in \cite{BalMalZech19} for the ReLU activations.

Mihailo Stojnic

We consider the capacity of \emph{treelike committee machines} (TCM) neural networks. Relying on Random Duality Theory (RDT), \cite{Stojnictcmspnncaprdt23} recently introduced a generic framework for their capacity analysis. An upgrade based on the so-called \emph{partially lifted} RDT (pl RDT) was then presented in \cite{Stojnictcmspnncapliftedrdt23}. Both lines of work focused on the networks with the most typical, \emph{sign}, activations. Here, on the other hand, we focus on networks with other, more general, types of activations and show that the frameworks of \cite{Stojnictcmspnncaprdt23,Stojnictcmspnncapliftedrdt23} are sufficiently powerful to enable handling of such scenarios as well. In addition to the standard \emph{linear} activations, we uncover that particularly convenient results can be obtained for two very commonly used activations, namely, the \emph{quadratic} and \emph{rectified linear unit (ReLU)} ones. In more concrete terms, for each of these activations, we obtain both the RDT and pl RDT based memory capacities upper bound characterization for \emph{any} given (even) number of the hidden layer neurons, $d$. In the process, we also uncover the following two, rather remarkable, facts: 1) contrary to the common wisdom, both sets of results show that the bounding capacity decreases for large $d$ (the width of the hidden layer) while converging to a constant value; and 2) the maximum bounding capacity is achieved for the networks with precisely \textbf{\emph{two}} hidden layer neurons! Moreover, the large $d$ converging values are observed to be in excellent agrement with the statistical physics replica theory based predictions.

Mihailo Stojnic

In \cite{Hop82}, Hopfield introduced a \emph{Hebbian} learning rule based neural network model and suggested how it can efficiently operate as an associative memory. Studying random binary patterns, he also uncovered that, if a small fraction of errors is tolerated in the stored patterns retrieval, the capacity of the network (maximal number of memorized patterns, $m$) scales linearly with each pattern's size, $n$. Moreover, he famously predicted $α_c=\lim_{n\rightarrow\infty}\frac{m}{n}\approx 0.14$. We study this very same scenario with two famous pattern's basins of attraction: \textbf{\emph{(i)}} The AGS one from \cite{AmiGutSom85}; and \textbf{\emph{(ii)}} The NLT one from \cite{Newman88,Louk94,Louk94a,Louk97,Tal98}. Relying on the \emph{fully lifted random duality theory} (fl RDT) from \cite{Stojnicflrdt23}, we obtain the following explicit capacity characterizations on the first level of lifting: \begin{equation} α_c^{(AGS,1)} = \left ( \max_{δ\in \left ( 0,\frac{1}{2}\right ) }\frac{1-2δ}{\sqrt{2} \mbox{erfinv} \left ( 1-2δ\right )} - \frac{2}{\sqrt{2π}} e^{-\left ( \mbox{erfinv}\left ( 1-2δ\right )\right )^2}\right )^2 \approx \mathbf{0.137906} \end{equation} \begin{equation} α_c^{(NLT,1)} = \frac{\mbox{erf}(x)^2}{2x^2}-1+\mbox{erf}(x)^2 \approx \mathbf{0.129490}, \quad 1-\mbox{erf}(x)^2- \frac{2\mbox{erf}(x)e^{-x^2}}{\sqrtπx}+\frac{2e^{-2x^2}}π=0. \end{equation} A substantial numerical work gives on the second level of lifting $α_c^{(AGS,2)} \approx \mathbf{0.138186}$ and $α_c^{(NLT,2)} \approx \mathbf{0.12979}$, effectively uncovering a remarkably fast lifting convergence. Moreover, the obtained AGS characterizations exactly match the replica symmetry based ones of \cite{AmiGutSom85} and the corresponding symmetry breaking ones of \cite{SteKuh94}.

Mihailo Stojnic

We consider the memorization capabilities of multilayered \emph{sign} perceptrons neural networks (SPNNs). A recent rigorous upper-bounding capacity characterization, obtained in \cite{Stojnictcmspnncaprdt23} utilizing the Random Duality Theory (RDT), demonstrated that adding neurons in a network configuration may indeed be very beneficial. Moreover, for particular \emph{treelike committee machines} (TCM) architectures with $d\leq 5$ neurons in the hidden layer, \cite{Stojnictcmspnncaprdt23} made a very first mathematically rigorous progress in over 30 years by lowering the previously best known capacity bounds of \cite{MitchDurb89}. Here, we first establish that the RDT bounds from \cite{Stojnictcmspnncaprdt23} scale as $\sim \sqrt{d}$ and can not on their own \emph{universally} (over the entire range of $d$) beat the best known $\sim \log(d)$ scaling of the bounds from \cite{MitchDurb89}. After recognizing that the progress from \cite{Stojnictcmspnncaprdt23} is therefore promising, but yet without a complete concretization, we then proceed by considering the recently developed fully lifted RDT (fl RDT) as an alternative. While the fl RDT is indeed a powerful juggernaut, it typically relies on heavy numerical evaluations. To avoid such heavy numerics, we here focus on a simplified, \emph{partially lifted}, variant and show that it allows for very neat, closed form, analytical capacity characterizations. Moreover, we obtain the concrete capacity bounds that \emph{universally} improve for \emph{any} $d$ over the best known ones of \cite{MitchDurb89}.

Mihailo Stojnic

We consider fully row/column-correlated linear regression models and study several classical estimators (including minimum norm interpolators (GLS), ordinary least squares (LS), and ridge regressors). We show that \emph{Random Duality Theory} (RDT) can be utilized to obtain precise closed form characterizations of all estimators related optimizing quantities of interest, including the \emph{prediction risk} (testing or generalization error). On a qualitative level out results recover the risk's well known non-monotonic (so-called double-descent) behavior as the number of features/sample size ratio increases. On a quantitative level, our closed form results show how the risk explicitly depends on all key model parameters, including the problem dimensions and covariance matrices. Moreover, a special case of our results, obtained when intra-sample (or time-series) correlations are not present, precisely match the corresponding ones obtained via spectral methods in [6,16,17,24].

Mihailo Stojnic

We consider correlated \emph{factor} regression models (FRM) and analyze the performance of classical ridge interpolators. Utilizing powerful \emph{Random Duality Theory} (RDT) mathematical engine, we obtain \emph{precise} closed form characterizations of the underlying optimization problems and all associated optimizing quantities. In particular, we provide \emph{excess prediction risk} characterizations that clearly show the dependence on all key model parameters, covariance matrices, loadings, and dimensions. As a function of the over-parametrization ratio, the generalized least squares (GLS) risk also exhibits the well known \emph{double-descent} (non-monotonic) behavior. Similarly to the classical linear regression models (LRM), we demonstrate that such FRM phenomenon can be smoothened out by the optimally tuned ridge regularization. The theoretical results are supplemented by numerical simulations and an excellent agrement between the two is observed. Moreover, we note that ``ridge smootenhing'' is often of limited effect already for over-parametrization ratios above $5$ and of virtually no effect for those above $10$. This solidifies the notion that one of the recently most popular neural networks paradigms -- \emph{zero-training (interpolating) generalizes well} -- enjoys wider applicability, including the one within the FRM estimation/prediction context.

Mihailo Stojnic

Perceptrons have been known for a long time as a promising tool within the neural networks theory. The analytical treatment for a special class of perceptrons started in seminal work of Gardner \cite{Gar88}. Techniques initially employed to characterize perceptrons relied on a statistical mechanics approach. Many of such predictions obtained in \cite{Gar88} (and in a follow-up \cite{GarDer88}) were later on established rigorously as mathematical facts (see, e.g. \cite{SchTir02,SchTir03,TalBook,StojnicGardGen13,StojnicGardSphNeg13,StojnicGardSphErr13}). These typically related to spherical perceptrons. A lot of work has been done related to various other types of perceptrons. Among the most challenging ones are what we will refer to as the discrete perceptrons. An introductory statistical mechanics treatment of such perceptrons was given in \cite{GutSte90}. Relying on results of \cite{Gar88}, \cite{GutSte90} characterized many of the features of several types of discrete perceptrons. We in this paper, consider a similar subclass of discrete perceptrons and provide a mathematically rigorous set of results related to their performance. As it will turn out, many of the statistical mechanics predictions obtained for discrete predictions will in fact appear as mathematically provable bounds. This will in a way emulate a similar type of behavior we observed in \cite{StojnicGardGen13,StojnicGardSphNeg13,StojnicGardSphErr13} when studying spherical perceptrons.

Mihailo Stojnic

It has been known for a long time that the classical spherical perceptrons can be used as storage memories. Seminal work of Gardner, \cite{Gar88}, started an analytical study of perceptrons storage abilities. Many of the Gardner's predictions obtained through statistical mechanics tools have been rigorously justified. Among the most important ones are of course the storage capacities. The first rigorous confirmations were obtained in \cite{SchTir02,SchTir03} for the storage capacity of the so-called positive spherical perceptron. These were later reestablished in \cite{TalBook} and a bit more recently in \cite{StojnicGardGen13}. In this paper we consider a variant of the spherical perceptron that operates as a storage memory but allows for a certain fraction of errors. In Gardner's original work the statistical mechanics predictions in this directions were presented sa well. Here, through a mathematically rigorous analysis, we confirm that the Gardner's predictions in this direction are in fact provable upper bounds on the true values of the storage capacity. Moreover, we then present a mechanism that can be used to lower these bounds. Numerical results that we present indicate that the Garnder's storage capacity predictions may, in a fairly wide range of parameters, be not that far away from the true values.

Mihailo Stojnic

Under-determined systems of linear equations with sparse solutions have been the subject of an extensive research in last several years above all due to results of \cite{CRT,CanRomTao06,DonohoPol}. In this paper we will consider \emph{noisy} under-determined linear systems. In a breakthrough \cite{CanRomTao06} it was established that in \emph{noisy} systems for any linear level of under-determinedness there is a linear sparsity that can be \emph{approximately} recovered through an SOCP (second order cone programming) optimization algorithm so that the approximate solution vector is (in an $\ell_2$-norm sense) guaranteed to be no further from the sparse unknown vector than a constant times the noise. In our recent work \cite{StojnicGenSocp10} we established an alternative framework that can be used for statistical performance analysis of the SOCP algorithms. To demonstrate how the framework works we then showed in \cite{StojnicGenSocp10} how one can use it to precisely characterize the \emph{generic} (worst-case) performance of the SOCP. In this paper we present a different set of results that can be obtained through the framework of \cite{StojnicGenSocp10}. The results will relate to \emph{problem dependent} performance analysis of SOCP's. We will consider specific types of unknown sparse vectors and characterize the SOCP performance when used for recovery of such vectors. We will also show that our theoretical predictions are in a solid agreement with the results one can get through numerical simulations.