arXiv subjects

Susama Agarwala

Publications and source records attributed to Susama Agarwala.

At least 19 recordsLinked to original sources

arXiv2025

Computing positroid cells in the Grassmannian of lines, their boundaries and their intersections

Positroids are families of matroids introduced by Postnikov in the study of non-negative Grassmannians. In particular, positroids enumerate a CW decomposition of the totally non-negative Grassmannian. Furthermore, Postnikov has identified several families of combinatorial objects in bijections with positroids. We will provide yet another characterization of positroids for Gr$_{\geq 0}(2,n)$, the Grassmannians of lines, in terms of certain graphs. We use this characterization to compute the dimension and the boundary of positroid cells. This also leads to a combinatorial description of the intersection of positroid cells, that is easily computable. Our techniques rely on determining different ways to enlarge a given collection of subsets of $\{1,\ldots,n\}$ to represent the dependent sets of a positroid, that is the dependencies among the columns of a matrix with non-negative maximal minors. Furthermore, we provide an algorithm to compute all the maximal positroids contained in a set.

math.CO↗

arXiv2024

Rado matroids and a graphical calculus for boundaries of Wilson loop diagrams

We study the boundaries of the positroid cells which arise from N = 4 super Yang Mills theory. Our main tool is a new diagrammatic object which generalizes the Wilson loop diagrams used to represent interactions in the theory. We prove conditions under which these new generalized Wilson loop diagrams correspond to positroids and give an explicit algorithm to calculate the Grassmann necklace of said positroids. Then we develop a graphical calculus operating directly on noncrossing generalized Wilson loop diagrams. In this paradigm, applying diagrammatic moves to a generalized Wilson loop diagram results in new diagrams that represent boundaries of its associated positroid, without passing through cryptomorphisms. We provide a Python implementation of the graphical calculus and use it to show that the boundaries of positroids associated to ordinary Wilson loop diagram are generated by our diagrammatic moves in certain cases.

math-ph↗

arXiv2022

Structural Similarity for Improved Transfer in Reinforcement Learning

Transfer learning is an increasingly common approach for developing performant RL agents. However, it is not well understood how to define the relationship between the source and target tasks, and how this relationship contributes to successful transfer. We present an algorithm called Structural Similarity for Two MDPS, or SS2, that calculates a state similarity measure for states in two finite MDPs based on previously developed bisimulation metrics, and show that the measure satisfies properties of a distance metric. Then, through empirical results with GridWorld navigation tasks, we provide evidence that the distance measure can be used to improve transfer performance for Q-Learning agents over previous implementations.

cs.LG↗

arXiv2022

Geometric instability of out of distribution data across autoencoder architecture

We study the map learned by a family of autoencoders trained on MNIST, and evaluated on ten different data sets created by the random selection of pixel values according to ten different distributions. Specifically, we study the eigenvalues of the Jacobians defined by the weight matrices of the autoencoder at each training and evaluation point. For high enough latent dimension, we find that each autoencoder reconstructs all the evaluation data sets as similar \emph{generalized characters}, but that this reconstructed \emph{generalized character} changes across autoencoder. Eigenvalue analysis shows that even when the reconstructed image appears to be an MNIST character for all out of distribution data sets, not all have latent representations that are close to the latent representation of MNIST characters. All told, the eigenvalue analysis demonstrated a great deal of geometric instability of the autoencoder both as a function on out of distribution inputs, and across architectures on the same set of inputs.

cs.LG↗

arXiv2022

Eigenvalues of Autoencoders in Training and at Initialization

In this paper, we investigate the evolution of autoencoders near their initialization. In particular, we study the distribution of the eigenvalues of the Jacobian matrices of autoencoders early in the training process, training on the MNIST data set. We find that autoencoders that have not been trained have eigenvalue distributions that are qualitatively different from those which have been trained for a long time ($>$100 epochs). Additionally, we find that even at early epochs, these eigenvalue distributions rapidly become qualitatively similar to those of the fully trained autoencoders. We also compare the eigenvalues at initialization to pertinent theoretical work on the eigenvalues of random matrices and the products of such matrices.

cs.LG↗

arXiv2021

A Geometric Chung Lu model and the Drosophila Medulla connectome

Many real world graphs have edges correlated to the distance between them, but, in an inhomogeneous manner. While the Chung-Lu model and the geometric random graph models both are elegant in their simplicity, they are insufficient to capture the complexity of these networks. In this paper, we develop a generalized geometric random graph model that preserves many graph theoretic aspects of these real world networks. We test the validity of this model on a graphical representation of the Drosophila Medulla connectome.

math.CO↗

arXiv2021

Geometry and Generalization: Eigenvalues as predictors of where a network will fail to generalize

We study the deformation of the input space by a trained autoencoder via the Jacobians of the trained weight matrices. In doing so, we prove bounds for the mean squared errors for points in the input space, under assumptions regarding the orthogonality of the eigenvectors. We also show that the trace and the product of the eigenvalues of the Jacobian matrices is a good predictor of the MSE on test points. This is a dataset independent means of testing an autoencoder's ability to generalize on new input. Namely, no knowledge of the dataset on which the network was trained is needed, only the parameters of the trained model.

cs.LG↗

arXiv2021

Cancellation of spurious poles in N=4 SYM: physical and geometric

This paper shows that not only do the codimension one spurious poles of $N^k MHV$ tree level diagrams in N=4 SYM theory cancel in the tree level amplitude as expected, but their vanishing loci have a geometric interpretation that is tightly connected to their representation in the positive Grassmannians. In general, given a positroid variety, $Σ$, and a minimal matrix representation of it in terms of independent variable valued matrices, $M_{\mathcal{V}}$, one can define a polynomial, $R({\mathcal{V}})$ that is uniquely defined by the Grassmann necklace, ${\mathcal{I}}$, of the positroid cell. The vanishing locus of $R({\mathcal{V}})$ lies on the boundary of the positive variety $\overlineΣ \setminus Σ$, but not all boundaries intersect the vanishing loci of a factor of $R({\mathcal{V}})$. We use this to show that the codimension one spurious poles of N=4 SYM, represented in twistor space, cancel in the tree level amplitude.

math-ph↗

arXiv2021

Combinatorics of the geometry of Wilson loop diagrams I: equivalence classes via matroids and polytopes

Wilson loop diagrams are an important tool in studying scattering amplitudes of SYM $N=4$ theory and are known by previous work to be associated to positroids. We characterize the conditions under which two Wilson loop diagrams give the same positroid, prove that an important subclass of subdiagrams (exact subdiagrams) correspond to uniform matroids, and enumerate the number of different Wilson loop diagrams that correspond to each positroid cell. We also give a correspondence between those positroids which can arise from Wilson loop diagrams and directions in associahedra.

math-ph↗

arXiv2021

Combinatorics of the geometry of Wilson loop diagrams II: Grassmann necklaces, dimensions, and denominators

Wilson loop diagrams are an important tool in studying scattering amplitudes of SYM $N=4$ theory and are known by previous work to be associated to positroids. In this paper we study the structure of the associated positroids, as well as the structure of the denominator of the integrand defined by each diagram. We give an algorithm to derive the Grassmann necklace of the associated positroid directly from the Wilson loop diagram, and a recursive proof that the dimension of these cells is thrice the number of propagators in the diagram. We also show that the ideal generated by the denominator in the integrand is the radical of the ideal generated by the product of Grassmann necklace minors.

math-ph↗

arXiv2018

Wilson loops in SYM $N=4$ do not parametrize an orientable space

In this paper we explore the geometric space parametrized by (tree level) Wilson loops in SYM $N=4$. We show that, this space can be seen as a vector bundle over a totally non-negative subspace of the Grassmannian, $\mathcal{W}_{k,cn}$. Furthermore, we explicitly show that this bundle is non-orientable in the majority of the cases, and conjecture that it is non-orientable in the remaining situation. Using the combinatorics of the Deodhar decomposition of the Grassmannian, we identify subspaces $Σ(W) \subset \mathcal{W}_{k,n}$ for which the restricted bundle lies outside the positive Grassmannian. Finally, while probing the combinatorics of the Deodhar decomposition, we give a diagrammatic algorithm for reading equations determining each Deodhar component as a semialgebraic set.

math-ph↗

arXiv2018

An algorithm to construct the Le diagram associated to a Grassmann necklace

Le diagrams and Grassmann necklaces both index the collection of positroids in the nonnegative Grassmannian $Gr_{\geq 0}(k,n)$, but they excel at very different tasks: for example, the dimension of a positroid is easily extracted from its Le diagram, while the list of bases of a positroid is far more easily obtained from its Grassmann necklace. Explicit bijections between the two are therefore desirable. An algorithm for turning a Le diagram into a Grassmann necklace already exists; in this note we give the reverse algorithm.

math.CO↗

arXiv2018

A study in $\mathbb{G}_{\mathbb{R}, \geq 0}$: from the geometric case book of Wilson loop diagrams and SYM $N=4$

We study the geometry underlying the Wilson loop diagram approach to calculating scattering amplitudes in the gauge theory of Supersymmetric Yang Mills (SYM) $N=4$. By applying the tools developed to study total positivity in the real Grassmannian, we are able to systematically compute with all Wilson loop diagrams of a given size and find unexpected patterns and relationships between them. We focus on the smallest nontrivial multi-propagator case, consisting of 2 propagators on 6 vertices, and compute the positroid cells associated to each diagram and the homology of the subcomplex they generate in $\mathbb{G}_{\mathbb{R}, \geq 0}$. We also verify in this case the conjecture that the spurious singularities of the volume functional {\em do} all cancel on the codimension 1 boundaries of these cells.

math.CO↗

arXiv2016

Rational Mixed Tate Motivic Graphs

In this paper, we study the combinatorics of a subcomplex of the Bloch-Kriz cycle complex [4] used to construct the category of mixed Tate motives. The algebraic cycles we consider properly contain the subalgebra of cycles that correspond to multiple logarithms (as defined in [12]). We associate an algebra of graphs to our subalgebra of algebraic cycles. We give a purely graphical criterion for admissibilty. We show that sums of bivalent graphs correspond to coboundary elements of the algebraic cycle complex. Finally, we compute the Hodge realization for an infinite family of algebraic cycles represented by sums of graphs that are not describable in the combinatorial language of [12].

math.AG↗

arXiv2015

The geometric $β$-function in curved space-time under operator regularization

In this paper, I compare the generators of the renormalization group flow, or the geometric $β$-functions for dimensional regularization and operator regularization. I then extend the analysis to show that the geometric $β$-function for a scalar field theory on a closed compact Riemannian manifold is defined on the entire manifold. I then extend the analysis to find the generator of the renormalization group flow for a conformal scalar-field theories on the same manifolds. The geometric $β$-function in this case is not defined.

math-ph↗

arXiv2015

Dihedral symmetries of multiple logarithms

This paper finds relationships between multiple logarithms with a dihedral group action on the arguments. I generalize the combinatorics developed in Gangl, Goncharov and Levin's R-deco polygon representation of multiple logarithms to find these relations. By writing multiple logarithms as iterated integrals, my arguments are valid for iterated integrals as over an arbitrary field.

math.NT↗

arXiv2015

Generalizing the Connes Moscovici Hopf algebra to contain all rooted trees

This paper defines a generalization of the Connes-Moscovici Hopf algebra, $\mathcal{H}(1)$ that contains the entire Hopf algebra of rooted trees. A relationship between the former, a much studied object in non-commutative geometry, and the later, a much studied object in perturbative Quantum Field Theory, has been established by Connes and Kreimer. The results of this paper open the door to study the cohomology of the Hopf algebra of rooted trees.

math-ph↗

arXiv2015

Wilson Loop diagrams and Positroids

In this paper, we study a new application of the positive Grassmanian to Wilson loop diagrams (or MHV diagrams) for scattering amplitudes in N=4 Super Yang-Mill theory ($N=4$ SYM). There has been much interest in studying this theory via the positive Grassmanians using BCFW recursion. This is the first attempt to study MHV diagrams for planar Wilson loop calculations (or planar amplitudes) in terms of positive Grassmannians. We codify Wilson loop diagrams completely in terms of matroids. This allows us to apply the combinatorial tools in matroid theory used to identify positroids, (non-negative Grassmannians), to Wilson loop diagrams. In doing so, we find that certain non-planar Wilson loop diagrams define positive Grassmannians. While non-planar diagrams do not have physical meaning, this finding suggests that they may have value as an algebraic tool, and deserve further investigation.

math-ph↗