skip to main content
US FlagAn official website of the United States government
dot gov icon
Official websites use .gov
A .gov website belongs to an official government organization in the United States.
https lock icon
Secure .gov websites use HTTPS
A lock ( lock ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.
Attention:The NSF Public Access Repository (PAR) system and access will be unavailable from 11:00 PM ET on Thursday, May 14 until 2:00 AM ET on Friday, May 15 due to maintenance. We apologize for the inconvenience.


Title: FRAPPE: fast rank approximation with explainable features for tensors
Abstract Tensor decompositions have proven to be effective in analyzing the structure of multidimensional data. However, most of these methods require a key parameter: the number of desired components. In the case of the CANDECOMP/PARAFAC decomposition (CPD), the ideal value for the number of components is known as the canonical rank and greatly affects the quality of the decomposition results. Existing methods use heuristics or Bayesian methods to estimate this value by repeatedly calculating the CPD, making them extremely computationally expensive. In this work, we proposeFRAPPE, the first method to estimate the canonical rank of a tensor without having to compute the CPD. This method is the result of two key ideas. First, it is much cheaper to generate synthetic data with known rank compared to computing the CPD. Second, we can greatly improve the generalization ability and speed of our model by generating synthetic data that matches a given input tensor in terms of size and sparsity. We can then train a specialized single-use regression model on a synthetic set of tensors engineered to match a given input tensor and use that to estimate the canonical rank of the tensor—all without computing the expensive CPD.FRAPPEis over$$24\times $$ 24 × faster than the best-performing baseline, and exhibits a$$10\%$$ 10 % improvement in MAPE on a synthetic dataset. It also performs as well as or better than the baselines on real-world datasets.  more » « less
Award ID(s):
2112650
PAR ID:
10591465
Author(s) / Creator(s):
;
Publisher / Repository:
Springer Nature
Date Published:
Journal Name:
Data Mining and Knowledge Discovery
Volume:
38
Issue:
6
ISSN:
1384-5810
Page Range / eLocation ID:
4217 to 4232
Subject(s) / Keyword(s):
Rank estimation, tensor decomposition, self-supervised, CANDECOMP/PARAFAC
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract The electricE1 and magneticM1 dipole responses of the$$N=Z$$ N = Z nucleus$$^{24}$$ 24 Mg were investigated in an inelastic photon scattering experiment. The 13.0 MeV electrons, which were used to produce the unpolarised bremsstrahlung in the entrance channel of the$$^{24}$$ 24 Mg($$\gamma ,\gamma ^{\prime }$$ γ , γ ) reaction, were delivered by the ELBE accelerator of the Helmholtz-Zentrum Dresden-Rossendorf. The collimated bremsstrahlung photons excited one$$J^{\pi }=1^-$$ J π = 1 - , four$$J^{\pi }=1^+$$ J π = 1 + , and six$$J^{\pi }=2^+$$ J π = 2 + states in$$^{24}$$ 24 Mg. De-excitation$$\gamma $$ γ rays were detected using the four high-purity germanium detectors of the$$\gamma $$ γ ELBE setup, which is dedicated to nuclear resonance fluorescence experiments. In the energy region up to 13.0 MeV a total$$B(M1)\uparrow = 2.7(3)~\mu _N^2$$ B ( M 1 ) = 2.7 ( 3 ) μ N 2 is observed, but this$$N=Z$$ N = Z nucleus exhibits only marginalE1 strength of less than$$\sum B(E1)\uparrow \le 0.61 \times 10^{-3}$$ B ( E 1 ) 0.61 × 10 - 3  e$$^2 \, $$ 2 fm$$^2$$ 2 . The$$B(\varPi 1, 1^{\pi }_i \rightarrow 2^+_1)/B(\varPi 1, 1^{\pi }_i \rightarrow 0^+_{gs})$$ B ( Π 1 , 1 i π 2 1 + ) / B ( Π 1 , 1 i π 0 gs + ) branching ratios in combination with the expected results from the Alaga rules demonstrate thatKis a good approximative quantum number for$$^{24}$$ 24 Mg. The use of the known$$\rho ^2(E0, 0^+_2 \rightarrow 0^+_{gs})$$ ρ 2 ( E 0 , 0 2 + 0 gs + ) strength and the measured$$B(M1, 1^+ \rightarrow 0^+_2)/B(M1, 1^+ \rightarrow 0^+_{gs})$$ B ( M 1 , 1 + 0 2 + ) / B ( M 1 , 1 + 0 gs + ) branching ratio of the 10.712 MeV$$1^+$$ 1 + level allows, in a two-state mixing model, an extraction of the difference$$\varDelta \beta _2^2$$ Δ β 2 2 between the prolate ground-state structure and shape-coexisting superdeformed structure built upon the 6432-keV$$0^+_2$$ 0 2 + level. 
    more » « less
  2. Abstract Low-lying states in$$^{54}$$ 54 Cr have been investigated via the$$\alpha $$ α -transfer reaction$$^{50}$$ 50 Ti($$^{7}$$ 7 Li,t) at a bombarding energy of 20 MeV. The exclusive$$\alpha $$ α -transfer channel is separated from other reaction channels through the appropriate energy gate on the complementary particle, triton. Levels of$$^{54}$$ 54 Cr populated exclusively by the$$\alpha $$ α -transfer process could be identified up to$$\approx $$ 5 MeV excitation energy and angular momentum up to$$(8)^{+}$$ ( 8 ) + , by identifying the corresponding known$$\gamma $$ γ -rays. These include multiple low-lying non-yrast 2$$^+$$ + and 4$$^+$$ + states, which would otherwise be unfavorable via fusion evaporation reactions. The feeding-subtracted$$\gamma $$ γ -ray yields have been extracted to estimate the population of various excited states through the transfer process. The measured integrated transfer cross sections for all the observed yrast and non-yrast states are compared with Coupled Channels calculations usingfrescoto extract the$$\alpha $$ α +$$^{50}$$ 50 Ti core spectroscopic factors. For the yrast states, a higher$$\alpha $$ α +core overlap is seen for the$$2^+$$ 2 + and$$4^+$$ 4 + states, while it is seen to be less favorable for the$$6^+$$ 6 + and$$(8)^+$$ ( 8 ) + states when$$\alpha $$ α -transfer is considered to occur predominantly as a direct one-step process to the$$^{50}$$ 50 Ti core ground state. The yrast$$2^+$$ 2 + , and$$4^+$$ 4 + states are predominantly populated by single-step transfer, while for the states with spin$$\ge $$ 5, the possibility of core excitation followed by$$\alpha $$ α -transfer shows a larger$$\alpha $$ α -core overlap. For the non-yrast$$0^+$$ 0 + ,$$2^+$$ 2 + , and$$4^+$$ 4 + states, single-step transfer shows moderate to small$$\alpha $$ α -core overlap. No higher spin non-yrast states are observed. 
    more » « less
  3. Abstract The$$^{90}$$ 90 Zr(p,$$\gamma $$ γ )$$^{91}$$ 91 Nb reaction is one of the important reactions in the$$A\approx 90$$ A 90 mass region and part of the nucleosynthesis path responsible for production of$$^{92}$$ 92 Mo during the$$\gamma $$ γ -process. Discrepant data in the literature provide a cross section that varies up to 30% within the Gamow window for the$$^{90}$$ 90 Zr(p,$$\gamma $$ γ )$$^{91}$$ 91 Nb reaction. Thus, the cross section measurements of$$^{90}$$ 90 Zr(p,$$\gamma $$ γ )$$^{91}$$ 91 Nb reaction were revisited using the$$\gamma $$ γ -summing technique. The results are consistent with the lower-value cross sections found in the literature. Based on the new data an updated reaction rate for$$^{90}$$ 90 Zr(p,$$\gamma $$ γ )$$^{91}$$ 91 Nb is provided that is up to 20% higher than that obtained from thenon-smokercode. 
    more » « less
  4. Abstract For a smooth projective varietyXover an algebraic number fieldka conjecture of Bloch and Beilinson predicts that the kernel of the Albanese map ofXis a torsion group. In this article we consider a product$$X=C_1\times \cdots \times C_d$$ X = C 1 × × C d of smooth projective curves and show that if the conjecture is true for any subproduct of two curves, then it is true forX. For a product$$X=C_1\times C_2$$ X = C 1 × C 2 of two curves over$$\mathbb {Q} $$ Q with positive genus we construct many nontrivial examples that satisfy the weaker property that the image of the natural map$$J_1(\mathbb {Q})\otimes J_2(\mathbb {Q})\xrightarrow {\varepsilon }{{\,\textrm{CH}\,}}_0(C_1\times C_2)$$ J 1 ( Q ) J 2 ( Q ) ε CH 0 ( C 1 × C 2 ) is finite, where$$J_i$$ J i is the Jacobian variety of$$C_i$$ C i . Our constructions include many new examples of non-isogenous pairs of elliptic curves$$E_1, E_2$$ E 1 , E 2 with positive rank, including the first known examples of rank greater than 1. Combining these constructions with our previous result, we obtain infinitely many nontrivial products$$X=C_1\times \cdots \times C_d$$ X = C 1 × × C d for which the analogous map$$\varepsilon $$ ε has finite image. 
    more » « less
  5. We employ matrix product states (MPS) and tensor networks to study topological properties of the space of ground states of gapped many-body systems. We focus on families of states in one spatial dimension, where each state can be represented as an injective MPS of finite bond dimension. Such states are short-range entangled ground states of gapped local Hamiltonians. To such parametrized families overX X we associate a gerbe, which generalizes the line bundle of ground states in zero-dimensional families ( in few-body quantum mechanics). The nontriviality of the gerbe is measured by a class inH^3(X, \Z), which is believed to classify one-dimensional parametrized systems. We show that when the gerbe is nontrivial, there is an obstruction to representing the family of ground states with an MPS tensor that is continuous everywhere onX X . We illustrate our construction with two examples of nontrivial parametrized systems overX=S^3 X = S 3 andX = \R P^2 × S^1. Finally, we sketch using tensor network methods how the construction extends to higher dimensional parametrized systems, with an example of a two-dimensional parametrized system that gives rise to a nontrivial 2-gerbe overX = S^4 X = S 4
    more » « less