[
  {
    "id": 0,
    "page": 2,
    "bounding_box": [
      113.38600158691406,
      419.5950012207031,
      505.23199462890625,
      497.9010009765625
    ],
    "latex_content": "\\begin{table}[htpb]\n\\begin{center}\n\\begin{footnotesize}\n\\newcommand{\\tabincell}[2]{\\begin{tabular}{@{}#1@{}}#2\\end{tabular}}\n\\begin{tabular}{c|c|c|c|c}\\hline\nApproach & Communication & Assumptions  & Min signal strength & Strength type \\\\\\hline\nLasso &0 &\\tabincell{c}{Mutual Incoherence \\\\Sparse Eigenvalue}  & $\\sqrt{\\frac{\\log p}{n}}$ & Element-wise\\\\\\hline\nGroup lasso &$\\Ocal(np)$  & \\tabincell{c}{Mutual Incoherence \\\\Sparse Eigenvalue}  &$\\sqrt{\\frac{1}{n}\\rbr{1 + \\frac{\\log p}{m}}}$ &Row-wise \\\\\\hline \nDSML & $\\Ocal(p)$  &\\tabincell{c}{Generalized Coherence \\\\ Restricted Eigenvalue}   &$\\sqrt{\\frac{1}{n}\\rbr{1 + \\frac{\\log p}{m}}} + \\frac{|S|\\log p}{n}$ &Row-wise \\\\\\hline\n\\end{tabular}\n\\small \\caption{\\small Lower bound on coefficients required to\n  ensure support recovery with $p$ variables, $m$ tasks, $n$\n  samples per task and a true support of size $|S|$.}\n\\label{tab:comparison_sparsistency}\n\\end{footnotesize}\n\\end{center}\n\\end{table}",
    "extracted_content": [
      [
        "Approach",
        "Communication",
        "Assumptions",
        "Min signal strength",
        "Strength type"
      ],
      [
        "Lasso",
        "0",
        "Mutual Incoherence\nSparse Eigenvalue",
        "q\nlog p\nn",
        "Element-wise"
      ],
      [
        "Group lasso",
        "O(np)",
        "Mutual Incoherence\nSparse Eigenvalue",
        "q\n1 \u00001 + log p\u0001\nn m",
        "Row-wise"
      ],
      [
        "DSML",
        "O(p)",
        "Generalized Coherence\nRestricted Eigenvalue",
        "q\nn1 \u00001 + lo mg p\u0001+ |S| nlog p",
        "Row-wise"
      ]
    ],
    "similarity_score": 0.6642335766423357,
    "table_image": "images/1510.00633v1/table_0.png",
    "page_image": "pages/1510.00633v1/page_2.png"
  },
  {
    "id": 1,
    "page": 3,
    "bounding_box": [
      116.75,
      95.239990234375,
      495.25,
      162.74700927734375
    ],
    "latex_content": "\\begin{table}[htpb]\n\\begin{footnotesize}\n\\newcommand{\\tabincell}[2]{\\begin{tabular}{@{}#1@{}}#2\\end{tabular}}\n\\begin{center}\n\\begin{tabular}{c|c|c|c}\\hline\nApproach  & Assumptions & $\\ell_1/\\ell_2$ estimation error & Prediction error \\\\\\hline\nLasso  &Restricted Eigenvalue & $ \\sqrt{\\frac{|S|^2 \\log p}{n}}$ & $\\frac{|S| \\log p}{n}$ \\\\\\hline\nGroup lasso   & Restricted Eigenvalue & $\\frac{|S|}{\\sqrt{n}} \\sqrt{1 + \\frac{\\log p}{m}}$ &$\\frac{|S|}{n} \\rbr{1 + \\frac{\\log p}{m}}$ \\\\\\hline \nDSML  &\\tabincell{c}{Generalized Coherence \\\\ Restricted Eigenvalue}  & $\\frac{|S|}{\\sqrt{n}} \\sqrt{1 + \\frac{\\log p}{m}} + \\frac{|S|^2 \\log p}{n}$ &$\\frac{|S|}{n} \\rbr{1 + \\frac{\\log p}{m}} + \\frac{|S|^3 (\\log p)^2}{n^2}$ \\\\\\hline\n\\end{tabular}\n\\end{center}\n\\small \\caption{\\small Comparison of parameter estimation errors and\n  prediction errors. The DSML guarantees improve over Lasso and have\n  the same leading term as the Group lasso as long as $m<n/(|S|^2\\log\n  p)$.}\n\\label{tab:comparison_estimation}\n\\end{footnotesize}\n\\end{table}",
    "extracted_content": [
      [
        "Approach",
        "Assumptions",
        "\u2113 1/\u2113 estimation error\n2",
        "Prediction error"
      ],
      [
        "Lasso",
        "Restricted Eigenvalue",
        "q\n|S|2 log p\nn",
        "|S| log p\nn"
      ],
      [
        "Group lasso",
        "Restricted Eigenvalue",
        "q\n\u221a|S n| 1 + lo mg p",
        "|S|\u00001 + log p\u0001\nn m"
      ],
      [
        "DSML",
        "Generalized Coherence\nRestricted Eigenvalue",
        "\u221a|S n|q 1 + lo mg p + |S|2 log p\nn",
        "|S n|\u00001 + lo mg p\u0001+ |S|3( nlo 2g p)2"
      ]
    ],
    "similarity_score": 0.3971631205673759,
    "table_image": "images/1510.00633v1/table_1.png",
    "page_image": "pages/1510.00633v1/page_3.png"
  }
]