make_table2.py

#!/usr/bin/env python3
"""Regenerate Table2.tex — the per-instance numerical tables for Experiment A.

Input : resultsA_raw.csv written by run_experimentA.m
Output: Table2.tex, a standalone LaTeX document with one landscape table per
        benchmark problem, in the column layout of the previous archive
        (\\#Dim, Ip, then \\#Itr / \\#Fev / \\#Tm / \\#Nrm for each of the four
        solvers).

Differences from the previous archive, all deliberate:
  * seven tables rather than six; Problem 7 was never in the old archive
  * six dimensions rather than five; n = 150000 was never run before
  * the \\multicolumn spans in the header are 4, matching the four data
    columns per solver (the old file said 3)
  * a Status column is not added, but any run that did not meet the common
    stopping criterion is flagged with a dagger and named in the caption
"""
import sys
import os
from collections import OrderedDict

SOLVERS = ["BBSANME", "Modified Algorithm 2.1", "Algorithm 2.1", "DLLA"]
SOLVER_TEX = ["\\texttt{BBSANME}", "Modified Algorithm 2.1",
              "\\texttt{Algorithm 2.1}", "DLLA"]

PROBLEM_NOTE = {
    "P1": "not globally Lipschitz",
    "P4": "not globally Lipschitz",
    "P7": "pseudomonotone, $n=2$",
}


def read_raw(path):
    rows = []
    with open(path) as fh:
        for line in fh:
            if line.startswith("#") or line.startswith("Solver,"):
                continue
            q = line.rstrip("\n").split(",")
            if len(q) < 10:
                continue
            fam, dim, guess = q[1].split("_")
            rows.append(dict(solver=q[0], family=fam,
                             dim=int(dim[3:]), guess=int(guess[5:]),
                             it=int(float(q[3])), fe=int(float(q[4])),
                             tm=float(q[6]), nrm=float(q[7]),
                             status=q[8], ok=bool(int(q[9]))))
    return rows


def sci(x):
    """1.23E-07, matching the previous archive's format."""
    s = "%.2E" % x
    m, e = s.split("E")
    return "%sE%s%02d" % (m, e[0], abs(int(e)))


def table_for(rows, fam):
    sub = [r for r in rows if r["family"] == fam]
    dims = sorted({r["dim"] for r in sub})
    guesses = sorted({r["guess"] for r in sub})
    index = {(r["solver"], r["dim"], r["guess"]): r for r in sub}

    failures = [r for r in sub if not r["ok"]]

    out = []
    out.append("\\begin{sidewaystable}[htbp]")
    cap = ("Numerical results for \\texttt{BBSANME}, Modified Algorithm~2.1, "
           "\\texttt{Algorithm~2.1} and DLLA on Problem~%s" % fam[1:])
    if fam in PROBLEM_NOTE:
        cap += " (%s)" % PROBLEM_NOTE[fam]
    cap += "."
    if failures:
        cap += (" Runs marked $\\dagger$ did not reach $\\|\\mathcal{H}(x_k)\\|"
                "\\le 10^{-6}$.")
    else:
        cap += " Every run reached $\\|\\mathcal{H}(x_k)\\|\\le 10^{-6}$."
    out.append("\t\\caption{%s}" % cap)
    out.append("\t\\label{tab:A-%s}" % fam)
    # Scale to fit, never stretch. The previous archive forced
    # height=7.5cm, width=24cm on every table, which inflated the
    # six-row Problem 7 table until its text was unreadable.
    out.append("\t\\adjustbox{max width=24cm, max totalheight=17cm, "
               "keepaspectratio}")
    out.append("\t{")
    out.append("\t\t\\footnotesize")
    out.append("\t\t\\setlength{\\tabcolsep}{4pt}")
    out.append("\t\t\\begin{tabular}{cc" + "cccc" * 4 + "}")
    out.append("\t\t\t\\hline")
    header = " & " * 2 + " & ".join(
        "\\multicolumn{4}{c}{%s}" % s for s in SOLVER_TEX) + " \\\\"
    out.append("\t\t\t" + header)
    out.append("\t\t\t\\cline{3-6}\\cline{7-10}\\cline{11-14}\\cline{15-18}")
    sub_header = ("\\#Dim & Ip & " +
                  " & ".join(["\\#Itr & \\#Fev & \\#Tm & \\#Nrm"] * 4) +
                  " \\\\")
    out.append("\t\t\t" + sub_header)
    out.append("\t\t\t\\hline")

    for dim in dims:
        mid = len(guesses) // 2
        for i, g in enumerate(guesses):
            first = ("%d" % dim) if i == mid else ""
            cells = [first, "$x_{%d}$" % g]
            for s in SOLVERS:
                r = index.get((s, dim, g))
                if r is None:
                    cells += ["--", "--", "--", "--"]
                    continue
                mark = "$^{\\dagger}$" if not r["ok"] else ""
                cells += ["%d" % r["it"], "%d" % r["fe"],
                          "%.6f" % r["tm"], sci(r["nrm"]) + mark]
            out.append("\t\t\t" + " & ".join(cells) + " \\\\")
        out.append("\t\t\t\\hline")

    out.append("\t\t\\end{tabular}%")
    out.append("\t}")
    out.append("\\end{sidewaystable}")
    out.append("")
    return "\n".join(out)


PREAMBLE = r"""\documentclass[11pt]{article}
\usepackage{amsmath,amsfonts,amssymb}
\usepackage[a4paper, left=0.5in, right=0.5in, top=0.9in, bottom=0.9in]{geometry}
\usepackage{rotating}
\usepackage{adjustbox}
\usepackage{booktabs, multirow}
\usepackage{longtable}
\usepackage{caption}
\usepackage{hyperref}
\usepackage[T1]{fontenc}

\title{Numerical tables for \\
\emph{Barzilai--Borwein-like spectral algorithm with norm descent line search
for monotone operator equations}}
\author{H. Mohammad \and H. Choi \and A. B. Abubakar \and M. Abdullahi
        \and M. S. Sarkinbai}
\date{%(date)s}

\begin{document}
\maketitle

\noindent
This archive holds the per-instance numerical results behind Section~4.1 of
the manuscript. Four solvers are compared on seven benchmark problems:
\texttt{BBSANME}, Modified Algorithm~2.1 (Muangchoo and Abubakar, 2024),
\texttt{Algorithm~2.1} (Feng, Sun and Wang, 2017) and DLLA (Awwal et al.,
2023).

\medskip
\noindent
\textbf{Design.} Problems~1--6 are run at $n \in \{1000, 5000, 10\,000,
50\,000, 100\,000, 150\,000\}$ from the six initial points $x_1,\dots,x_6$ of
Section~4.1; Problem~7 is two-dimensional and uses its own six initial
points. That is $222$ instances and $888$ runs.

\medskip
\noindent
\textbf{Settings.} \texttt{BBSANME} uses $\rho = 0.80$, $a = 10^{-4}$,
$r = 10^{-3}$ and $\eta_k = 1/\exp(k^2)$. The parameters of the three
competing methods are those recommended in their own papers. Every run stops
when $\|\mathcal{H}(x_k)\| \le 10^{-6}$ or after $1000$ iterations.

\medskip
\noindent
\textbf{Reported quantities.} \#Itr is the number of iterations, \#Fev the
number of evaluations of $\mathcal{H}$, \#Tm the CPU time in seconds and
\#Nrm the residual norm at the returned point. \#Fev counts every evaluation:
the initial $\mathcal{H}(x_0)$, every backtracking trial including rejected
ones, and, for the three projection methods, the evaluation at the projection
point. The same convention applies to all four solvers.

\medskip
\noindent
\textbf{Environment.} %(env)s

\clearpage

"""

POSTAMBLE = r"""
\end{document}
"""


def main():
    raw = sys.argv[1] if len(sys.argv) > 1 else "resultsA_raw.csv"
    out = sys.argv[2] if len(sys.argv) > 2 else "Table2.tex"
    env = sys.argv[3] if len(sys.argv) > 3 else "MATLAB, Microsoft Windows."
    date = sys.argv[4] if len(sys.argv) > 4 else "\\today"

    rows = read_raw(raw)
    fams = sorted({r["family"] for r in rows}, key=lambda f: int(f[1:]))
    body = "\n\\clearpage\n\n".join(table_for(rows, f) for f in fams)

    with open(out, "w") as fh:
        fh.write(PREAMBLE % dict(env=env, date=date))
        fh.write(body)
        fh.write(POSTAMBLE)

    nfail = sum(1 for r in rows if not r["ok"])
    print("%s written: %d runs, %d problems, %d not meeting the criterion"
          % (out, len(rows), len(fams), nfail))


if __name__ == "__main__":
    main()