R/176_reductions_dcp2cone_canonicalizers_quad_form_canon.R

In CVXR: Disciplined Convex Optimization

#####
## DO NOT EDIT THIS FILE!! EDIT THE SOURCE INSTEAD: rsrc_tree/reductions/dcp2cone/canonicalizers/quad_form_canon.R
#####

## CVXPY SOURCE: reductions/dcp2cone/canonicalizers/quad_form_canon.py
## Decomposes P via eigenvalues, then uses sum_squares -> quad_over_lin_canon


quad_form_canon <- function(expr, args, solver_context = NULL) {
  result <- decomp_quad(value(args[[2L]]))
  scale <- result$scale
  M1 <- result$M1
  M2 <- result$M2

  ## Special case: P == 0
  if (ncol(M1) == 0L && ncol(M2) == 0L) {
    return(list(Constant(0), list()))
  }

  ## CVXPY v1.8.2 fix: reject indefinite P (both M1 and M2 non-empty).
  ## Previously, the M2 block silently overwrote qexpr from M1, dropping
  ## the positive eigenvalue contribution entirely.
  if (ncol(M1) > 0L && ncol(M2) > 0L) {
    cli_abort(c(
      "{.fn quad_form} canonicalization does not support indefinite matrices.",
      "i" = "P must be positive or negative semidefinite.",
      "i" = "If P is intended to be PSD, check for numerical errors in its construction (e.g., use {.code P <- (P + t(P)) / 2} and clip negative eigenvalues)."
    ))
  }

  if (ncol(M1) > 0L) {
    qexpr <- sum_squares(Constant(t(M1)) %*% args[[1L]])
  } else {
    scale <- -scale
    qexpr <- sum_squares(Constant(t(M2)) %*% args[[1L]])
  }
  qol_result <- quad_over_lin_canon(qexpr, .args(qexpr), solver_context)
  list(scale * qol_result[[1L]], qol_result[[2L]])
}

method(dcp_canonicalize, QuadForm) <- quad_form_canon
method(has_dcp_canon, QuadForm) <- function(expr) TRUE