---
title: "What's New in CVXR 1.8"
author: "Anqi Fu, Balasubramanian Narasimhan, and Stephen Boyd"
date: "`r Sys.Date()`"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{What's New in CVXR 1.8}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
```

## Overview

CVXR 1.8 is a ground-up rewrite using R's
[S7](https://rconsortium.github.io/S7/) object system, designed to
mirror [CVXPY 1.8](https://www.cvxpy.org/) for long-term
maintainability. It is approximately 4--5x faster than the previous
S4-based release and adds many new features.

This vignette summarizes the key changes from CVXR 1.x that may
affect existing users. For the introductory tutorial, see
`vignette("cvxr_intro")`. For worked examples, visit the
[CVXR website](https://cvxr.rbind.io).

## New Solve Interface

The primary solve function is now `psolve()`, which returns the
optimal value directly. Variable values are extracted with `value()`
and problem status with `status()`:

```{r}
library(CVXR)
x <- Variable(2, name = "x")
prob <- Problem(Minimize(sum_squares(x)), list(x >= 1))
opt_val <- psolve(prob, solver = "CLARABEL")
opt_val
value(x)
status(prob)
```

The old `solve()` still works and returns a backward-compatible list:

```{r}
result <- solve(prob, solver = "CLARABEL")
result$value
result$status
```

## API Changes

| Old (CVXR 1.x) | New (CVXR 1.8) |
|-----------------|----------------|
| `solve(problem)` | `psolve(problem)` |
| `result$getValue(x)` | `value(x)` |
| `result$value` | return value of `psolve()` |
| `result$status` | `status(problem)` |
| `result$getDualValue(con)` | `dual_value(con)` |
| `get_problem_data(prob, solver)` | `problem_data(prob, solver)` |

Old function names still work but emit once-per-session deprecation
warnings.

## 13 Solvers

CVXR 1.8 ships with four built-in solvers and supports nine additional
solvers via optional packages:

```{r}
installed_solvers()
```

| Solver | R Package | Problem Classes |
|--------|-----------|-----------------|
| CLARABEL | `clarabel` | LP, QP, SOCP, SDP, ExpCone, PowCone |
| SCS | `scs` | LP, QP, SOCP, SDP, ExpCone, PowCone |
| OSQP | `osqp` | LP, QP |
| HIGHS | `highs` | LP, QP, MILP |
| MOSEK | `Rmosek` | LP, QP, SOCP, SDP, ExpCone, PowCone |
| GUROBI | `gurobi` | LP, QP, SOCP, MIP |
| ECOS / ECOS_BB | `ECOSolveR` | LP, SOCP, ExpCone (+ MIP) |
| GLPK / GLPK_MI | `Rglpk` | LP (+ MILP) |
| CPLEX | `Rcplex` | LP, QP, MIP |
| CVXOPT | `cccp` | LP, SOCP, SDP |
| PIQP | `piqp` | LP, QP |

CLARABEL is the default solver and handles the widest range of problem
types among the built-in solvers. You can specify a solver explicitly:

```{r, eval = FALSE}
psolve(prob, solver = "SCS")
```

### Solver exclusion

You can temporarily exclude solvers from automatic selection without
uninstalling them:

```{r}
available_solvers()
exclude_solvers("SCS")
available_solvers()
include_solvers("SCS")
```

## Disciplined Programming Extensions

### DGP (Geometric Programming)

Geometric programs are solved by converting to convex form via
log-log transformation:

```{r}
x <- Variable(pos = TRUE, name = "x")
y <- Variable(pos = TRUE, name = "y")
obj <- Minimize(x * y)
constr <- list(x * y >= 40, x <= 20, y >= 2)
prob <- Problem(obj, constr)
cat("Is DGP:", is_dgp(prob), "\n")
opt_val <- psolve(prob, gp = TRUE, solver = "CLARABEL")
cat("Optimal:", opt_val, " x =", value(x), " y =", value(y), "\n")
```

### DQCP (Quasiconvex Programming)

Quasiconvex problems are solved via bisection:

```{r}
x <- Variable(name = "x")
prob <- Problem(Minimize(ceil_expr(x)), list(x >= 0.7, x <= 1.5))
cat("Is DQCP:", is_dqcp(prob), "\n")
opt_val <- psolve(prob, qcp = TRUE, solver = "CLARABEL")
cat("Optimal:", opt_val, " x =", value(x), "\n")
```

### DPP (Parameterized Programming)

Parameters allow efficient re-solving when only data changes:

```{r}
n <- 5
x <- Variable(n, name = "x")
lam <- Parameter(nonneg = TRUE, name = "lambda", value = 1.0)

set.seed(42)
A <- matrix(rnorm(10 * n), 10, n)
b <- rnorm(10)

prob <- Problem(Minimize(sum_squares(A %*% x - b) + lam * p_norm(x, 1)))
cat("Is DPP:", is_dpp(prob), "\n")
psolve(prob, solver = "CLARABEL")
value(x)
```

Changing the parameter and re-solving reuses the cached compilation:

```{r}
value(lam) <- 10.0
psolve(prob, solver = "CLARABEL")
value(x)
```

## Mixed-Integer Programming

Variables can be declared boolean or integer:

```{r}
x <- Variable(3, integer = TRUE, name = "x")
prob <- Problem(Minimize(sum(x)), list(x >= 0.5, x <= 3.5))
psolve(prob, solver = "HIGHS")
value(x)
```

## Complex Variables

CVXR supports complex-valued optimization:

```{r}
z <- Variable(2, complex = TRUE, name = "z")
prob <- Problem(Minimize(p_norm(z, 2)), list(z[1] == 1 + 2i))
psolve(prob, solver = "CLARABEL")
value(z)
```

## Boolean Logic

For mixed-integer formulations, boolean logic atoms are available:

```{r}
b1 <- Variable(boolean = TRUE, name = "b1")
b2 <- Variable(boolean = TRUE, name = "b2")
## implies() returns an Expression; compare with == 1 to make a constraint
prob <- Problem(Maximize(b1 + b2),
                list(implies(b1, b2) == 1, b1 + b2 <= 1))
psolve(prob, solver = "HIGHS")
cat("b1 =", value(b1), " b2 =", value(b2), "\n")
```

## Decomposed Solve API

For bootstrapping or simulation, you can compile once and solve many
times:

```{r, eval = FALSE}
pd <- problem_data(prob, solver = "CLARABEL")
chain <- pd$chain

# In a loop:
raw <- solve_via_data(chain, warm_start = FALSE, verbose = FALSE)
result <- problem_unpack_results(prob, raw$solution, chain, raw$inverse_data)
```

## Visualization

`visualize()` provides problem introspection in text or interactive
HTML:

```{r}
x <- Variable(3, name = "x")
prob <- Problem(Minimize(p_norm(x, 1) + sum_squares(x)),
                list(x >= -1, sum(x) == 1))
visualize(prob)
```

For interactive HTML output with collapsible expression trees, DCP
analysis, and curvature coloring:

```{r, eval = FALSE}
visualize(prob, html = "my_problem.html")
```

## New Atoms

### Convenience atoms

| Function | Description |
|----------|-------------|
| `ptp(x)` | Peak-to-peak (range): `max(x) - min(x)` |
| `cvxr_mean(x)` | Arithmetic mean along an axis |
| `cvxr_std(x)` | Standard deviation |
| `cvxr_var(x)` | Variance |
| `vdot(x, y)` | Vector dot product |
| `cvxr_outer(x, y)` | Outer product |
| `inv_prod(x)` | Reciprocal of product |
| `loggamma(x)` | Log of gamma function |
| `log_normcdf(x)` | Log of standard normal CDF |
| `cummax_expr(x)` | Cumulative maximum |
| `dotsort(X, W)` | Weighted sorted dot product |

### Math function dispatch

Standard R math functions work directly on CVXR expressions:

```{r}
x <- Variable(3, name = "x")
abs(x)     # elementwise absolute value
sqrt(x)    # elementwise square root
sum(x)     # sum of entries
max(x)     # maximum entry
norm(x, "2")  # Euclidean norm
```

### Boolean logic atoms

`Not()`, `And()`, `Or()`, `Xor()`, `implies()`, `iff()` — for
mixed-integer programming with boolean variables.

### Other new atoms

- `perspective(f, s)` for perspective functions
- `FiniteSet(expr, values)` constraint for discrete optimization
- `ceil_expr()`, `floor_expr()` for DQCP problems
- `condition_number()`, `gen_lambda_max()`, `dist_ratio()` for DQCP

## Migration Guide

To migrate code from CVXR 1.x to 1.8:

1. Replace `result <- solve(problem)` with `opt_val <- psolve(problem)`
2. Replace `result$getValue(x)` with `value(x)`
3. Replace `result$status` with `status(problem)`
4. Replace `result$getDualValue(con)` with `dual_value(con)`
5. Replace `axis = NA` with `axis = NULL` (axis values 1 and 2 are unchanged)
6. Update solver preferences: the default is now CLARABEL (was ECOS)
7. Wrap Matrix package objects with `as_cvxr_expr()` before using them
   with CVXR operators (e.g., `as_cvxr_expr(A_sparse) %*% x` instead
   of `A_sparse %*% x`). This preserves sparsity --- unlike
   `as.matrix()`, which densifies. Base R `matrix` and `numeric`
   objects work natively without wrapping.
8. **Dimension-preserving operations.** CVXR 1.8 preserves 2D shapes
   throughout, matching CVXPY. In particular, axis reductions like
   `sum_entries(X, axis = 2)` now return a proper row vector of shape
   `(1, n)` rather than collapsing to a 1D vector.
   When comparing such a result with an R numeric vector (which CVXR
   treats as a column vector), you may need to use `t()` or
   `matrix(..., nrow = 1)` to match shapes. For example:
   ```r
   ## Old (worked in CVXR 1.x because axis reductions were 1D):
   sum_entries(X, axis = 2) == target_vec
   ## New (wrap target as row vector to match the (1, n) shape):
   sum_entries(X, axis = 2) == t(target_vec)
   ```
   Similarly, if you extract a scalar from a CVXR result and need a
   plain numeric value, use `as.numeric()` to drop the matrix
   dimensions.

All old function names continue to work with deprecation warnings.

## Further Reading

- [CVXR website](https://cvxr.rbind.io) — worked examples
- [Package reference](https://www.cvxgrp.org/CVXR/) — full API documentation
- [CVXPY documentation](https://www.cvxpy.org/) — mathematical framework
- Fu, Narasimhan, and Boyd (2020). "CVXR: An R Package for Disciplined
  Convex Optimization." _Journal of Statistical Software_, 94(14).
  doi:10.18637/jss.v094.i14