Edit on GitHub

Release notes

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

v0.16.0 (unreleased)

This release contains a large number of changes, including some breaking changes.

Breaking

  • This release involved a substantial rewrite of Convex.jl to integrate better with MathOptInterface. (#504)
    • x + A will error if x is a scalar variable and A is an array. Instead, use x * ones(size(A)) + A.
    • The RelativeEntropyAtom now returns a scalar value instead o elementwise values. This does not affect the result of relative_entropy.
    • The function constant should be used instead of the type Constant (which now refers to exclusively real constants).
    • The constraint a <= b now produces a - b in Nonpositives() instead of b - a in Nonnegatives(). The primal solutions are equivalent, but the dual variable associated with such constraints is now reversed in sign. (Following the convention in MathOptInterface, the dual of a <= b is always negative, regardless of optimization sense.) (#593)
    • The structs LtConstraint, GtConstraint, EqConstraint SOCConstraint and SDPConstraint have been replaced by GenericConstraint{S} where S<:MOI.AbstractSet (#590)
  • The syntaxes dot(*), dot(/) and dot(^) have been removed in favor of explicit broadcasting (x .* y, x ./ y, and x .^ y). These were (mild) type piracy. In addition, vecdot(x,y) has been removed. Call dot(vec(x), vec(y)) instead. (#524)
  • The function constraints, used to get constraints associated to an individual variable, has been renamed get_constraints (#527)
  • DCP violations now throw a DCPViolationError exception, rather than a warning. Relatedly, Convex.emit_dcp_warnings has been removed (#523)
  • Removed the undocumented an internal function latex_formulation (#551)
  • The strict inequalities > and < have been deprecated. They will be removed in the next breaking release. Note that these never enforced strict inequalities, but instead were equivalent to >= and <= respectively (#555)
  • The functions norm_inf, norm_1, and norm_fro have been deprecated. They will be removed in the next breaking release (#567)

Added

  • SDP, SOC, and exponential cone constraints now have dual values populated (#504)
  • geomean supports more than 2 arguments (#504)
  • Added Convex.Optimizer (#511), (#530), (#534)
  • Added write_to_file (#531)
  • Added entropy_elementwise (#570)
  • norm on AbstractExpr objects now supports matrices (treating them like vectors), matching Base's behavior (#528)

Fixed

  • sumlargesteigs now enforces that it's argument is hermitian. (#504)
  • Type piracy of imag and real has been removed. This should not affect use of Convex. (#504)
  • Bugfix: dot now correctly complex-conjugates its first argument (#524)
  • Add tests and fix a number of bugs in various atoms (#546), (#550), (#554), (#556), (#558), (#559), (#561), (#562), (#563), (#565), (#566), (#567) (#568)

Other

  • Improved the documentation (#517), (#529)
  • Refactored the tests into a functional form (#532)
  • Added test/Project.toml (#536)
  • Refactored imports to explicitly overload methods (#537)
  • Tidied and renamed various atoms and files clarity. This should be non-breaking as no public API was changed. (#538), (#539), (#540), (#541), (#543), (#545)
  • Removed the unused file src/problem_depot/problems/benchmark.jl (#560)

v0.15.4 (October 24, 2023)

  • Convex's piracy of hcat and vcat was made less severe, allowing precompilation of Convex.jl on Julia 1.10.

v0.15.3 (February 11, 2023)

  • Add support for LDLFactorizations v0.10 #496.
  • Replace randn(m, 1) with randn(m) to be more Julian #498.
  • Add support for indexing expressions with CartesianIndex #500.

v0.15.2 (August 10, 2022)

  • Add support for LDLFactorizations v0.9 #493.
  • Fix use of deprecated functions from AbstractTrees #494.

v0.15.1 (March 28, 2022)

  • Use OrderedDict internally for reproducible results, issue: #488, fix: #489.

v0.15.0 (March 2, 2022)

Breaking changes

  • Minimum required version of Julia is now v1.6
  • Updated to MathOptInterface v1.0
    • As a consequence, many previously deprecated solver calls may stop working. For example, instead of () -> SCS.Optimizer(verbose = 0), use MOI.OptimizerWithAttributes(SCS.Optimizer, "verbose" => 0).

v0.14.18 (November 14, 2021)

  • Fix typo in logisticloss for length-1 expressions which caused errors (reported in #458, fixed in #469).

v0.14.17 (November 14, 2021)

  • Updated to become compatible with MathOptInterface v0.10, which enables compatibility with the latest version of many solvers (#467, #468).

v0.14.16 (September 25, 2021)

  • Improve numerical stability when evaluating logsumexp (#457). Thanks @JinraeKim!

v0.14.15 (September 15, 2021)

  • Use sparse factorization for checking for positive semi-definiteness in quadform when possible (#457). Thanks @mtanneau!
  • Add assume_psd=false argument to skip checking for positive semi-definiteness in quadform (#456).

v0.14.14 (September 8, 2021)

  • Increase the tolerance used in checking if a matrix is positive-semi definite in quadform (#453). Thanks @numbermaniac!

v0.14.13 (July 25, 2021)

  • fix quadform for positive semi-definite matrices (fixes a regression introduced in v0.14.11 that required strictly positive semi-definite inputs) #450.

v0.14.12 (July 19, 2021)

  • fix size of result of evaluate on IndexAtoms #448. Thanks @hurak!

v0.14.11 (July 5, 2021)

  • fix quadform in the complex case #444. Thanks @lrnv!

v0.14.10 (May 20, 2021)

  • declare compatibility with BenchmarkTools v1.0 #441

v0.14.9 (May 18, 2021)

  • fix some tests in lp_dual_abs_atom #439. Thanks @moehle!

v0.14.8 (May 4, 2021)

  • a complete port of cvxquad thanks to @dstahlke, yielding new functions quantum_relative_entropy, quantum_entropy, trace_logm, trace_mpower, and lieb_ando, and cones GeomMeanHypoCone, GeomMeanEpiCone, and RelativeEntropyEpiCone (#418). Thanks a ton for the awesome contribution @dstahlke!

v0.14.7 (April 22, 2021)

  • declare compatibility with BenchmarkTools v0.7 #434

v0.14.6 (March 28, 2021)

v0.14.5 (March 14, 2021)

  • allow sumlargest(x,k), sumsmallest(x,k), and sumlargesteigs(x,k) for k=0 (simply returns Constant(0)). (#429).

v0.14.4 (March 14, 2021)

  • fixed a bug where the values of variables were being converted to Float64 even if the problem was solved in high precision. (#427).

v0.14.3 (March 10, 2021)

  • update compatibility bounds for BenchmarkTools 0.6

v0.14.2 (February 15, 2021)

  • added lasso, ridge, and elastic net regression examples (#420). Thanks to @PaulSoderlind!

v0.14.1 (January 24, 2021)

  • there was a bug causing conj to act in-place (reported in #416), which has been fixed (#417). This bug appears to have existed since the introduction of conj in Convex.jl v0.5.0.

v0.14.0 (January 17, 2021)

Breaking changes

  • Changes to the sign of atoms:
    • The sign of sumlargesteigs has been changed from Positive() to NoSign(), to allow non-positive-semidefinite inputs (#409). This has the potential to break code that required that sign to be positive. If you run into this problem, please file an issue so we can figure out a workaround.
    • The sign of eigmin and eigmax has been changed from Positive() to NoSign() (#413). This is a bugfix because in general eigmin and eigmax do not need to return a positive quantity (for non-positive-semidefinite inputs). Again, this has the potential to break code that required that sign to be positive. If you run into this problem, please file an issue so we can figure out a workaround.
  • Removal of deprecations:
    • lambdamin and lambdamax has been deprecated to eigmin and eigmax since Convex v0.13.0. This deprecation has been removed, so your code must be updated to call eigmin or eigmax instead (#412).
    • norm(x, p) where x is a matrix expression has been deprecated to opnorm(x,p) since Convex v0.8.0. This deprecation has been removed, so your code must be updated to call opnorm(x, p) instead (#412). Currently, norm(x,p) for a matrix
    expression x will error, but in Convex.jl v0.15.0 it will return norm(vec(x), p).
    • Convex.clearmemory() has been deprecated and unnecessary since Convex v0.12.5. This deprecation has been removed, so if this function is in your code, just delete it (#412).
    • vecnorm(x, p) has been deprecated to norm(vec(x), p) since Convex v0.8.0. This deprecation has been removed, so your code must be updated to call norm(vec(x),p) instead (#412).
  • Other changes:
    • Convex.DCP_WARNINGS was introduced in Convex v0.13.1 to allow turning off Convex.jl's DCP warnings. This has been removed in favor of the function Convex.emit_dcp_warnings() (Commit 481fa02).

Other changes

  • updated nuclearnorm and sumlargesteigs to allow complex variables, and allow the argument of sumlargesteigs to be non-positive-semi-definite (#409). Thanks to @dstahlke!

v0.13.8 (December 2, 2020)

  • add unary + for Sign and ComplexSign to allow single-argument hcat and vcat to work (#405). Thanks to @dstahlke!

v0.13.7 (September 11, 2020)

  • fix #403 by adding the keyword argument silent_solver to solve!.

v0.13.6 (September 8, 2020)

  • fix #401 by allowing diagm(x).

v0.13.5 (August 25, 2020)

  • fix #398 by allowing fix!'d variables in quadform.

v0.13.4 (July 27, 2020)

  • You can now create your own variable types by subtyping AbstractVariable. See the docs for more information. You can also add constraints directly to a variable using add_constraint! (#358).
  • Functions vexity(x::Variable), sign(x::Variable), and evaluate(x::Variable) should now be the preferred way to access properties of a variable; likewise use set_value! to set the initial value of a variable (#358).
  • To create integer or binary constraints, use the VarType enum (for example, Variable(BinVar)). Access or set this via vartype and vartype! (#358).

v0.13.3 (March 22, 2020)

v0.13.2 (March 14, 2020)

v0.13.1 (March 6, 2020)

  • Allow disabling DCP warnings (#372)
  • Restore export of Constraint (#371)

v0.13.0 (February 28, 2020)

Major changes

  • The intermediate layer has changed from MathProgBase.jl to MathOptInterface.jl (#330). To solve problems, one should pass a MathOptInterface optimizer constructor, such as SCS.Optimizer, or MOI.OptimizerWithAttributes(SCS.Optimizer, "verbose" => 0).
  • lambdamin and lambdamax have been deprecated in favor of eigmin and eigmax (#357).
  • Many "internal" functions and types are no longer exported, such as the atoms, types corresponding to constraints and vexities, etc. (#357).
  • evaluate(x::Variable) and evaluate(c::Constant) now return scalars and vectors as appropriate, instead of (1,1)- and (d,1)-matrices (#359). This affects functions which used to return (1,1)-matrices; for example, now evaluate(quadform(...)) yields a scalar.