Error when nonzero coefficient close to zero

[guberger]

Hello! I get an error when I run the following code:

using JuMP
using HiGHS
model = Model(HiGHS.Optimizer)
@variable(model, x)
@constraint(model, 1e-9*x <= 1) # works with @constraint(model, 0*x <= 1)
optimize!(model)

WARNING: LP matrix packed vector contains 1 |values| in [1e-09, 1e-09] less than 1e-09: ignored
ERROR: Encountered an error in HiGHS (Status 1). Check the log for details.

(log file empty) Version info:

Julia Version 1.7.0
Commit 3bf9d17731 (2021-11-30 12:12 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i7-7600U CPU @ 2.80GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, skylake)

  [87dc4568] HiGHS v0.3.0
  [4076af6c] JuMP v0.22.1

[odow]

Interesting edge case. I guess this makes sense. HiGHS ignores very small coefficients, and then once you've ignored the small coefficient it's left with a constraint matrix that has no non-zeros.

What happens if you have other non-zeros? i.e., other variables or constraints with coefficients > 1e-09

cc @jajhall

[guberger]

Similar code with

@variable(model, x[1:2])
@constraint(model, [1, 1e-9]'*x <= 1)

fails as well :-/ (same for

@variable(model, x)
@variable(model, y)
@constraint(model, x + 1e-9*y <= 1)

)

[jajhall]

Anything less than or equal to small_matrix_value (which is 1e-9 by default) is ignored by HiGHS. The wording of the warning message could be better.

If you really want values of 1e-9 to be used, reduce the value of small_matrix_value using a setOptions call

[odow]

I think the problem is the error code at some point. I'll need to come up with a reproducible example in C, but it shouldn't be too difficult.

[odow]

@guberger do you have the full stack trace?

[guberger]

Here :-)

WARNING: LP matrix packed vector contains 1 |values| in [1e-09, 1e-09] less than 1e-09: ignored
ERROR: Encountered an error in HiGHS (Status 1). Check the log for details.
Stacktrace:
  [1] error(s::String)
    @ Base .\error.jl:33
  [2] _check_ret
    @ C:\Users\guberger\.julia\packages\HiGHS\gxmTb\src\MOI_wrapper.jl:347 [inlined]
  [3] add_constraint(model::HiGHS.Optimizer, f::MathOptInterface.ScalarAffineFunction{Float64}, s::MathOptInterface.LessThan{Float64})
    @ HiGHS C:\Users\guberger\.julia\packages\HiGHS\gxmTb\src\MOI_wrapper.jl:1279
  [4] _copy_constraints(dest::HiGHS.Optimizer, src::MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}, index_map::MathOptInterface.Utilities.IndexMap, cis_src::Vector{MathOptInterface.ConstraintIndex{MathOptInterface.ScalarAffineFunction{Float64}, MathOptInterface.LessThan{Float64}}})
    @ MathOptInterface.Utilities C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Utilities\copy.jl:248
  [5] pass_nonvariable_constraints_fallback(dest::HiGHS.Optimizer, src::MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}, index_map::MathOptInterface.Utilities.IndexMap, constraint_types::Vector{Any})
    @ MathOptInterface.Utilities C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Utilities\copy.jl:261
  [6] pass_nonvariable_constraints
    @ C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Utilities\copy.jl:287 [inlined]
  [7] pass_nonvariable_constraints(dest::MathOptInterface.Bridges.LazyBridgeOptimizer{HiGHS.Optimizer}, src::MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}, idxmap::MathOptInterface.Utilities.IndexMap, constraint_types::Vector{Any})
    @ MathOptInterface.Bridges C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Bridges\bridge_optimizer.jl:406
  [8] _pass_constraints(dest::MathOptInterface.Bridges.LazyBridgeOptimizer{HiGHS.Optimizer}, src::MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}, index_map::MathOptInterface.Utilities.IndexMap, variable_constraints_not_added::Vector{Any})
    @ MathOptInterface.Utilities C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Utilities\copy.jl:309
  [9] default_copy_to(dest::MathOptInterface.Bridges.LazyBridgeOptimizer{HiGHS.Optimizer}, src::MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}})
    @ MathOptInterface.Utilities C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Utilities\copy.jl:442
 [10] #copy_to#7
    @ C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Bridges\bridge_optimizer.jl:421 [inlined]
 [11] copy_to
    @ C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Bridges\bridge_optimizer.jl:421 [inlined]
 [12] optimize!
    @ C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\MathOptInterface.jl:80 [inlined]
 [13] optimize!(m::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.Bridges.LazyBridgeOptimizer{HiGHS.Optimizer}, MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}})
    @ MathOptInterface.Utilities C:\Users\guberger\.julia\packages\MathOptInterface\eoIu0\src\Utilities\cachingoptimizer.jl:285
 [14] optimize!(model::Model, optimizer_factory::Nothing; bridge_constraints::Bool, ignore_optimize_hook::Bool, kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ JuMP C:\Users\guberger\.julia\packages\JuMP\2IF9U\src\optimizer_interface.jl:195
 [15] optimize! (repeats 2 times)
    @ C:\Users\guberger\.julia\packages\JuMP\2IF9U\src\optimizer_interface.jl:167 [inlined]
 [16] top-level scope
    @ REPL[13]:1

[guberger]

The problem seems to be in addRow in the C code of HiGHS, which returns an error code, following which _check_ret throws an error.

[jajhall]

The problem seems to be in addRow in the C code of HiGHS, which returns an error code, following which _check_ret throws an error.

I've reproduced your example with a C++ call to HiGHS. It loads the problem using a call to addCol and then addRow, and then solves it OK. For sure, addRow will return 1=(int)HighsStatus::kWarning rather than 0=HighsStatus::kOk, since we want to warn users that at least one of their values has been ignored. Maybe the Julia interface didn't anticipate having to handle anything other than kOk.

I've inserted "or equal to" into the warning message. :-)

[odow]

Okay, we need to update https://github.com/jump-dev/HiGHS.jl/blob/567ed76f40fe4fa0e65979219367054afca6b287/src/MOI_wrapper.jl#L345-L353 to not error on warnings. But potentially double check all usages to check there aren't any instances where 1 is a boolean error code.

[Mastomaki]

Hello!

I get this same error. It would help if I knew which coefficient is causing the problem. Is the coefficient as given by the user or is there already some automatic scaling?

WARNING: LP matrix packed vector contains 1 |values| in [2.14431e-10, 2.14431e-10] less than 1e-09: ignored
ERROR: LoadError: Encountered an error in HiGHS (Status 1). Check the log for details.

[jajhall]

This coefficient is given by the user. When there's only one small coefficient, clearly its row and column indices could be given, but if there are vast numbers, it doesn't seem sensible to report them all individually.

What might be helpful is a message that coefficient (row, col) is in [a, b] less than 1e-09, as are N others. Then a user would have somewhere to go to look for the small value, and that might explain all the rest?

[Mastomaki]

That's one possiblity. But I suppose it is possible to list more coefficients in the log file?

[jajhall]

If all small coefficients were listed by default, we'd be sure to have users complaining about excessive reporting. There is the option log_dev_level that is primarily for development logging so has a default value of zero. I could put out a full list of small values if it were set to a positive value. It wouldn't yield excessive development logging - principally a statement of feasibility etc every time the simplex basis matrix is refactorized - which was too "chatty" for some users to see by default!