[docs] clarify that JuMP can reformulate indicator constraints

[odow]

Closes #3701

Is this better @schlichtanders?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 98.35%. Comparing base (434940d) to head (d61127b).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #3703   +/-   ##
=======================================
  Coverage   98.35%   98.35%           
=======================================
  Files          43       43           
  Lines        5707     5707           
=======================================
  Hits         5613     5613           
  Misses         94       94           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[schlichtanders]

Thank you so much, yes that is better.

The Big-M trick only works if the domain is limited (in this case by the Big-M), hence I find it a bit self-contradictory.
But I am fine with leaving it. Who knows who gets which benefits by seeing the Big-M formulation there

[odow]

I guess the M is chosen by the user. Whereas JuMP cannot automatically perform the same reformulation only if we know the bounds to infer a suitable M.

In practice, all decision variables in reality have bounds (you cannot have infinite anything). It's just that the user may not directly specify them, or know an appropriate value.

[remi-garcia]

I just saw this with the new release. The ``Trick 2'' bothers me as it is not possible to use big M constraints if variables have not a finite domain. Doing so actually add bounds.

MWE for which this changes the result:

model = Model()
@variable(model, x[1:2])
@variable(model, z, Bin)
@constraint(model, z --> {sum(x) <= 1})
@objective(model, Max, sum(x))

Termination status: unbounded while

model = Model()
@variable(model, x[1:2])
@variable(model, z, Bin)
M = 100   # the specific value does not matter
@constraint(model, sum(x) <= 1 + M * (1 - z))
@objective(model, Max, sum(x))

Termination status: optimal, objective value: M+1