[baggepinnen]

Related to JuliaComputing/DyadControlSystems.jl#837 ("analysis point on array applies to first element only").

Problem

named_ss, named_ss(sys, linfun, outputs) and the get_named_* functions counted one input and one output per entry of inputs and outputs. An analysis point on an array-valued connection, or an array variable, contributes one channel per element to the linearization:

• For a 2-vector input analysis point, nu = 1 and B has two columns. The condition size(matrices.B, 2) == 2nu then holds, the second column is taken to be the derivative of the first, and causal_simplification eliminates it. The result is a one-input system that contains only the first element, together with a spurious nonzero D.
• For other sizes, constructing the NamedStateSpace fails with Length of input names must match size(sys,2).

Changes

• The number of channels of each input and output is determined from the variable that drives it: the analysis point's input variable, obtained through GetInput, or the variable itself.
• Signal names are expanded to one name per channel, u1, u2, ..., following expand_symbol, so the channels of a signal remain selectable by its name (P[:y, :u]).
• The input operating point is flattened to one entry per channel.
• If an expanded channel name coincides with another signal's name, for example an array-valued analysis point x with two elements next to a scalar analysis point x1, an ArgumentError names the coinciding channels and the signals they originate from. Previously this would have failed in the NamedStateSpace constructor with a message that did not mention the expansion.
• ControlSystemsMTK.signal_channel_names(sys, signals) (not exported) returns the channel names in the order used by named_ss. Downstream packages can use it to get the channel count of each signal from the model instead of inferring it from the channel names. JuliaComputing/DyadControlSystems.jl#840 uses it.

Tests

test/test_ODESystem.jl gains a testset with a 2×2 loop whose connections carry 2-vector analysis points. It checks named_ss with and without allow_input_derivatives, with array-valued and with scalar outputs, and get_named_sensitivity, as well as signal_channel_names and the error for coinciding channel names. The full test suite passes locally (60 passed, 1 broken).

Loop openings at array-valued analysis points of components built directly in MTK additionally require SciML/ModelingToolkit.jl#5349; the tests here do not use them.

Verified end-to-end on a vector-analysis-point variant of the F16 LQG tutorial (LQGAnalysis through DyadControlSystems): the resulting controller agrees with the design obtained with scalar analysis points to a relative difference of 1e-7.

---

Host: demeter2, session 649523b8-6ed5-4d11-864f-e0f78038c754, working directory ~/.julia/dev/ModelingToolkit/.claude/worktrees/soft-wiggling-wreath (sources edited in a scratch copy and committed through the GitHub API)

🤖 Generated with Claude Code