control-toolbox · ocots · Jul 4, 2026 · Jul 4, 2026 · Jul 4, 2026 · Jul 4, 2026
diff --git a/BREAKINGS.md b/BREAKINGS.md
@@ -26,6 +26,22 @@ This document outlines all breaking changes introduced in CTBase v0.18.0-beta co
 - `NotStored` / `NotStoredType` are unchanged and remain extraction-internal to
   `CTBase.Options` (the defining file was renamed `not_provided.jl` → `not_stored.jl`).
 
+## Non-breaking note (0.26.1-beta)
+
+- **Traits**: New `Feedback` trait family added for encoding how a control law closes the loop
+  - **New trait family**: `AbstractFeedback` with tags `OpenLoopFeedback`, `ClosedLoopFeedback`, `DynClosedLoopFeedback`
+  - **Type-parameter-only contract**: trait value read from a type parameter by the `feedback` accessor; no `has_feedback_trait` guard
+  - **Derived predicates**: `is_open_loop(obj)`, `is_closed_loop(obj)`, `is_dyn_closed_loop(obj)`
+  - **No breaking changes**: Purely additive. Existing code unaffected.
+- **Data**: New `PseudoHamiltonian` and `ControlLaw` data types added
+  - **PseudoHamiltonian**: scalar function `H̃(t, x, p, u[, v]) → ℝ` with explicit control argument; `AbstractPseudoHamiltonian` supertype and `PseudoHamiltonian` concrete struct
+  - **ControlLaw**: control law function `u(⋯) → 𝒰` with feedback trait; `AbstractControlLaw` supertype, `ControlLaw` concrete struct, and `OpenLoop`/`ClosedLoop`/`DynClosedLoop` user-facing constructors
+  - **No breaking changes**: Purely additive. Existing data types unchanged.
+- **Differentiation**: New pseudo-Hamiltonian gradient methods added
+  - `pseudo_hamiltonian_gradient(backend, h̃, t, x, p, u, v) → (∂H̃/∂x, ∂H̃/∂p)`
+  - `pseudo_hamiltonian_control_gradient(backend, h̃, t, x, p, u, v) → ∂H̃/∂u`
+  - **No breaking changes**: Purely additive. Existing gradient methods unchanged.
+
 ## Non-breaking note (0.26.0-beta)
 
 - **Traits**: New `ControlDependence` family added for encoding control presence in optimal control problems

diff --git a/CHANGELOGS.md b/CHANGELOGS.md
@@ -5,6 +5,62 @@ All notable changes to CTBase will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.26.1-beta] - 2026-07-04
+
+### ✨ New Features
+
+#### **Feedback Trait Family**
+
+- **New `AbstractFeedback` trait family** in `CTBase.Traits` for encoding how a control law closes the loop:
+  - **Tags**: `OpenLoopFeedback` (time/variable only), `ClosedLoopFeedback` (time + state), `DynClosedLoopFeedback` (time + state + costate)
+  - **Type-parameter-only contract**: the trait value is read from a type parameter of the concrete data type by the `feedback` accessor; no `has_feedback_trait` guard
+  - **Derived predicates**: `is_open_loop(obj)`, `is_closed_loop(obj)`, `is_dyn_closed_loop(obj)` dispatch on the feedback type parameter
+  - **Dynamics trait mapping**: open-loop and closed-loop → `StateDynamics`; dynamic closed-loop → `HamiltonianDynamics`
+
+#### **PseudoHamiltonian**
+
+- **New `PseudoHamiltonian` data type** in `CTBase.Data` for scalar functions `H̃(t, x, p, u[, v]) → ℝ` that extend the standard Hamiltonian with an explicit control argument `u`:
+  - **Abstract supertype**: `AbstractPseudoHamiltonian{TD,VD}` with trait accessors
+  - **Concrete struct**: `PseudoHamiltonian{F,TD,VD}` with keyword constructor `PseudoHamiltonian(f; is_autonomous, is_variable)`
+  - **Natural and uniform call signatures** matching the time/variable dependence traits
+  - **Dynamics trait**: always `HamiltonianDynamics`
+
+#### **ControlLaw**
+
+- **New `ControlLaw` data type** in `CTBase.Data` for control law functions `u(⋯) → 𝒰`:
+  - **Abstract supertype**: `AbstractControlLaw{FB,TD,VD}` with feedback, time, and variable dependence trait accessors
+  - **Concrete struct**: `ControlLaw{F,FB,TD,VD}` with typed constructor
+  - **User-facing constructors**: `OpenLoop(f; ...)`, `ClosedLoop(f; ...)`, `DynClosedLoop(f; ...)` fixing the feedback trait
+  - **Feedback-dependent call signatures**: natural and uniform signatures vary by feedback trait
+  - **Dynamics trait**: determined by feedback (open/closed-loop → `StateDynamics`; dynamic closed-loop → `HamiltonianDynamics`)
+
+#### **Pseudo-Hamiltonian Gradient Methods**
+
+- **New AD contract methods** in `CTBase.Differentiation` for pseudo-Hamiltonian gradients:
+  - `pseudo_hamiltonian_gradient(backend, h̃, t, x, p, u, v) → (∂H̃/∂x, ∂H̃/∂p)`: non-negated partial derivatives
+  - `pseudo_hamiltonian_control_gradient(backend, h̃, t, x, p, u, v) → ∂H̃/∂u`: control gradient (vanishes along optimal trajectories by PMP stationarity)
+  - **Rationale**: splitting `∂H̃/∂u` from `(∂H̃/∂x, ∂H̃/∂p)` reflects the distinct role of the stationarity condition and avoids computing the control gradient when only the state/costate gradients are needed
+  - **Extension implementation**: `CTBaseDifferentiationInterface` extension provides concrete implementations using `DifferentiationInterface.jl`
+  - **Fallback**: `NotImplemented` on `AbstractADBackend` without the extension
+
+### 📚 Documentation
+
+- **Guide pages updated**:
+  - `traits.md`: added Feedback trait family section with type-parameter-only contract note, predicates, and dynamics trait mapping; added Feedback to Other families table and accessor/predicate summary
+  - `data.md`: added PseudoHamiltonian section (construction, calling, dynamics trait) and ControlLaw section (constructors, call signatures, feedback→dynamics mapping); updated overview table, typed constructors, and See also
+  - `differentiation.md`: updated contract count to nine methods; added Pseudo-Hamiltonian gradients section with examples; updated extension note and See also
+- **Docstring compliance**: aligned all docstrings in `feedback.jl`, `abstract_control_law.jl`, `control_law.jl`, `abstract_pseudo_hamiltonian.jl`, `pseudo_hamiltonian.jl`, `abstract_ad_backend.jl`, and `CTBaseDifferentiationInterface.jl` with the Handbook conventions (`See also:` inline lines, `# Returns`/`# Arguments` sections, `# Output` for show methods)
+- **API reference**: added `feedback.jl`, `abstract_control_law.jl`, `control_law.jl`, `abstract_pseudo_hamiltonian.jl`, `pseudo_hamiltonian.jl` to auto-generated API pages
+
+### 🧪 Testing
+
+- Added tests for `pseudo_hamiltonian_gradient` (return signature `(grad_x, grad_p)`) and `pseudo_hamiltonian_control_gradient` (return `grad_u`), including `NotImplemented` error tests and `DifferentiationInterface` extension tests
+- Added `pseudo_hamiltonian_control_gradient` to export-list checks in `test_differentiation_module.jl`
+
+### ✅ Compatibility
+
+- **No breaking changes**: purely additive. Existing code unaffected. See [BREAKINGS.md](BREAKINGS.md).
+
 ## [0.26.0-beta] - 2026-06-28
 
 ### ✨ New Features

diff --git a/Project.toml b/Project.toml
@@ -1,6 +1,6 @@
 name = "CTBase"
 uuid = "54762871-cc72-4466-b8e8-f6c8b58076cd"
-version = "0.26.0-beta"
+version = "0.26.1-beta"
 authors = ["Olivier Cots <olivier.cots@irit.fr>", "Jean-Baptiste Caillau <caillau@univ-cotedazur.fr>"]
 
 [deps]

diff --git a/docs/api_reference.jl b/docs/api_reference.jl
@@ -54,6 +54,10 @@ function generate_api_reference(src_dir::String)
                 joinpath("Data", "hamiltonian.jl"),
                 joinpath("Data", "abstract_hamiltonian_vector_field.jl"),
                 joinpath("Data", "hamiltonian_vector_field.jl"),
+                joinpath("Data", "abstract_control_law.jl"),
+                joinpath("Data", "control_law.jl"),
+                joinpath("Data", "abstract_pseudo_hamiltonian.jl"),
+                joinpath("Data", "pseudo_hamiltonian.jl"),
             ),
         ),
         (
@@ -175,6 +179,7 @@ function generate_api_reference(src_dir::String)
                 joinpath("Traits", "time_dependence.jl"),
                 joinpath("Traits", "variable_dependence.jl"),
                 joinpath("Traits", "control_dependence.jl"),
+                joinpath("Traits", "feedback.jl"),
             ),
         ),
         (

diff --git a/docs/src/guide/data.md b/docs/src/guide/data.md
@@ -22,10 +22,13 @@ using LinearAlgebra
 |---|---|---|
 | [`Data.VectorField`](@ref CTBase.Data.VectorField) | ``X : \mathcal{X} \to \mathbb{R}^n`` | `X(x)` |
 | [`Data.Hamiltonian`](@ref CTBase.Data.Hamiltonian) | ``H : T^*\mathcal{X} \to \mathbb{R}`` | `H(x, p)` |
+| [`Data.PseudoHamiltonian`](@ref CTBase.Data.PseudoHamiltonian) | ``\tilde{H} : T^*\mathcal{X} \times \mathcal{U} \to \mathbb{R}`` | `H̃(x, p, u)` |
 | [`Data.HamiltonianVectorField`](@ref CTBase.Data.HamiltonianVectorField) | ``\vec{H} : T^*\mathcal{X} \to \mathbb{R}^{2n}`` | `HVF(x, p)` |
+| [`Data.ControlLaw`](@ref CTBase.Data.ControlLaw) | ``u : \cdots \to \mathcal{U}`` | `u()` / `u(x)` / `u(x, p)` |
 
-All three share the same trait axes — time dependence, variable dependence and
-mutability. See [Traits](traits.md) for the full call-signature tables.
+All five share the same trait axes — time dependence and variable dependence.
+`VectorField` and `HamiltonianVectorField` also carry a mutability trait;
+`ControlLaw` carries a feedback trait (see [Traits](traits.md)).
 
 ---
 
@@ -204,6 +207,117 @@ For a plain user-supplied function that does not accept `variable_costate`, pass
 
 ---
 
+## PseudoHamiltonian
+
+A `PseudoHamiltonian` wraps a scalar function ``\tilde{H}(x, p, u) \in \mathbb{R}``
+that extends the standard Hamiltonian with an explicit control argument ``u``.
+
+```math
+\tilde{H} : T^*\mathcal{X} \times \mathcal{U} \to \mathbb{R}, \quad (x, p, u) \mapsto \tilde{H}(x, p, u).
+```
+
+Unlike [`Data.Hamiltonian`](@ref CTBase.Data.Hamiltonian), which encodes the
+control implicitly, a pseudo-Hamiltonian takes the control as an additional
+argument. This enables dynamic closed-loop flows where the control is computed
+from the pseudo-Hamiltonian's maximisation condition (PMP stationarity:
+``\partial \tilde{H} / \partial u = 0``).
+
+### Construction
+
+```@example data
+# Autonomous, fixed (default): H̃(x, p, u)
+ph1 = Data.PseudoHamiltonian((x, p, u) -> dot(p, x) + u^2)
+
+# Non-autonomous, fixed: H̃(t, x, p, u)
+ph2 = Data.PseudoHamiltonian((t, x, p, u) -> t * dot(p, x) + u^2; is_autonomous=false)
+
+# Autonomous, non-fixed: H̃(x, p, u, v)
+ph3 = Data.PseudoHamiltonian((x, p, u, v) -> dot(p, x) + u^2 + v[1]; is_variable=true)
+
+ph1
+```
+
+### Calling
+
+```@example data
+x0, p0, u0 = [1.0, 0.5], [0.3, 0.7], 2.0
+
+ph1(x0, p0, u0)                    # natural call
+ph1(0.0, x0, p0, u0, nothing)      # uniform call
+ph2(0.5, x0, p0, u0)               # non-autonomous
+```
+
+The dynamics trait of a `PseudoHamiltonian` is always
+[`Traits.HamiltonianDynamics`](@ref CTBase.Traits.HamiltonianDynamics), since
+pseudo-Hamiltonians involve both state and costate.
+
+---
+
+## ControlLaw
+
+A `ControlLaw` wraps a function ``u(\cdots)`` that provides the control input for
+an optimal control problem. The feedback trait (see [Traits](traits.md))
+determines which arguments the control law depends on.
+
+Three user-facing constructors fix the feedback trait:
+
+| Constructor | Feedback trait | Natural signature (autonomous/fixed) |
+|---|---|---|
+| [`Data.OpenLoop`](@ref CTBase.Data.OpenLoop) | `OpenLoopFeedback` | `u()` |
+| [`Data.ClosedLoop`](@ref CTBase.Data.ClosedLoop) | `ClosedLoopFeedback` | `u(x)` |
+| [`Data.DynClosedLoop`](@ref CTBase.Data.DynClosedLoop) | `DynClosedLoopFeedback` | `u(x, p)` |
+
+### Construction
+
+```@example data
+# Open-loop, autonomous, fixed (default): u()
+u_ol = Data.OpenLoop(() -> 1.0)
+
+# Closed-loop, autonomous, fixed: u(x)
+u_cl = Data.ClosedLoop(x -> -x)
+
+# Dynamic closed-loop, autonomous, fixed: u(x, p)
+u_dcl = Data.DynClosedLoop((x, p) -> -x - p)
+
+u_ol
+```
+
+Non-autonomous and variable variants are constructed with the same keywords as
+other data types:
+
+```@example data
+u_ol_na = Data.OpenLoop((t, v) -> t * v; is_autonomous=false, is_variable=true)
+```
+
+### Calling
+
+Every `ControlLaw` is callable via its **natural** signature (matching the
+traits) and via a **uniform** signature that depends on the feedback trait:
+
+| Feedback | Natural `(Aut, Fixed)` | Uniform |
+|---|---|---|
+| `OpenLoop` | `u()` | `u(t, v)` |
+| `ClosedLoop` | `u(x)` | `u(t, x, v)` |
+| `DynClosedLoop` | `u(x, p)` | `u(t, x, p, v)` |
+
+```@example data
+u_ol()                         # natural call
+u_ol(0.0, nothing)             # uniform call
+
+u_cl(x0)                       # natural call
+u_cl(0.0, x0, nothing)         # uniform call
+
+u_dcl(x0, p0)                  # natural call
+u_dcl(0.0, x0, p0, nothing)    # uniform call
+```
+
+The feedback trait determines the dynamics trait: open-loop and closed-loop
+control laws carry [`Traits.StateDynamics`](@ref CTBase.Traits.StateDynamics),
+while dynamic closed-loop control laws carry
+[`Traits.HamiltonianDynamics`](@ref CTBase.Traits.HamiltonianDynamics).
+
+---
+
 ## Querying traits
 
 Because the trait metadata lives in the type, it can be recovered from any data
@@ -251,15 +365,19 @@ Traits.time_dependence(vf_typed)
 ```
 
 The same positional form exists for [`Data.Hamiltonian`](@ref CTBase.Data.Hamiltonian)
-(`Hamiltonian(f, TD, VD)`) and
+(`Hamiltonian(f, TD, VD)`),
+[`Data.PseudoHamiltonian`](@ref CTBase.Data.PseudoHamiltonian)
+(`PseudoHamiltonian(f, TD, VD)`),
 [`Data.HamiltonianVectorField`](@ref CTBase.Data.HamiltonianVectorField)
-(`HamiltonianVectorField(f, TD, VD, MD)`).
+(`HamiltonianVectorField(f, TD, VD, MD)`), and
+[`Data.ControlLaw`](@ref CTBase.Data.ControlLaw)
+(`ControlLaw(f, FB, TD, VD)`).
 
 ---
 
 ## See also
 
-- [`Data.VectorField`](@ref CTBase.Data.VectorField), [`Data.Hamiltonian`](@ref CTBase.Data.Hamiltonian), [`Data.HamiltonianVectorField`](@ref CTBase.Data.HamiltonianVectorField) — concrete data types.
-- [`Data.AbstractVectorField`](@ref CTBase.Data.AbstractVectorField), [`Data.AbstractHamiltonian`](@ref CTBase.Data.AbstractHamiltonian), [`Data.AbstractHamiltonianVectorField`](@ref CTBase.Data.AbstractHamiltonianVectorField) — abstract supertypes.
+- [`Data.VectorField`](@ref CTBase.Data.VectorField), [`Data.Hamiltonian`](@ref CTBase.Data.Hamiltonian), [`Data.PseudoHamiltonian`](@ref CTBase.Data.PseudoHamiltonian), [`Data.HamiltonianVectorField`](@ref CTBase.Data.HamiltonianVectorField), [`Data.ControlLaw`](@ref CTBase.Data.ControlLaw) — concrete data types.
+- [`Data.AbstractVectorField`](@ref CTBase.Data.AbstractVectorField), [`Data.AbstractHamiltonian`](@ref CTBase.Data.AbstractHamiltonian), [`Data.AbstractPseudoHamiltonian`](@ref CTBase.Data.AbstractPseudoHamiltonian), [`Data.AbstractHamiltonianVectorField`](@ref CTBase.Data.AbstractHamiltonianVectorField), [`Data.AbstractControlLaw`](@ref CTBase.Data.AbstractControlLaw) — abstract supertypes.
 - [Traits](traits.md) — the three trait axes, the dynamics trait, and call-signature tables.
 - [Exceptions](exceptions.md) — `PreconditionError` and `IncorrectArgument`, raised by the constructors and the `variable_costate` path.
diff --git a/docs/src/guide/differentiation.md b/docs/src/guide/differentiation.md
@@ -42,14 +42,15 @@ The concrete strategy stores a single option, `:ad_backend`, holding an
 `AutoForwardDiff()`). `ADTypes` is a hard dependency of CTBase, so the default is
 always available.
 
-The contract has seven methods.
+The contract has nine methods.
 [`Differentiation.ad_backend`](@ref CTBase.Differentiation.ad_backend) — the accessor
 returning the wrapped ADTypes backend — is resolved in core and always available; the
-six differentiation primitives below live in an extension.
+eight differentiation primitives below live in an extension.
 
 !!! note "The differentiation methods live in an extension"
     The contract methods (`gradient`, `derivative`, `differentiate`,
-    `pushforward`, `hamiltonian_gradient`, `variable_gradient`) are implemented in
+    `pushforward`, `hamiltonian_gradient`, `variable_gradient`,
+    `pseudo_hamiltonian_gradient`, `pseudo_hamiltonian_control_gradient`) are implemented in
     the `CTBaseDifferentiationInterface` **package extension**. They become
     available only once `DifferentiationInterface` *and* a concrete AD package
     (e.g. `ForwardDiff`) are loaded. Without the extension, every contract method
@@ -164,6 +165,38 @@ hv = Data.Hamiltonian((x, p, v) -> 0.5 * (sum(abs2, x) + sum(abs2, p) + sum(abs2
 Differentiation.variable_gradient(backend, hv, 0.0, [1.0, 2.0], [3.0, 4.0], [5.0, 6.0])
 ```
 
+## Pseudo-Hamiltonian gradients
+
+The two pseudo-Hamiltonian methods operate directly on a
+[`Data.PseudoHamiltonian`](@ref CTBase.Data.PseudoHamiltonian) (see the [Data](data.md)
+guide).
+[`Differentiation.pseudo_hamiltonian_gradient`](@ref CTBase.Differentiation.pseudo_hamiltonian_gradient)
+returns `(∂H̃/∂x, ∂H̃/∂p)`, **non-negated** — the caller applies the signs of
+Hamilton's equations:
+
+```@example diff
+# H̃(x, p, u) = ½(‖x‖² + ‖p‖²) + u²  →  ∂H̃/∂x = x, ∂H̃/∂p = p
+ph = Data.PseudoHamiltonian((x, p, u) -> 0.5 * (sum(abs2, x) + sum(abs2, p)) + u^2)
+∂x, ∂p = Differentiation.pseudo_hamiltonian_gradient(backend, ph, 0.0, [1.0, 2.0], [3.0, 4.0], 2.0, nothing)
+(∂x, ∂p)
+```
+
+[`Differentiation.pseudo_hamiltonian_control_gradient`](@ref CTBase.Differentiation.pseudo_hamiltonian_control_gradient)
+returns `∂H̃/∂u`, which vanishes along optimal trajectories by the Pontryagin
+Maximum Principle stationarity condition:
+
+```@example diff
+# ∂H̃/∂u = 2u = 4.0
+Differentiation.pseudo_hamiltonian_control_gradient(backend, ph, 0.0, [1.0, 2.0], [3.0, 4.0], 2.0, nothing)
+```
+
+For a non-fixed pseudo-Hamiltonian, the `v` argument carries the variable:
+
+```@example diff
+phv = Data.PseudoHamiltonian((x, p, u, v) -> 0.5 * (sum(abs2, x) + sum(abs2, p)) + u^2 + v[1]; is_variable=true)
+Differentiation.pseudo_hamiltonian_control_gradient(backend, phv, 0.0, [1.0, 2.0], [3.0, 4.0], 2.0, [5.0])
+```
+
 ## The contract without the extension
 
 Every contract method has a fallback on `AbstractADBackend` that throws
@@ -186,5 +219,5 @@ end # hide
 
 - [`Differentiation.AbstractADBackend`](@ref CTBase.Differentiation.AbstractADBackend), [`Differentiation.DifferentiationInterface`](@ref CTBase.Differentiation.DifferentiationInterface) — the strategy types.
 - [Implementing a Strategy](@ref), [Options System](@ref) — the contract these build on.
-- [Data](data.md) — `Hamiltonian` wrappers consumed by the gradient methods.
+- [Data](data.md) — `Hamiltonian` and `PseudoHamiltonian` wrappers consumed by the gradient methods.
 - [Exceptions](exceptions.md) — `NotImplemented`, raised by the contract fallbacks.