Physlib

Physlib.Mathematics.VariationalCalculus.HasVarAdjoint

Variational adjoint

Definition of adjoint of linear function between function spaces. It is inspired by the definition of distributional adjoint of linear maps between test functions as described here: https://en.wikipedia.org/wiki/Distribution_(mathematics) under 'Preliminaries: Transpose of a linear operator' but we require that the adjoint is function between test functions too.

The key results are: - variational adjoint is unique on test functions - variational adjoint of identity is identity, `HasVarAdjoint.id` - variational adjoint of composition is composition of adjoint in reverse order, `HasVarAdjoint.comp` - variational adjoint of deriv is `- deriv`, `HasVarAdjoint.deriv` - variational adjoint of algebraic operations is algebraic operation of adjoints, `HasVarAdjoint.neg`, `HasVarAdjoint.add`, `HasVarAdjoint.sub`, `HasVarAdjoint.mul_left`, `HasVarAdjoint.mul_right`, `HasVarAdjoint.smul_left`, `HasVarAdjoint.smul_right`

27 declarations

theorem

Symmetry of the variational adjoint for localized transforms

Let F:(XU)(YV)F: (X \to U) \to (Y \to V) and F:(YV)(XU)F': (Y \to V) \to (X \to U) be transformations between function spaces. If FF' is the variational adjoint of FF, and FF is a localized function transform, then FF is the variational adjoint of FF'. A function transformation FF is said to be **localized** if for every compact set KYK \subseteq Y, there exists a compact set LXL \subseteq X such that for any two functions ϕ,ϕ:XU\phi, \phi' : X \to U, if ϕ(x)=ϕ(x)\phi(x) = \phi'(x) for all xLx \in L, then (Fϕ)(y)=(Fϕ)(y)(F\phi)(y) = (F\phi')(y) for all yKy \in K.

theorem

The variational adjoint of the identity is the identity

The identity transformation id:(XU)(XU)\text{id} : (X \to U) \to (X \to U), defined by id(ϕ)=ϕ\text{id}(\phi) = \phi for any function ϕ\phi, is its own variational adjoint.

theorem

The variational adjoint of the zero operator is the zero operator

The zero transformation F:(XU)(YV)F: (X \to U) \to (Y \to V), defined by F(ϕ)=0F(\phi) = 0 for all functions ϕ:XU\phi: X \to U, has a variational adjoint given by the zero transformation F:(YV)(XU)F': (Y \to V) \to (X \to U), defined by F(ψ)=0F'(\psi) = 0 for all functions ψ:YV\psi: Y \to V.

theorem

The variational adjoint of FGF \circ G is GFG' \circ F'

Let F:(YV)(ZW)F: (Y \to V) \to (Z \to W) and G:(XU)(YV)G: (X \to U) \to (Y \to V) be transformations between function spaces. Suppose FF has a variational adjoint FF', and GG has a variational adjoint GG'. Then the composition of these transformations FGF \circ G (where (FG)(ϕ)=F(G(ϕ))(F \circ G)(\phi) = F(G(\phi))) has a variational adjoint given by the composition of their adjoints in reverse order, GFG' \circ F' (where (GF)(ψ)=G(F(ψ))(G' \circ F')(\psi) = G'(F'(\psi))).

theorem

Equality on test functions implies the same variational adjoint

Let F,G:(XU)(YV)F, G: (X \to U) \to (Y \to V) and F:(YV)(XU)F': (Y \to V) \to (X \to U) be operators between function spaces. If FF' is the variational adjoint of GG, and for every test function ϕ\phi, the equality F(ϕ)=G(ϕ)F(\phi) = G(\phi) holds, then FF' is also the variational adjoint of FF.

theorem

Equality on test functions implies variational adjointness for localized transforms

Let F:(XU)(YV)F: (X \to U) \to (Y \to V) and F,G:(YV)(XU)F', G': (Y \to V) \to (X \to U) be transformations between function spaces. Suppose that FF' is a variational adjoint of FF, and that for every test function ϕ:YV\phi: Y \to V, the condition F(ϕ)=G(ϕ)F'(\phi) = G'(\phi) holds. If GG' is a localized function transform, then GG' is also a variational adjoint of FF. A function transformation G:(YV)(XU)G': (Y \to V) \to (X \to U) is said to be **localized** if for every compact set KXK \subseteq X, there exists a compact set LYL \subseteq Y such that for any two functions ϕ,ϕ:YV\phi, \phi': Y \to V, if ϕ\phi and ϕ\phi' agree on LL (i.e., ϕ(y)=ϕ(y)\phi(y) = \phi'(y) for all yLy \in L), then GϕG'\phi and GϕG'\phi' agree on KK (i.e., (Gϕ)(x)=(Gϕ)(x)(G'\phi)(x) = (G'\phi')(x) for all xKx \in K).

theorem

Uniqueness of variational adjoint on test functions

Let XX be a space equipped with a measure μ\mu (the volume measure) such that open sets are measurable, μ\mu is finite on compact sets, and μ\mu is strictly positive on non-empty open sets. Let FF be an operator mapping functions from XX to UU to functions from YY to VV. If FF' and GG' are both variational adjoints of FF, then for every test function ϕ:YV\phi: Y \to V (a smooth function with compact support), the equality F(ϕ)=G(ϕ)F'(\phi) = G'(\phi) holds.

theorem

Variational Adjoints of FF are Equal on Smooth Functions

Let XX and YY be real inner product spaces equipped with measures, and let YY be finite-dimensional. Assume that the volume measure on XX is finite on compact sets and strictly positive on non-empty open sets. Let FF be an operator mapping functions from XX to UU to functions from YY to VV. If FF' and GG' are both variational adjoints of FF, then for any smooth (CC^\infty) function f:YVf: Y \to V, it holds that F(f)=G(f)F'(f) = G'(f).

theorem

The variational adjoint of F-F is F-F'

Let FF be a function transformation from functions on XX to functions on VV, and let FF' be its variational adjoint. Then the transformation defined by (F)(ϕ)=F(ϕ)(-F)(\phi) = -F(\phi) has a variational adjoint defined by (F)(ψ)=F(ψ)(-F')(\psi) = -F'(\psi).

theorem

If F-F has variational adjoint F-F', then FF has variational adjoint FF'

Let FF be a function transformation mapping functions from XX to UU into functions from XX to VV, and let FF' be a function transformation mapping functions from XX to VV into functions from XX to UU. If the transformation defined by (F)(ϕ)=F(ϕ)(-F)(\phi) = -F(\phi) has the variational adjoint (F)(ψ)=F(ψ)(-F')(\psi) = -F'(\psi), then FF has the variational adjoint FF'.

theorem

The variational adjoint of F+GF + G is F+GF' + G'

Let FF and GG be function transformations mapping functions from XX to UU into functions from XX to VV. If FF has a variational adjoint FF' and GG has a variational adjoint GG', then the pointwise sum transformation (F+G)(F + G), defined by ((F+G)ϕ)(x)=Fϕ(x)+Gϕ(x)((F + G)\phi)(x) = F\phi(x) + G\phi(x), has the variational adjoint (F+G)(F' + G'), defined by ((F+G)ψ)(x)=Fψ(x)+Gψ(x)((F' + G')\psi)(x) = F'\psi(x) + G'\psi(x).

theorem

The variational adjoint of Fi\sum F_i is Fi\sum F'_i

Let ι\iota be a finite index set. For each iιi \in \iota, let FiF_i be an operator mapping functions from XX to UU to functions from XX to VV, and let FiF'_i be its corresponding variational adjoint mapping functions from XX to VV to functions from XX to UU. If for every iιi \in \iota, FiF_i has the variational adjoint FiF'_i, then the sum of these operators iιFi\sum_{i \in \iota} F_i has a variational adjoint given by the sum of their individual adjoints iιFi\sum_{i \in \iota} F'_i. Specifically, the operator mapping ϕ\phi to the pointwise sum iιFi(ϕ)(x)\sum_{i \in \iota} F_i(\phi)(x) has the variational adjoint mapping ψ\psi to iιFi(ψ)(x)\sum_{i \in \iota} F'_i(\psi)(x).

theorem

The variational adjoint of FGF - G is FGF' - G'

Let FF and GG be transformations mapping functions from XX to UU into functions from XX to VV. If FF has a variational adjoint FF' and GG has a variational adjoint GG', then the pointwise difference transformation (FG)(F - G), defined by ((FG)ϕ)(x)=Fϕ(x)Gϕ(x)((F - G)\phi)(x) = F\phi(x) - G\phi(x), has the variational adjoint (FG)(F' - G'), defined by ((FG)ψ)(x)=Fψ(x)Gψ(x)((F' - G')\psi)(x) = F'\psi(x) - G'\psi(x).

theorem

Variational Adjoint of ϕψF(ϕ)\phi \mapsto \psi F(\phi) is φF(ψφ)\varphi \mapsto F'(\psi \varphi)

Suppose FF is an operator that maps functions ϕ:XU\phi: X \to U to functions F(ϕ):XRF(\phi): X \to \mathbb{R}, and FF' is its variational adjoint. If ψ:XR\psi: X \to \mathbb{R} is a smooth (CC^\infty) function, then the operator mapping ϕ\phi to the pointwise product ψF(ϕ)\psi F(\phi) has a variational adjoint which maps a function φ\varphi to F(ψφ)F'(\psi \varphi), where ψφ\psi \varphi is the pointwise product of ψ\psi and φ\varphi.

theorem

Variational adjoint of right multiplication by a smooth function

Let XX and UU be spaces. Let F:(XU)(XR)F: (X \to U) \to (X \to \mathbb{R}) be an operator that possesses a variational adjoint F:(XR)(XU)F': (X \to \mathbb{R}) \to (X \to U). If ψ:XR\psi: X \to \mathbb{R} is a smooth function (i.e., ψ\psi is CC^\infty), then the operator defined by the pointwise product ϕF(ϕ)ψ\phi \mapsto F(\phi) \cdot \psi also has a variational adjoint. This adjoint is given by the mapping φF(φψ)\varphi \mapsto F'(\varphi \cdot \psi), where (φψ)(x)=φ(x)ψ(x)(\varphi \cdot \psi)(x) = \varphi(x) \psi(x).

theorem

Variational adjoint of ϕψF(ϕ)\phi \mapsto \psi \cdot F(\phi) is φF(ψφ)\varphi \mapsto F'(\psi \cdot \varphi)

Let FF be an operator mapping functions ϕ:XU\phi: X \to U to functions F(ϕ):XVF(\phi): X \to V, and let FF' be its variational adjoint. If ψ:XR\psi: X \to \mathbb{R} is a smooth (CC^\infty) function, then the operator that maps ϕ\phi to the pointwise scalar product ψF(ϕ)\psi \cdot F(\phi) has a variational adjoint. This adjoint maps a function φ:XV\varphi: X \to V to F(ψφ)F'(\psi \cdot \varphi), where (ψφ)(x)=ψ(x)φ(x)(\psi \cdot \varphi)(x) = \psi(x) \cdot \varphi(x).

theorem

Variational adjoint of scalar multiplication by a smooth function

Let F:(XU)(XV)F: (X \to U) \to (X \to V) be an operator that possesses a variational adjoint F:(XV)(XU)F': (X \to V) \to (X \to U). If ψ:XR\psi: X \to \mathbb{R} is a smooth (CC^\infty) function, then the operator defined by the pointwise scalar multiplication ϕψF(ϕ)\phi \mapsto \psi \cdot F(\phi) also has a variational adjoint. This adjoint is given by the mapping φF(ψφ)\varphi \mapsto F'(\psi \cdot \varphi), where (ψφ)(x)=ψ(x)φ(x)(\psi \cdot \varphi)(x) = \psi(x) \cdot \varphi(x).

theorem

The variational adjoint of xf(x)(ϕ(x))x \mapsto f(x)(\phi(x)) is xf(x)(ψ(x))x \mapsto f(x)^\dagger(\psi(x))

Let UU and VV be complete generalized inner product spaces over R\mathbb{R}. Let f:XL(U,V)f: X \to \mathcal{L}(U, V) be an infinitely differentiable (CC^\infty) function that maps each point xXx \in X to a continuous linear map f(x)L(U,V)f(x) \in \mathcal{L}(U, V). Then the operator that maps a function ϕ:XU\phi: X \to U to the function (fϕ)(x)=f(x)(ϕ(x))(f\phi)(x) = f(x)(\phi(x)) has a variational adjoint. This variational adjoint maps a function ψ:XV\psi: X \to V to the function defined by xf(x)(ψ(x))x \mapsto f(x)^\dagger(\psi(x)), where f(x):VUf(x)^\dagger: V \to U is the adjoint of the continuous linear map f(x)f(x).

theorem

The variational adjoint of the derivative is the negative derivative

Let UU be a real generalized inner product space. The derivative operator, which maps a function ϕ:RU\phi: \mathbb{R} \to U to its derivative ϕ(x)=dϕdx(x)\phi'(x) = \frac{d\phi}{dx}(x), possesses a variational adjoint. This adjoint operator maps a function ψ:RU\psi: \mathbb{R} \to U to its negative derivative ψ(x)=dψdx(x)-\psi'(x) = -\frac{d\psi}{dx}(x). This relationship corresponds to the identity ϕ(x),ψ(x)dx=ϕ(x),ψ(x)dx\int_{-\infty}^{\infty} \langle \phi'(x), \psi(x) \rangle \, dx = \int_{-\infty}^{\infty} \langle \phi(x), -\psi'(x) \rangle \, dx for all test functions ϕ\phi and ψ\psi.

theorem

The variational adjoint of D()(v)\text{D}(\cdot)(v) is D()(v)-\text{D}(\cdot)(v)

Let XX be a finite-dimensional real vector space equipped with an additive Haar measure (volume), and let UU be a real generalized inner product space. For any fixed vector vXv \in X, the linear operator LL that maps a test function ϕ:XU\phi: X \to U to its directional Fréchet derivative (Lϕ)(x)=Dϕ(x)(v)(L\phi)(x) = \text{D}\phi(x)(v) has a variational adjoint LL^\dagger. This adjoint operator is given by (Lψ)(x)=Dψ(x)(v)(L^\dagger\psi)(x) = -\text{D}\psi(x)(v), where Dψ(x)\text{D}\psi(x) denotes the Fréchet derivative of the test function ψ\psi at xx.

theorem

The variational adjoint of ϕ(Dϕ)δy\phi \mapsto (D\phi)^\dagger \delta y is ψ(div ψ)δy\psi \mapsto -(\text{div } \psi) \delta y

Let XX and YY be real generalized inner product spaces, where XX is finite-dimensional, proper, and equipped with an additive Haar measure. For a fixed vector δyY\delta y \in Y, let LL be the linear operator that maps a function ϕ:XY\phi: X \to Y to the function x(Dϕ(x))(δy)x \mapsto (D\phi(x))^\dagger(\delta y), where (Dϕ(x)):YX(D\phi(x))^\dagger: Y \to X is the adjoint of the Fréchet derivative of ϕ\phi at xx. The variational adjoint of LL is the operator LL^* that maps a function ψ:XX\psi: X \to X to the function x(div ψ(x))δyx \mapsto -(\text{div } \psi(x)) \delta y. This relationship is characterized by the identity: X(Dϕ(x))(δy),ψ(x)Xdx=Xϕ(x),(div ψ(x))δyYdx \int_X \langle (D\phi(x))^\dagger(\delta y), \psi(x) \rangle_X \, dx = \int_X \langle \phi(x), -(\text{div } \psi(x)) \delta y \rangle_Y \, dx for all test functions ϕ\phi and ψ\psi.

theorem

The variational adjoint of \nabla is divbasis.repr-\text{div} \circ \text{basis.repr}

Let XX be the dd-dimensional real inner product space denoted by `Space d`. Let LL be the operator that maps a scalar function ϕ:XR\phi: X \to \mathbb{R} to its gradient field ϕ:XX\nabla \phi: X \to X. The variational adjoint of LL is the operator LL^* that maps a vector field ψ:XX\psi: X \to X to the scalar function xdiv(repr(ψ(x)))x \mapsto -\text{div}(\text{repr}(\psi(x))), where repr\text{repr} denotes the representation of a vector in the canonical basis of XX. This relationship is defined by the integral identity: Xϕ(x),ψ(x)Xdx=Xϕ(x)(div(reprψ)(x))dx \int_X \langle \nabla \phi(x), \psi(x) \rangle_X \, dx = \int_X \phi(x) \left( -\text{div}(\text{repr} \circ \psi)(x) \right) \, dx for all test functions ϕ\phi and ψ\psi.

theorem

The variational adjoint of \nabla is div-\text{div}

Let XX be a dd-dimensional real inner product space (denoted as `Space d`). The operator that maps a scalar function ϕ:XR\phi: X \to \mathbb{R} to its gradient field ϕ:XX\nabla \phi: X \to X has a variational adjoint. This adjoint maps a vector field ψ:XX\psi: X \to X to the scalar function defined by xdiv ψ(x)x \mapsto -\text{div } \psi(x), where div\text{div} is the divergence operator. This relationship is characterized by the integral identity: Xϕ(x),ψ(x)dx=Xϕ(x)(div ψ(x))dx\int_X \langle \nabla \phi(x), \psi(x) \rangle \, dx = \int_X \phi(x) (-\text{div } \psi(x)) \, dx for all appropriate test functions ϕ\phi and ψ\psi.

theorem

The variational adjoint of div\text{div} is -\nabla

Let XX be a dd-dimensional real inner product space (denoted as `Space d`). The divergence operator, which maps a vector field ϕ:XRd\phi: X \to \mathbb{R}^d to a scalar function div ϕ:XR\text{div } \phi: X \to \mathbb{R}, has a variational adjoint. This adjoint maps a scalar function ψ:XR\psi: X \to \mathbb{R} to the vector field defined by xψ(x)x \mapsto -\nabla \psi(x), where \nabla is the gradient operator. This relationship is characterized by the integral identity: X(div ϕ(x))ψ(x)dx=Xϕ(x),ψ(x)dx\int_X (\text{div } \phi(x)) \psi(x) \, dx = \int_X \langle \phi(x), -\nabla \psi(x) \rangle \, dx for all appropriate test functions ϕ\phi and ψ\psi.

theorem

The variational adjoint of (F,G)(F, G) is F(proj1)+G(proj2)F'(\text{proj}_1) + G'(\text{proj}_2)

Let XX be a measurable space where open sets are measurable, and let the volume measure on XX be finite on compact sets. Let F:(XU)(XV)F: (X \to U) \to (X \to V) and G:(XU)(XW)G: (X \to U) \to (X \to W) be operators with variational adjoints FF' and GG', respectively. Then the product operator HH, which maps a function ϕ:XU\phi: X \to U to the pair of functions (Fϕ,Gϕ)(F\phi, G\phi), has a variational adjoint HH'. For any function ψ:XV×W\psi: X \to V \times W, this adjoint is given by H(ψ)=F(ψ1)+G(ψ2)H'(\psi) = F'(\psi_1) + G'(\psi_2) where ψ1(x)\psi_1(x) and ψ2(x)\psi_2(x) are the first and second components of ψ(x)\psi(x) in V×WV \times W, respectively.

theorem

Variational Adjoint of the First Component of an Operator

Let XX be a space and U,W,VU, W, V be target spaces. Let FF be an operator that maps functions ϕ:XU\phi: X \to U to functions F(ϕ):XW×VF(\phi): X \to W \times V. If FF has a variational adjoint FF', then the operator G:(XU)(XW)G: (X \to U) \to (X \to W) defined by the first component of FF, G(ϕ)(x)=(F(ϕ)(x))1G(\phi)(x) = (F(\phi)(x))_1, also has a variational adjoint GG'. For any function ϕ:XW\phi: X \to W, this adjoint is given by G(ϕ)=F(ψ)G'(\phi) = F'(\psi), where ψ:XW×V\psi: X \to W \times V is the function defined by ψ(x)=(ϕ(x),0)\psi(x') = (\phi(x'), 0) for all xXx' \in X.

theorem

Variational Adjoint of the Second Component Operator

Let FF be an operator that maps functions ϕ:XU\phi: X \to U to functions Fϕ:XW×VF\phi: X \to W \times V. Suppose FF has a variational adjoint FF'. Then the operator F2F_2 defined by F2(ϕ)(x)=(Fϕ(x))2F_2(\phi)(x) = (F\phi(x))_2 (the projection onto the second component of the output) also has a variational adjoint. This adjoint maps a test function ψ:XV\psi: X \to V to F(ψ~)F'(\tilde{\psi}), where ψ~:XW×V\tilde{\psi}: X \to W \times V is the function defined by ψ~(x)=(0,ψ(x))\tilde{\psi}(x) = (0, \psi(x)).