Physlib

Physlib.Mathematics.VariationalCalculus.HasVarAdjDeriv

Variational adjoint derivative

Variational adjoint derivative of `F` at `u` is a generalization of `(fderiv ℝ F u).adjoint` to function spaces. In particular, variational gradient is an analog of `gradient F u := (fderiv ℝ F u).adjoint 1`.

The definition of `HasVarAdjDerivAt` is constructed exactly such that we can prove composition theorem saying ``` HasVarAdjDerivAt F F' (G u)) → HasVarAdjDerivAt G G' u → HasVarAdjDerivAt (F ∘ G) (G' ∘ F') u ``` This theorem is the main tool to mechanistically compute variational gradient.

32 declarations

theorem

Smoothness of F(v)F(v) from existence of variational adjoint derivative of FF

Let F:(X→U)→(X→V)F: (X \to U) \to (X \to V) be an operator between function spaces that possesses a variational adjoint derivative F′F' at uu. If v:X→Uv: X \to U is a smooth function (of class C∞C^\infty), then its image under the operator FF, given by F(v):X→VF(v): X \to V, is also a smooth function.

theorem

Smoothness of F(u)F(u) from existence of variational adjoint derivative at uu

Let F:(X→U)→(X→V)F: (X \to U) \to (X \to V) be an operator between function spaces. Suppose FF possesses a variational adjoint derivative F′F' at a function u:X→Uu: X \to U. Then the function F(u):X→VF(u): X \to V is infinitely differentiable (of class C∞C^\infty).

theorem

s↦F(ϕ(s,⋅))(x)s \mapsto F(\phi(s, \cdot))(x) is C∞C^\infty for smooth families ϕ\phi and operators with variational adjoint derivatives

Suppose an operator F:(X→U)→(X→V)F: (X \to U) \to (X \to V) has a variational adjoint derivative at a function u:X→Uu: X \to U. Let ϕ:R×X→U\phi: \mathbb{R} \times X \to U be an infinitely differentiable (C∞C^\infty) family of functions, where the mapping (s,x)↦ϕ(s,x)(s, x) \mapsto \phi(s, x) is smooth. Then, for any fixed point x∈Xx \in X, the function mapping the parameter s∈Rs \in \mathbb{R} to the value of the operator FF evaluated at the function ϕ(s,⋅)\phi(s, \cdot) and then evaluated at xx, given by s↦F(ϕ(s,⋅))(x), s \mapsto F(\phi(s, \cdot))(x), is infinitely differentiable (C∞C^\infty) with respect to ss.

theorem

s↦F(u0+sδu)(x)s \mapsto F(u_0 + s \delta u)(x) is C∞C^\infty for operators with variational adjoint derivatives

Suppose an operator F:(X→U)→(X→V)F: (X \to U) \to (X \to V) has a variational adjoint derivative at u:X→Uu: X \to U. Let ϕ:R×X→U\phi: \mathbb{R} \times X \to U be an infinitely differentiable (C∞C^\infty) family of functions. Let u0:X→Uu_0: X \to U be the function u0(x)=ϕ(0,x)u_0(x) = \phi(0, x) and let δu:X→U\delta u: X \to U be the partial derivative of ϕ\phi with respect to its first argument at 00, given by δu(x)=∂ϕ(s,x)∂s∣s=0\delta u(x) = \left. \frac{\partial \phi(s, x)}{\partial s} \right|_{s=0}. Then, for any fixed point x∈Xx \in X, the function mapping s′∈Rs' \in \mathbb{R} to the value of the operator FF evaluated at the linear variation u0+s′δuu_0 + s' \delta u and then evaluated at xx, given by s′↦F(u0+s′δu)(x), s' \mapsto F(u_0 + s' \delta u)(x), is infinitely differentiable (C∞C^\infty) with respect to s′s'.

theorem

s↦F(u+sδu)(x)s \mapsto F(u + s \delta u)(x) is C∞C^\infty for operators with variational adjoint derivatives

Suppose an operator F:(X→U)→(X→V)F: (X \to U) \to (X \to V) has a variational adjoint derivative at the function u:X→Uu: X \to U. Then, for any infinitely differentiable (smooth) function δu:X→U\delta u: X \to U and any fixed point x∈Xx \in X, the function mapping s∈Rs \in \mathbb{R} to the value of the operator at the perturbed function u+sδuu + s \delta u evaluated at xx, given by s↦F(u+sδu)(x), s \mapsto F(u + s \delta u)(x), is infinitely differentiable (C∞C^\infty) with respect to ss.

theorem

s↦F(u0+sδu)(x)s \mapsto F(u_0 + s \delta u)(x) is differentiable for operators with variational adjoint derivatives

Suppose an operator F:(X→U)→(X→V)F: (X \to U) \to (X \to V) has a variational adjoint derivative at u:X→Uu: X \to U. Let ϕ:R×X→U\phi: \mathbb{R} \times X \to U be an infinitely differentiable (C∞C^\infty) family of functions. Let u0:X→Uu_0: X \to U be the function u0(x)=ϕ(0,x)u_0(x) = \phi(0, x) and let δu:X→U\delta u: X \to U be the partial derivative of ϕ\phi with respect to its first argument at s=0s=0, given by δu(x)=∂ϕ(s,x)∂s∣s=0\delta u(x) = \left. \frac{\partial \phi(s, x)}{\partial s} \right|_{s=0}. Then, for any fixed point x∈Xx \in X, the function mapping s′∈Rs' \in \mathbb{R} to the value of the operator FF evaluated at the linear variation u0+s′δuu_0 + s' \delta u and then evaluated at xx, given by s′↦F(u0+s′δu)(x), s' \mapsto F(u_0 + s' \delta u)(x), is differentiable with respect to s′s'.

theorem

ddsF(ϕ(s))=ddsF(ϕ(0)+sϕ˙(0))\frac{d}{ds} F(\phi(s)) = \frac{d}{ds} F(\phi(0) + s \dot{\phi}(0)) for linear operators FF

Let F:(X→U)→(X→V)F: (X \to U) \to (X \to V) be an operator mapping between function spaces. Suppose FF is linear on smooth functions, meaning that for all smooth ϕ1,ϕ2:X→U\phi_1, \phi_2: X \to U and c∈Rc \in \mathbb{R}, F(ϕ1+ϕ2)=F(ϕ1)+F(ϕ2)F(\phi_1 + \phi_2) = F(\phi_1) + F(\phi_2) and F(c⋅ϕ1)=c⋅F(ϕ1)F(c \cdot \phi_1) = c \cdot F(\phi_1). Additionally, assume that for any smooth family of functions ϕ:R×X→U\phi: \mathbb{R} \times X \to U, the derivative at s=0s=0 commutes with FF at any point x∈Xx \in X: dds∣s=0F(ϕ(s,⋅))(x)=F(∂ϕ∂s∣s=0)(x). \left. \frac{d}{ds} \right|_{s=0} F(\phi(s, \cdot))(x) = F\left( \left. \frac{\partial \phi}{\partial s} \right|_{s=0} \right)(x). Then, for any such smooth ϕ\phi and any x∈Xx \in X, the derivative of FF applied to the path ϕ(s,⋅)\phi(s, \cdot) is equal to the derivative of FF applied to the first-order linear approximation of ϕ\phi at s=0s=0: dds∣s=0F(ϕ(s,⋅))(x)=dds∣s=0F(y↦ϕ(0,y)+s⋅∂ϕ∂s∣s=0(y))(x). \left. \frac{d}{ds} \right|_{s=0} F(\phi(s, \cdot))(x) = \left. \frac{d}{ds} \right|_{s=0} F\left( y \mapsto \phi(0, y) + s \cdot \left. \frac{\partial \phi}{\partial s} \right|_{s=0}(y) \right)(x).

theorem

The variational adjoint of the linearization of a linear operator FF is its adjoint F′F'

Let F:(X→U)→(X→V)F: (X \to U) \to (X \to V) be an operator between function spaces that is linear on the space of smooth functions, meaning F(ϕ1+ϕ2)=F(ϕ1)+F(ϕ2)F(\phi_1 + \phi_2) = F(\phi_1) + F(\phi_2) and F(c⋅ϕ)=c⋅F(ϕ)F(c \cdot \phi) = c \cdot F(\phi) for any c∈Rc \in \mathbb{R} and smooth functions ϕ,ϕ1,ϕ2:X→U\phi, \phi_1, \phi_2: X \to U. Suppose u:X→Uu: X \to U is a smooth function and FF has a variational adjoint F′F'. Then the linear operator that maps a variation δu:X→U\delta u: X \to U to the directional derivative of FF at uu, defined by δu↦(x↦dds∣s=0F(u+sδu)(x)), \delta u \mapsto \left( x \mapsto \left. \frac{d}{ds} \right|_{s=0} F(u + s \delta u)(x) \right), has F′F' as its variational adjoint.

theorem

The Variational Adjoint Derivative of a Linear Operator is its Variational Adjoint

Let F:(X→U)→(X→V)F: (X \to U) \to (X \to V) be an operator between function spaces. Suppose FF satisfies the following conditions: 1. **Linearity:** FF is linear on the space of smooth functions, i.e., F(ϕ1+ϕ2)=F(ϕ1)+F(ϕ2)F(\phi_1 + \phi_2) = F(\phi_1) + F(\phi_2) and F(c⋅ϕ)=c⋅F(ϕ)F(c \cdot \phi) = c \cdot F(\phi) for any c∈Rc \in \mathbb{R} and smooth functions ϕ,ϕ1,ϕ2:X→U\phi, \phi_1, \phi_2: X \to U. 2. **Smoothness:** FF maps any smooth family of functions ϕ:R×X→U\phi: \mathbb{R} \times X \to U to a smooth function on R×X\mathbb{R} \times X defined by (s,x)↦F(ϕ(s,⋅))(x)(s, x) \mapsto F(\phi(s, \cdot))(x). 3. **Commutativity of Differentiation:** For any smooth family ϕ:R×X→U\phi: \mathbb{R} \times X \to U, the derivative at s=0s=0 commutes with FF at any point x∈Xx \in X: dds∣s=0F(ϕ(s,⋅))(x)=F(∂ϕ∂s∣s=0)(x). \left. \frac{d}{ds} \right|_{s=0} F(\phi(s, \cdot))(x) = F\left( \left. \frac{\partial \phi}{\partial s} \right|_{s=0} \right)(x). If u:X→Uu: X \to U is a smooth function and FF has a variational adjoint F′F', then the variational adjoint derivative of FF at uu is F′F'.

theorem

The variational adjoint derivative of the identity operator is the identity operator

Let u:X→Uu: X \to U be a C∞C^\infty smooth function. The identity operator F:(X→U)→(X→U)F : (X \to U) \to (X \to U) defined by F(ϕ)=ϕF(\phi) = \phi has a variational adjoint derivative at uu given by the identity operator F′(ψ)=ψF'(\psi) = \psi.

theorem

Variational Adjoint Derivative of a Constant Functional is Zero

Let u:X→Uu: X \to U and v:X→Vv: X \to V be C∞C^\infty smooth functions. The constant functional F:(X→U)→(X→V)F: (X \to U) \to (X \to V) defined by F(ϕ)=vF(\phi) = v for all ϕ\phi has a variational adjoint derivative at uu given by the zero operator F′(ψ)=0F'(\psi) = 0.

theorem

Chain Rule for Variational Adjoint Derivatives: (F∘G)′=G′∘F′(F \circ G)' = G' \circ F'

Let F:(Y→V)→(Z→W)F: (Y \to V) \to (Z \to W) and G:(X→U)→(Y→V)G: (X \to U) \to (Y \to V) be operators between function spaces, and let u:X→Uu: X \to U be a function. If FF has a variational adjoint derivative F′F' at the point G(u)G(u) and GG has a variational adjoint derivative G′G' at the point uu, then the composition F∘GF \circ G has a variational adjoint derivative at uu given by the composition of the adjoint derivatives in reverse order, G′∘F′G' \circ F'.

theorem

Agreement on Smooth Functions Preserves Variational Adjoint Derivative

Let F,G:(X→U)→(Y→V)F, G: (X \to U) \to (Y \to V) be functionals and u:X→Uu: X \to U be a function. Suppose that FF has a variational adjoint derivative F′F' at uu (denoted by HasVarAdjDerivAt F F′ u\text{HasVarAdjDerivAt } F \ F' \ u). If FF and GG agree on all smooth functions, i.e., F(ϕ)=G(ϕ)F(\phi) = G(\phi) for every smooth function ϕ:X→U\phi: X \to U (ϕ∈C∞\phi \in C^\infty), then GG also has F′F' as its variational adjoint derivative at uu.

theorem

Uniqueness of Variational Adjoint Derivative on Test Functions

Let XX and YY be spaces, where XX is equipped with a measure that is finite on compact sets and strictly positive on non-empty open sets. Let F:(X→U)→(Y→V)F: (X \to U) \to (Y \to V) be a functional. If F′F' and G′G' are both variational adjoint derivatives of FF at u:X→Uu: X \to U (satisfying the `HasVarAdjDerivAt` condition), then for any test function ϕ:Y→V\phi: Y \to V (i.e., a smooth function with compact support), it holds that F′ϕ=G′ϕF' \phi = G' \phi.

theorem

Uniqueness of Variational Adjoint Derivative on Smooth Functions

Let XX and YY be normed real inner product spaces equipped with measures, where XX has a measure that is finite on compact sets and strictly positive on non-empty open sets, and YY is finite-dimensional. Let F:(X→U)→(Y→V)F: (X \to U) \to (Y \to V) be a functional. If FF has two variational adjoint derivatives F′F' and G′G' at a point u:X→Uu: X \to U (denoted by HasVarAdjDerivAt F F′ u\text{HasVarAdjDerivAt } F \ F' \ u and HasVarAdjDerivAt F G′ u\text{HasVarAdjDerivAt } F \ G' \ u), then for every smooth function ϕ:Y→V\phi: Y \to V (ϕ∈C∞\phi \in C^\infty), it holds that F′ϕ=G′ϕF' \phi = G' \phi.

theorem

Variational Adjoint Derivative of the Product Operator (F,G)(F, G)

Let XX be a space equipped with a measure that is finite on compact sets. Let F:(X→U)→(X→V)F: (X \to U) \to (X \to V) and G:(X→U)→(X→W)G: (X \to U) \to (X \to W) be operators. Suppose FF has a variational adjoint derivative F′F' at the function u:X→Uu: X \to U, and GG has a variational adjoint derivative G′G' at uu. Then the product operator H:(X→U)→(X→V×W)H: (X \to U) \to (X \to V \times W), defined by H(ϕ)(x)=(F(ϕ)(x),G(ϕ)(x))H(\phi)(x) = (F(\phi)(x), G(\phi)(x)), has a variational adjoint derivative H′H' at uu given by: H′(ψ)(x)=F′(ψ1)(x)+G′(ψ2)(x) H'(\psi)(x) = F'(\psi_1)(x) + G'(\psi_2)(x) where ψ1:X→V\psi_1: X \to V and ψ2:X→W\psi_2: X \to W are the component functions of ψ:X→V×W\psi: X \to V \times W such that ψ(x)=(ψ1(x),ψ2(x))\psi(x) = (\psi_1(x), \psi_2(x)).

theorem

Variational adjoint of the first component of an operator FF is F′F' applied to the zero-padded input

Suppose an operator F:(X→U)→(X→W×V)F: (X \to U) \to (X \to W \times V) has a variational adjoint derivative F′F' at a function u:X→Uu: X \to U. Then the operator that maps a function ϕ:X→U\phi: X \to U to the first component of F(ϕ)F(\phi), given by x↦(F(ϕ)(x))1x \mapsto (F(\phi)(x))_1, also has a variational adjoint derivative at uu. This variational adjoint derivative is the operator that maps a function ψ:X→W\psi: X \to W to the value of F′F' evaluated at the function x′↦(ψ(x′),0)x' \mapsto (\psi(x'), 0).

theorem

Variational adjoint of the second component of an operator FF is F′F' applied to the zero-prefixed input

Suppose an operator F:(X→U)→(X→W×V)F: (X \to U) \to (X \to W \times V) has a variational adjoint derivative F′F' at a function u:X→Uu: X \to U. Then the operator that maps a function ϕ:X→U\phi: X \to U to the second component of F(ϕ)F(\phi), given by x↦(F(ϕ)(x))2x \mapsto (F(\phi)(x))_2, also has a variational adjoint derivative at uu. This variational adjoint derivative is the operator that maps a function ψ:X→V\psi: X \to V to the value of F′F' evaluated at the function x′↦(0,ψ(x′))x' \mapsto (0, \psi(x')).

theorem

The variational adjoint of the derivative operator is the negative derivative operator

Let u:R→Uu: \mathbb{R} \to U be a C∞C^\infty smooth function. Let FF be the operator that maps a function ϕ:R→U\phi: \mathbb{R} \to U to its derivative dϕdx\frac{d\phi}{dx}. Then the variational adjoint derivative of FF at uu is the operator that maps a function ϕ\phi to its negative derivative −dϕdx-\frac{d\phi}{dx}.

theorem

The variational adjoint derivative of ddxF\frac{d}{dx} F is F′∘(−ddx)F' \circ (-\frac{d}{dx})

Let F:(R→U)→(R→V)F: (\mathbb{R} \to U) \to (\mathbb{R} \to V) be an operator between function spaces. If FF has a variational adjoint derivative F′F' at the function u:R→Uu: \mathbb{R} \to U, then the operator mapping ϕ\phi to the derivative ddx(F(ϕ))\frac{d}{dx}(F(\phi)) has a variational adjoint derivative at uu given by the mapping ψ↦F′(−dψdx)\psi \mapsto F'(-\frac{d\psi}{dx}).

theorem

Variational Adjoint Derivative of Pointwise Mapping (fmap)

Let UU and VV be complete generalized inner product spaces over R\mathbb{R}. Let u:X→Uu: X \to U be a smooth (C∞C^\infty) function. Let f:X→U→Vf: X \to U \to V be a map such that the uncurried function (x,y)↦f(x,y)(x, y) \mapsto f(x, y) is C∞C^\infty, and for each x∈Xx \in X, the map f(x,⋅):U→Vf(x, \cdot): U \to V has an adjoint Fréchet derivative f′(x,u(x))f'(x, u(x)) at the point u(x)u(x). Consider the operator FF acting on functions ϕ:X→U\phi: X \to U defined by pointwise application: (Fϕ)(x)=f(x,ϕ(x)).(F\phi)(x) = f(x, \phi(x)). Then, the variational adjoint derivative of FF at uu is the operator that maps a function ψ:X→V\psi: X \to V to the function: x↦f′(x,u(x))(ψ(x)).x \mapsto f'(x, u(x))(\psi(x)).

theorem

The variational adjoint derivative of −F-F is −F′-F'

Suppose FF is an operator mapping functions from X→UX \to U to functions from X→VX \to V. If F′F' is the variational adjoint derivative of FF at uu, then the operator mapping ϕ\phi to −F(ϕ)-F(\phi) has a variational adjoint derivative at uu given by the mapping ψ↦−F′(ψ)\psi \mapsto -F'(\psi).

theorem

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

Suppose FF and GG are operators mapping functions from X→UX \to U to functions from X→VX \to V. If FF has a variational adjoint derivative F′F' at u:X→Uu: X \to U, and GG has a variational adjoint derivative G′G' at uu, then the sum operator defined by ϕ↦F(ϕ)+G(ϕ)\phi \mapsto F(\phi) + G(\phi) has a variational adjoint derivative at uu given by the operator ψ↦F′(ψ)+G′(ψ)\psi \mapsto F'(\psi) + G'(\psi).

theorem

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

Let ι\iota be a finite index set. For each i∈ιi \in \iota, let FiF_i be an operator mapping functions from X→UX \to U to functions from X→VX \to V. Suppose that for each i∈ιi \in \iota, FiF_i has a variational adjoint derivative Fi′F'_i at a C∞C^\infty smooth function u:X→Uu: X \to U. Then the sum operator SS, defined by (Sϕ)(x)=∑i∈ιFi(ϕ)(x)(S\phi)(x) = \sum_{i \in \iota} F_i(\phi)(x), has a variational adjoint derivative at uu given by the operator S′S', where for any function ψ:X→V\psi: X \to V, S′(ψ)(x)=∑i∈ιFi′(ψ)(x).S'(\psi)(x) = \sum_{i \in \iota} F'_i(\psi)(x).

theorem

Product Rule for Variational Adjoint Derivatives

Let F,G:(X→U)→(X→R)F, G: (X \to U) \to (X \to \mathbb{R}) be operators. Suppose FF and GG have variational adjoint derivatives F′F' and G′G' at a function u:X→Uu: X \to U, respectively. Then the pointwise product operator HH, defined by (Hϕ)(x)=F(ϕ)(x)⋅G(ϕ)(x)(H\phi)(x) = F(\phi)(x) \cdot G(\phi)(x), has a variational adjoint derivative at uu given by the operator H′H', where for any function ψ:X→R\psi: X \to \mathbb{R}, H′(ψ)(x)=F′(x↦ψ(x)G(u)(x))(x)+G′(x↦F(u)(x)ψ(x))(x).H'(\psi)(x) = F'(x \mapsto \psi(x)G(u)(x))(x) + G'(x \mapsto F(u)(x)\psi(x))(x). In more concise notation, the variational adjoint derivative is H′(ψ)=F′(ψ⋅G(u))+G′(ψ⋅F(u))H'(\psi) = F'(\psi \cdot G(u)) + G'(\psi \cdot F(u)), where the products inside the arguments are pointwise products of functions.

theorem

The Variational Adjoint Derivative of cFc F is F′(cψ)F'(c \psi)

Let F:(X→U)→(X→R)F: (X \to U) \to (X \to \mathbb{R}) be an operator. Suppose FF has a variational adjoint derivative F′F' at a function u:X→Uu: X \to U. For any constant scalar c∈Rc \in \mathbb{R}, the operator HH defined by the pointwise multiplication (Hϕ)(x)=c⋅F(ϕ)(x)(H\phi)(x) = c \cdot F(\phi)(x) has a variational adjoint derivative H′H' at uu, given by: H′(ψ)=F′(c⋅ψ)H'(\psi) = F'(c \cdot \psi) where c⋅ψc \cdot \psi denotes the function x′↦cψ(x′)x' \mapsto c \psi(x').

theorem

Variational Adjoint Derivative of Pointwise Multiplication by a Fixed Smooth Function ff

Let XX and UU be spaces. Let f:X→Rf: X \to \mathbb{R} be an infinitely differentiable function (f∈C∞(X,R)f \in C^\infty(X, \mathbb{R})). Suppose an operator F:(X→U)→(X→R)F: (X \to U) \to (X \to \mathbb{R}) possesses a variational adjoint derivative F′F' at a function u:X→Uu: X \to U. Then the operator HH, defined by the pointwise product (Hϕ)(x)=f(x)F(ϕ)(x)(H\phi)(x) = f(x)F(\phi)(x), has a variational adjoint derivative at uu given by the operator H′H', where for any function ψ:X→R\psi: X \to \mathbb{R}: H′(ψ)=F′(f⋅ψ)H'(\psi) = F'(f \cdot \psi) Here, f⋅ψf \cdot \psi denotes the pointwise product x′↦f(x′)ψ(x′)x' \mapsto f(x')\psi(x').

theorem

The variational adjoint derivative of the directional derivative Dϕ[dx]D\phi[dx] is −Dψ[dx]-D\psi[dx].

Let XX be a finite-dimensional real vector space equipped with an additive Haar measure (volume). Let u:X→Uu: X \to U be a smooth function. For a fixed vector dx∈Xdx \in X, consider the operator that maps a function ϕ:X→U\phi: X \to U to its Fréchet derivative at xx in the direction dxdx, denoted by Dϕ(x)[dx]D\phi(x)[dx]. The variational adjoint derivative of this operator at uu is the operator that maps a function ψ\psi to the negative of its directional derivative, −Dψ(x)[dx]-D\psi(x)[dx].

theorem

The variational adjoint derivative of D(F(ϕ))[dx]D(F(\phi))[dx] is F′(−Dψ[dx])F'(-D\psi[dx])

Let XX be a finite-dimensional real vector space equipped with an additive Haar measure. Let F:(X→U)→(X→V)F: (X \to U) \to (X \to V) be an operator between function spaces, and let F′F' be its variational adjoint derivative at a point uu. For a fixed vector dx∈Xdx \in X, consider the operator that maps a function ϕ\phi to the directional derivative of F(ϕ)F(\phi) in the direction dxdx, denoted by x↦D(F(ϕ))(x)[dx]x \mapsto D(F(\phi))(x)[dx]. The variational adjoint derivative of this operator at uu is given by the operator that maps a function ψ\psi to F′(ψ∗)F'(\psi^*), where ψ∗(x′)=−Dψ(x′)[dx]\psi^*(x') = -D\psi(x')[dx] is the negative directional derivative of ψ\psi.

theorem

The variational adjoint derivative of the gradient ∇ϕ\nabla \phi is −divψ-\text{div} \psi

Let u:Space d→Ru: \text{Space } d \to \mathbb{R} be an infinitely smooth (C∞C^\infty) function. Consider the operator that maps a scalar function ϕ:Space d→R\phi: \text{Space } d \to \mathbb{R} to its gradient field ∇ϕ\nabla \phi. The variational adjoint derivative of this gradient operator at uu is the operator that maps a vector field ψ\psi to the negative of its divergence, −divψ-\text{div} \psi.

theorem

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

Let u:Space d→Ru: \text{Space } d \to \mathbb{R} be an infinitely smooth (C∞C^\infty) function. The variational adjoint derivative of the gradient operator FF, defined by F(ϕ)=∇ϕF(\phi) = \nabla \phi for scalar functions ϕ:Space d→R\phi: \text{Space } d \to \mathbb{R}, at the point uu is the negative divergence operator F′F', defined by F′(ψ)=−divψF'(\psi) = -\text{div} \psi for vector fields ψ\psi.

theorem

The Variational Adjoint Derivative of div\text{div} is −∇-\nabla

Let dd be a natural number representing the dimension of the space, and let u:Space d→Rdu: \text{Space } d \to \mathbb{R}^d be a smooth (C∞C^\infty) function. The variational adjoint derivative of the divergence operator ϕ↦div ϕ\phi \mapsto \text{div } \phi at uu is the operator that maps a scalar field ψ\psi to its negative gradient −∇ψ-\nabla \psi.