# Changes between Version 3 and Version 4 of Working Groups/TAM/Reference Manual/Nemo Tam Basics

Ignore:
Timestamp:
2010-02-22T17:06:20+01:00 (11 years ago)
Comment:

### Legend:

Unmodified
 v3 = Tangent and Adjoint coding techniques = the original program {{{P}}}, whatever its size and run time, computes a function [[Image(ad_1)]] which is the composition of the elementary functions computed by each run-time instruction. In other words if {{{P}}} executes a sequence of elementary statements [[Image(ad_2.png)]], then {{{P}}} actually evaluates [[Image(ad_3.png)]] where each ''f,,k,,'' is the function implemented by ''I,,k,,''. Therefore one can apply the chain rule of derivative calculus to get the Jacobian matrix ''F!''', i.e. the partial derivatives of each component of ''Y'' with respect to each component of ''X''. Calling ''X,,0,,=X'' and ''X,,k,,=f,,k,,(X,,{k-1},,)'' the successive values of all intermediate variables, i.e. the successive '''states''' of the memory throughout execution of {{{P}}}, we get [[Image(ad_4.png)]] The derivatives ''f!',,k,,'' of each elementary instruction are easily built, and must be inserted in the differentiated program so that each of them has the values ''X,,k-1,,'' directly available for use. This process yields analytic derivatives, that are exact up to numerical accuracy. In practice, two sorts of derivatives are of particular importance in scientific computing: the tangent (or directional) derivatives, and the adjoint (or reverse) derivatives. The tangent derivative is the product $\dot{Y} = F'(X) \times \dot{X}$ of the full Jacobian times a direction $\dot{X}$ in the input space. >From equation above, we find [[Image(ad_5)]] which is most cheaply executed from right to left because matrix {{{x}}} vector products are much cheaper than matrix {{{x}}} matrix products. This is also the most convenient execution order because it uses the intermediate values ''X,,k,,'' in the same order as the program {{{P}}} builds them. On the other hand the adjoint derivative is the product [[Image(ad_6)]] of the ''transposed'' Jacobian times a weight vector $\overline{Y}$ in the output space. The resulting $\overline{X}$ is the gradient of the dot product $(Y \cdot \overline{Y})$. >From equation (\ref{eqchainrule}), we find \label{eqadjmode} \overline{X} = F'^{*}(X) \times \overline{Y} = f'^{*}_1(X_0) \times \dots \times f'^{*}_{p-1}(X_{p-2}) \times f'^{*}_p(X_{p-1}) \times \overline{Y} which is also most cheaply executed from right to left. However, this uses the intermediate values $X_k$ in the inverse of their building order in {\tt P}. the original program P, whatever its size and run time, computes a function F, X∈IRm 􏰄→ Y ∈IRn which is the composition of the elementary functions computed by each run-time instruction. In other words if P executes a sequence of elementary statements Ik,k ∈ [1..p], then P actually evaluates F =fp ◦fp−1 ◦···◦f1 , where each fk is the function implemented by Ik. Therefore one can apply the chain rule of derivative calculus to get the Jacobian matrix F′, i.e. the partial derivatives of each component of Y with respect to each component of X. Calling X0 = X and Xk = fk(Xk−1) the successive values of all intermediate variables, i.e. the successive states of the memory throughout execution of P, we get F′(X)=fp′(Xp−1)×fp′−1(Xp−2)×···×f1′(X0) .       (1) The derivatives fk′ of each elementary instruction are easily built, and must be inserted in the differentiated program so that each of them has the values Xk−1 directly available for use. This process yields analytic derivatives, that are exact up to numerical accuracy. In practice, two sorts of derivatives are of particular importance in scien- tific computing: the tangent (or directional) derivatives, and the adjoint (or reverse) derivatives. In particular, tangent and adjoint are the two sorts of derivative programs required for OPA, and TAPENADE provides both. The tangent derivative is the product Y ̇ = F ′(X) × X ̇ of the full Jacobian times a direction X ̇ in the input space. >From equation (1), we find Y ̇ =F′(X)×X ̇ =fp′(Xp−1)×fp′−1(Xp−2)×···×f1′(X0)×X ̇   (2) which is most cheaply executed from right to left because matrix×vector prod- ucts are much cheaper than matrix×matrix products. This is also the most convenient execution order because it uses the intermediate values Xk in the same order as the program P builds them. On the other hand the adjoint derivative is the product X = F ′∗(X) × Y of the transposed Jacobian times a weight vector Y in the output space. The resulting X is the gradient of the dot product (Y · Y ). >From equation (1), we find X =F′∗(X)×Y =f′∗(X )×···×f′∗  (X      )×f′∗(X )×Y     (3) 1 0     p−1 p−2 p p−1 which is also most cheaply executed from right to left. However, this uses the intermediate values Xk in the inverse of their building order in P. = Potential issues =