Implicit differentiation and the inverse Jacobian
Introduced| English | Meaning |
|---|---|
| implicit differentiation/ɪmˈplɪsɪt ˌdɪfəˌrenʃɪˈeɪʃn/ | Differentiating defining equations while dependent variables change with the chosen independent variable. |
| Jacobian/dʒæˈkəʊbɪən/ | The derivative matrix of a coordinate map, or its determinant when used as a signed local scaling factor. |
| local inverse/ˈləʊkl ɪnˈvɜːs/ | A map that reverses a function between suitable neighbourhoods of corresponding source and output points. |
A coupled control cannot always be reversed one dial at a time
- Two controls produce outputs $x=u+v$, $y=u+2v$. Increasing $u$ while keeping $v$ fixed changes both outputs. Reversing the change in $x$ while keeping y fixed is a different experiment, so its sensitivity need not be a scalar reciprocal.
- We will use implicit differentiation 隐函数求导 to track the coupled changes. The derivative matrix controls sensitivities; its determinant, often called the Jacobian 雅可比行列式 in integration, controls signed local area scaling.
Separate the matrix from its determinant
- For $F(u,v)=(f(u,v),g(u,v))$, the Jacobian matrix is $J=\begin{pmatrix}f_u&f_v\\g_u&g_v\end{pmatrix}$. Rows follow the output order $(x,y)$; columns follow the input order $(u,v)$. State this order before using an entry.
- Its determinant is the scalar $\Delta=f_ug_v-f_vg_u$. Some texts use “Jacobian” for the matrix, others for this determinant. Here $J$ always means the matrix and $\Delta=\det J$ the determinant; an inverse matrix is not the reciprocal of each entry.
For $J=\begin{pmatrix}2&1\\1&-1\end{pmatrix}$, what is the Jacobian determinant?
Difficulty 1/5 · 100 coins + a small bonus on first completion today
The determinant is the scalar 2(-1)-1(1)=-3; J itself is a matrix.
Check a determinant at its source point
- For the original map $F(u,v)=(uv,u+v)$, $J=\begin{pmatrix}v&u\\1&1\end{pmatrix}$ and $\Delta=v-u$. At source $(1,2)$, the determinant is 1 and the output is $(2,3)$.
- Swapping the source coordinates gives $(2,1)$, with the same output but determinant $-1$. Each source has its own smooth local inverse branch; the full map is not one-to-one. A determinant at one source does not choose between distant preimages.
At $(u, v) = (1, 2)$, the Jacobian determinant of $(x, y) = (uv, u + v)$ is
Difficulty 1/5 · 100 coins + a small bonus on first completion today
$v \cdot 1 - u \cdot 1 = 2 - 1$.
Columns describe forward changes
- At a specified source point, the first column gives $(x_u,y_u)$ with $v$ fixed; the second gives $(x_v,y_v)$ with $u$ fixed. The differential equations are $dx=f_u\,du+f_v\,dv$ and $dy=g_u\,du+g_v\,dv$.
- For a nonlinear map these describe the first-order change: small actual increments are approximated by $J(\Delta u,\Delta v)^T$. For a linear map, the same matrix relation for increments is exact. Evaluate the partial derivatives at the source point.
Hold the other output fixed when reversing
- To find $u_x$ with $y$ fixed, differentiate both defining equations: $f_uu_x+f_vv_x=1$ and $g_uu_x+g_vv_x=0$. The zero comes from holding $y$ constant; both input variables may change.
- In the opening map, solve explicitly: $u=2x-y$, $v=y-x$. Thus $u_x=2$, $v_x=-1$, $u_y=-1$ and $v_y=1$. The forward entry $x_u=1$ and reverse entry $u_x=2$ belong to different held-fixed conditions.
For a map of two variables, $\partial u/\partial x$ always equals $1/(\partial x/\partial u)$.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
The forward derivative holds v fixed, while the inverse derivative holds y fixed. In x=u+v,y=u+2v, x_u=1 but u_x=2. Special triangular or decoupled maps can satisfy a reciprocal relation.
State the local inverse theorem
- Let $F$ be $C^1$ on an open neighbourhood of a source point $p=(u_0,v_0)$. If $\det J(p)\ne0$, there are neighbourhoods of $p$ and $q=F(p)$ on which $F$ has a unique $C^1$ local inverse 局部逆映射 $G$.
- Differentiating $G(F(p))=p$ gives $DG(q)J(p)=I$, hence $DG(q)=J(p)^{-1}$. The inverse derivative is evaluated at the output point $q$, using the forward derivative at its corresponding source point $p$.
Read the inverse entries in order
- When $\Delta\ne0$, $J^{-1}=\dfrac1\Delta\begin{pmatrix}g_v&-f_v\\-g_u&f_u\end{pmatrix}$. Its rows now follow $(u,v)$ and its columns follow $(x,y)$: $\begin{pmatrix}u_x&u_y\\v_x&v_y\end{pmatrix}=J^{-1}$.
- Therefore $u_x=g_v/\Delta$, $u_y=-f_v/\Delta$, $v_x=-g_u/\Delta$ and $v_y=f_u/\Delta$. In each x-column derivative hold y fixed, and in each y-column derivative hold x fixed. Check the full matrix product, not just one entry.
Work the original linear exercise
- For $x=u+2v$, $y=3u+4v$, $J=\begin{pmatrix}1&2\\3&4\end{pmatrix}$ and $\Delta=-2$. The inverse is $\begin{pmatrix}-2&1\\3/2&-1/2\end{pmatrix}$. Thus the original answers are $u_x=-2$ and $v_y=-1/2$.
- The explicit inverse is $u=-2x+y$, $v=(3x-y)/2$, so all four derivatives can be checked directly. This linear map is globally one-to-one because its constant matrix is invertible; that global conclusion uses more than a theorem at one point.
$x = u + 2v$, $y = 3u + 4v$. Find $\partial u/\partial x$ (with $y$ fixed).
Difficulty 2/5 · 140 coins + a small bonus on first completion today
$\det J = -2$; $u_x = g_v/\det J = 4/(-2)$.
For $x=u+2v$, $y=3u+4v$, find $\partial v/\partial y$ with x fixed.
Difficulty 2/5 · 140 coins + a small bonus on first completion today
By the inverse function theorem, J⁻¹ = (1/det J)·adj J, so ∂v/∂y = f_u / det J. With f_u = 1 and det J = −2 this is −0.5.
For $x=2u+3v$, $y=u+v$, find the output corresponding to source $(1,-1)$ and all four inverse partial derivatives there. Give the forward area factor.
Check the matrix order, corresponding point, held-fixed variable and regularity against the model.
Match points in a nonlinear example
- For $x=u^2+v$, $y=u-v$, the source point $(1,0)$ maps to $(1,1)$. At that source, $J=\begin{pmatrix}2&1\\1&-1\end{pmatrix}$ and $\Delta=-3$.
- At the corresponding output point, $DG=\dfrac13\begin{pmatrix}1&1\\1&-2\end{pmatrix}$. In particular $u_x=1/3$, while $1/x_u=1/2$. The original table compares this with the opening linear map.
For $F(u,v)=(u^2+v,u-v)$, the inverse derivative obtained from $J(1,0)$ is evaluated at which output point?
Difficulty 1/5 · 100 coins + a small bonus on first completion today
F(1,0)=(1,1). The derivative of the inverse belongs to the corresponding output point, on the branch through source (1,0).
Choose the branch before differentiating
- In that nonlinear map, $v=u-y$ gives $u^2+u=x+y$. Near source $(1,0)$, use $u=(-1+\sqrt{1+4x+4y})/2$, then $v=u-y$. The positive square-root branch recovers $u=1$ at output $(1,1)$.
- Differentiating gives $u_x=u_y=1/\sqrt{1+4x+4y}$, hence $1/3$ there. The other branch gives source $(-2,-3)$ at the same output. Thus each branch is locally valid away from the turning point, but the unrestricted map is not globally one-to-one.
The map $x=u^2+v$, $y=u-v$ sends $(1,0)$ to $(1,1)$. What is the other source value of u for that same output?
Difficulty 2/5 · 140 coins + a small bonus on first completion today
Elimination gives u²+u=2, so (u-1)(u+2)=0. The other u is -2 and v=u-y=-3.
For $x=u^2-v$, $y=u+v$, choose the inverse branch through source $(1,0)$. Find $u_x$ and $v_y$ at the corresponding output, and show why the unrestricted map is not one-to-one.
Check the matrix order, corresponding point, held-fixed variable and regularity against the model.
Nonzero determinants everywhere still need global checks
- The map $F(u,v)=(e^u\cos v,e^u\sin v)$ has determinant $e^{2u}>0$ everywhere. It has a smooth local inverse near each source point, yet $v$ and $v+2\pi$ give the same output.
- Local invertibility does not select a global angle branch. Restricting v to an open interval of length at most $2\pi$ removes these repetitions, with the image restricted accordingly. A nonzero determinant alone is not a global injectivity proof.
A singular derivative can have a continuous inverse
- For $F(u,v)=(u^3,v)$, the determinant is $3u^2$, zero when $u=0$. Nevertheless $G(x,y)=(\sqrt[3]x,y)$ is a global continuous inverse. Its first component is not differentiable in x at zero.
- For $F(u,v)=(u^2,v)$, no open neighbourhood of $(0,0)$ is one-to-one: sources $(t,0)$ and $(-t,0)$ have the same output for arbitrarily small nonzero t. A zero determinant alone distinguishes neither of these outcomes.
A differentiable inverse would force a nonsingular matrix
- If both $F$ and a two-sided local inverse $G$ are differentiable at corresponding points, the chain rule gives $DG\,DF=I$. A singular $DF$ cannot satisfy this identity. Thus zero determinant rules out a differentiable inverse there.
- The inverse function theorem is a sufficient test for a $C^1$ inverse. When the determinant is zero, it gives no such inverse; the cube example shows why this is different from ruling out every continuous inverse. State the regularity required by the question.
A zero Jacobian determinant does not by itself rule out a continuous local inverse.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
The map F(u,v)=(u³,v) has continuous inverse (cube root of x,y), despite determinant zero at u=0. That inverse is not differentiable there. A differentiable two-sided inverse would force an invertible derivative by the chain rule.
The map $F(u,v)=(u^3,v)$ has a differentiable two-sided inverse at output $(0,0)$.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
Its continuous inverse has first component cube root of x, which has no finite x derivative at zero. Also DG DF=I is impossible when DF is singular.
A scalar reciprocal can hold in special maps
- Suppose $f_u\ne0$ and $\Delta\ne0$. The equality $u_x=1/f_u$ holds precisely when $f_vg_u=0$, because $u_x=g_v/\Delta$ and $\Delta=f_ug_v-f_vg_u$. This is a special structural condition, not a general rule for partial derivatives.
- For $x=2u+v$, $y=3v$, holding y fixed forces v fixed, so $u_x=1/2=1/x_u$. The map still has two inputs and x depends on both. Two-variable maps can satisfy this reciprocal relation when the held-fixed conditions agree.
For $x=2u+v$, $y=3v$, the inverse partial $u_x$ with y fixed equals $1/x_u$ with v fixed.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
Holding y fixes v, so 2u+v=x gives u_x=1/2. The forward derivative x_u is 2; this triangular map is a valid special case.
Differentiate an implicit system as a system
- Let $H(u,v,x,y)=0$ and $K(u,v,x,y)=0$ hold at a point, with both functions $C^1$ nearby. If the dependent-variable block $M=\begin{pmatrix}H_u&H_v\\K_u&K_v\end{pmatrix}$ is invertible there, u and v are locally $C^1$ functions of x and y.
- With y fixed, solve $H_uu_x+H_vv_x=-H_x$ and $K_uu_x+K_vv_x=-K_x$. With x fixed, replace the right sides by $-H_y,-K_y$ and solve for $u_y,v_y$. Include each equation's direct dependence on the chosen independent variable.
Work a coupled implicit derivative
- Take $H=u^2+v-x-y$ and $K=u+v^2-x+y$. At $(u,v,x,y)=(1,1,2,0)$, both equations vanish and $M=\begin{pmatrix}2&1\\1&2\end{pmatrix}$ has determinant 3.
- For x derivatives the right side is $(1,1)^T$, giving $u_x=v_x=1/3$. For y derivatives it is $(1,-1)^T$, giving $u_y=1$, $v_y=-1$. Substituting these pairs into both differentiated equations checks every sign.
For $u^2+v-x-y=0$ and $u+v^2-x+y=0$, find $u_x$ with y fixed at $(u,v,x,y)=(1,1,2,0)$.
Difficulty 2/5 · 140 coins + a small bonus on first completion today
Differentiate both equations: 2u_x+v_x=1 and u_x+2v_x=1. Solving gives u_x=v_x=1/3.
For $H=u^2+v-x-y=0$ and $K=u+v^2-x+y=0$, find all four partial derivatives of u,v with respect to x,y at $(u,v,x,y)=(0,1,1,0)$.
Check the matrix order, corresponding point, held-fixed variable and regularity against the model.
Singular equation blocks can still define a smooth graph
- The equations $H=(u-x)^3=0$ and $K=v-y=0$ define the unique smooth graph $u=x,v=y$. At every solution the dependent block is singular: $H_u=0$. This representation fails the implicit theorem's sufficient test even though the graph exists.
- This does not contradict the previous inverse result: an invertible graph map and a nonsingular defining-equation block are different objects. Singular implicit equations need direct analysis; for $u^2=x$, near zero there can be two real branches, one, or no real solution depending on x.
What does the stated example justify?
Distinguish smooth inversion, continuous-only inversion and an implicit graph.
At the origin compare (i) the inverse of $F(u,v)=(u^3,v)$; (ii) the inverse of $F(u,v)=(u^2,v)$; and (iii) the graph defined by $(u-x)^3=0$, $v-y=0$. All have a singular relevant matrix. What exists in each case?
Check the matrix order, corresponding point, held-fixed variable and regularity against the model.
Use a determinant for area, with the map direction stated
- For a one-to-one $C^1$ change of variables with nonzero determinant on the region's interior, $dA_{xy}=|\Delta|\,du\,dv$. Substitute the integrand and region too. A negative determinant reverses orientation; its absolute value scales ordinary area.
- The original linear map has $\Delta=-2$, so a unit input square maps to a parallelogram of area 2. Reversing the map gives area factor $1/2$, rather than an individual sensitivity entry. In polar coordinates the forward factor is r for $r>0$, on a chosen angle branch; the origin requires separate treatment.
Which derivative quantity answers this task?
Keep direction, sensitivity, orientation and ordinary area distinct.
The map $x=u+2v$, $y=3u+4v$ sends a unit square in the uv plane to a parallelogram. Find its ordinary area.
Difficulty 2/5 · 140 coins + a small bonus on first completion today
The forward determinant is -2, so the ordinary area factor is |-2|=2. The inverse area factor would be 1/2.
Recap: direction, point and regularity
- State which variable stays fixed, the source/output points and the type of inverse needed. Use the full inverse matrix for sensitivities, the appropriate implicit block for coupled equations, and an absolute determinant for ordinary area scaling.
- Jacobian: The derivative matrix of a coordinate map, or its determinant when used as a signed local scaling factor; state which meaning is intended.
- implicit differentiation: Differentiating defining equations while dependent variables change with the chosen independent variable.
- local inverse: A map that reverses a function between suitable neighbourhoods of corresponding source and output points.