RobotGuide
Math foundations/The Jacobian: Rotate a Joint One Degree — How Far Does the End Effector Move?

Foundations T9 / 27 · Linear Algebra and Geometry

The Jacobian: Rotate a Joint One Degree — How Far Does the End Effector Move?

The Jacobian is the conversion table for «each joint moves a little — how much does the end effector follow» — the multivariable chain rule arranged as a matrix. Velocity control, force control, singularity analysis: all of it is read off this one table.

About 45 min3 theorems7 proof steps4 self-test questions

One-line takeaway

Two-link arm: you rotate the shoulder 1° and the elbow 2° — how far does the tip (the pen point) travel, and in which direction? Not one number but a 2×2 table of partial derivatives: the sensitivity of x to θ1, of x to θ2, of y to θ1, of y to θ2. That table is the Jacobian J. Robots compute it, use it, and check when it fails — a thousand times per second.

Animated · drag the arm joints and watch |det J| collapse — the moment it stretches straight is the singular pose

In the previous six bridges you learned the single-variable derivative (slope → gradient). Now two variables determine two outputs — the derivative upgrades to a «table», and that table is the Jacobian. It is nearly the highest-frequency matrix in robotics: velocity transmission, force mapping, singularity, least-squares inversion — all pass through its hands.

Why this tool exists

High-school math is entirely sufficient: a partial derivative = hold every other variable fixed, let one variable move a tiny bit, and look at the ratio of the output's movement. You do not need to «finish multivariable calculus» to read this — the page computes a two-link arm completely.

Definitions

Definition 3.1 Partial derivative (intuitive version)

The partial derivative of f(x_1, …, x_n) with respect to x_i, ∂f/∂x_i: hold all other variables fixed, let x_i increase a little, and look at the ratio of how much f moved. It is «the single-variable derivative», except you release only one knob at a time.

Definition 3.2 The Jacobian matrix

For a system with m outputs f_1…f_m and n inputs x_1…x_n, the Jacobian J is the m×n matrix whose (i,j) entry = ∂f_i/∂x_j. Row i = the sensitivity of output f_i to all inputs; column j = the reaction of all outputs when input x_j moves alone. For this page's arm: inputs (θ1, θ2), outputs (x, y), and J is 2×2.

Definition 3.3 Singular configuration

A joint-angle combination where det J = 0. At that moment the Jacobian «flattens» some direction — some joint-velocity combination produces zero end-effector velocity, and there is an end-effector direction no joint velocity can reach. Near a singularity the arm loses control, jitter, and overloads its actuators — all of it is this one determinant's fault.

Theorems and proofs

Theorem 3.1 The chain rule in matrix form

If z = g(y) and y = h(x), then the Jacobian of the composite map is J_z|_x = J_z|_y · J_y|_x — Jacobians multiply along the chain.

Proof

Verify the one-dimensional core case componentwise and it generalizes: z = g(h(x)); a small perturbation δx moves y by δy = h′(x)δx, which moves z by δz = g′(y)δy = g′(h(x))·h′(x)·δx.

The proportionality factor g′(h(x))·h′(x) is exactly the product of the two derivatives — the matrix form is J_z J_y. With multiple inputs, each (output, input) pair contributes one partial derivative by the same logic, and arranging them as a matrix is precisely the (i,j)-entry definition of matrix multiplication.∎

NoteThis is the same fact as T8's associativity: transformations in series, bookkeeping matrices multiplied. The differential version of a kinematic chain (base→shoulder→elbow→wrist) is a string of Jacobians multiplied.

Theorem 3.2 The Jacobian and determinant of the two-link arm

With link lengths l_1 = l_2 = 1, the tip is (x, y) = (cos θ1 + cos(θ1+θ2), sin θ1 + sin(θ1+θ2)). Then J = [[−sin θ1 − sin(θ1+θ2), −sin(θ1+θ2)], [cos θ1 + cos(θ1+θ2), cos(θ1+θ2)]], and det J = sin θ2.

Proof

Differentiate x with respect to θ1 (θ2 fixed): ∂x/∂θ1 = −sin θ1 − sin(θ1+θ2). With respect to θ2: −sin(θ1+θ2). Same for y with cos.

Expand det J by definition: [−sin θ1 − sin(θ1+θ2)]·cos(θ1+θ2) − [−sin(θ1+θ2)]·[cos θ1 + cos(θ1+θ2)].

Expand and collect: = −sin θ1 cos(θ1+θ2) + sin(θ1+θ2) cos θ1 = sin(θ1+θ2 − θ1) = sin θ2 (the sine angle-addition formula, high-school material).∎

NoteThe reason the result is startlingly clean: det J = sin θ2 depends only on the elbow angle. The shoulder angle θ1 only decides «where the arm points», not «how straight it extends» — the singular configuration is decided entirely by the elbow.

Theorem 3.3 Velocity transmission and singularity

The tip velocity is v = J q̇ (q̇ is the vector of joint angular velocities). When det J ≠ 0, J is invertible and any tip velocity is realizable by some joint velocities; when det J = 0 there are «unreachable» directions, and the required joint velocities blow up to infinity.

Proof

Take the total time derivative of the pose equations; each partial derivative contributes one term via the chain rule: ẋ = (∂x/∂θ1)θ̇1 + (∂x/∂θ2)θ̇2, likewise ẏ — arranged as a matrix this is v = J q̇. A direct corollary of the «first-order approximation of small perturbations».

det J = sin θ2 (Theorem 3.2). As θ2 → 0 or π, det J → 0: J⁻¹ does not exist. To move along a «flattened direction», the inverse solution's q̇ grows like 1/|det J|, tending to infinity — this is why arms near singularity spin wildly and motors overcurrent.∎

NoteEngineering mantra: the straighter the arm, the harder it turns. A human reaching for something behind and to the side with a straight arm feels «awkward»; a robot arm at the same pose experiences numerical explosion. Planners avoid singular configurations not because they are slow, but because they lose control.

Work it by hand

l_1 = l_2 = 1, θ1 = 30°, θ2 = 60°. Compute J and det J; if the joint velocities are q̇ = (0.5, −0.5) rad/s, how fast does the tip move?

  1. θ1 + θ2 = 90°. sin θ1 = 0.5, cos θ1 = √3/2 ≈ 0.866; sin 90° = 1, cos 90° = 0.
  2. [object Object]
  3. det J = (−1.5)(0) − (−1)(0.866) = 0.866 = sin 60° ✓ (consistent with Theorem 3.2).
  4. [object Object]
  5. Tip speed = √(0.25² + 0.433²) = 0.5 m/s, biased toward positive y — shoulder forward, elbow backward, the tip is «sweeping sideways».
Answer

J = [[−1.5, −1], [0.866, 0]], det J = 0.866, tip velocity 0.5 m/s pointing along (−0.25, 0.433). The whole computation uses only trigonometry and multiplication — the Jacobian is not hard; what is hard is knowing what you are computing at every step.

Run the numbers

Where the numbers come from

Actually run at generation time (l_1 = l_2 = 1):

  • At θ2 = 90°, |det J| = 1.000 — the elbow angle with the best manipulability
  • At θ2 = 10-6 rad (nearly straight), |det J| = 1.0e-6 — collapsed by six orders of magnitude
  • Inverse-solution joint velocities grow like 1/|det J|: for the same tip velocity, the near-straight pose needs joint speeds on the order of a million times larger

Where this shows up

The object hereThe counterpart in RLUsed in
Velocity controlCartesian velocity command → joint velocities: q̇ = J−1 vused in →
Force controlTip force → joint torques: τ = J^T F (the duality)used in →
Policy gradientIn RL, «the sensitivity table of actions to parameters» is the gradient — a special case of the Jacobianused in →

Boundaries: where this article stops

Where this page stops: only 2×2. The 6×6 generalized Jacobian of space arms (floating base), the full theory of the manipulability ellipsoid, and null-space projection for redundant arms are not developed — but every 2×2 conclusion here (v = J q̇, τ = J^T F, det J collapse) holds verbatim in higher dimensions.

Self-test: 4 questions

All hand-computation and «find the flaw in this proof» questions — nothing to memorize. Stuck? Go back to the section.

01What physical quantity is the (i,j) entry of the Jacobian?

The rate at which unit angular velocity of the j-th joint changes the i-th output (say, the tip's x coordinate). The whole table = the complete sensitivity ledger of «joints move a little → how the tip moves».

02Why is a singular configuration dangerous?

det J → 0: J is not invertible. To command tip velocity along a flattened direction, the inverse-solution joint velocities diverge like 1/|det J| — motor overcurrent, numerical explosion, control divergence.

03Why does force control use J^T?

The duality of the virtual-work principle: if the velocity map is v = J q̇, the force map runs backward, τ = J^T F. Same matrix, transposed — the velocity table becomes the force table.

04What is the relation between partial derivatives and the high-school derivative?

The same thing: hold the other variables, open one knob, read the output's rate of change. The high-school derivative is the n=1 special case; the partial derivative is the version that opens one knob at a time — no new mathematical tools at all.

Up next

The full series · 27 articles

T1The Matrix Is a Ledger for Rotation≈ 25 minT2From Slope to Gradient: The First Line of Every Robot Learning Algorithm≈ 25 minT3Vectors and forces: the physical reading of the dot productAbout 20 minT4Two Rulers and a Better Rulerabout 25 minT5The Ledger of One Motor Revolution: Radians, Speed, and Powerabout 20 minT6Complex multiplication is rotation: the one-dimensional version of quaternionsAbout 20 minT7Bases and coordinates: why robots put everything in matrices45 minT8Matrix Multiplication: Chaining Transformations, Not a Table OperationAbout 40 minT9The Jacobian: Rotate a Joint One Degree — How Far Does the End Effector Move?About 45 minT10Eigenvectors: The Directions a Matrix Dare Not Touchabout 45 minT11SVD and the Pseudoinverse: When There Is No Exact Solution, Find the Nearest Oneabout 50 minT12Taylor expansion: putting a straight line to work in place of a curveAbout 40 minT13Probability spaces: why «random» needs axioms50 minT14Expectation: A Rigorous Construction of «Average»About 55 minT15Conditional Expectation: What the Vertical Bar Really Means≈ 60 minT16Four inequalities: the quantitative answer to «how many samples are enough»45 minT17Norms and fixed points: where «it always converges» comes from55 minT18Markov Chains: When the Time Average Equals the Ensemble Averageabout 55 minT19The Mathematical Construction of the MDP: What One Axiom Is Worthabout 50 minT20Dynamic Programming: The Accounting for Two Iterations≈ 50 minT21Optimization: why constraints cannot rely on penalties aloneAbout 55 minT22Stochastic Approximation: Every Term in the Learning-Rate Formulaabout 55 minT23The Mathematics of Contact: Hertz Theoryabout 45 minT24Friction and the Friction Cone: The Geometry of Constraintsabout 40 minT25Force Closure and Grasp Metrics: ε Is a Linear ProgramAbout 50 minT26Contact Kinematics: Why Rolling Is Not «Sliding Along»≈ 45 minT27Slip detection: from «I felt it» to «I computed it»About 50 min