RobotGuide
Math foundations/Friction and the Friction Cone: The Geometry of Constraints

Foundations T24 / 27 · Contact and Convex Geometry

Friction and the Friction Cone: The Geometry of Constraints

Whether something slips does not depend on how big the force is — it depends on the direction. Coulomb friction draws a cone in force space, and stuffing that cone into an optimizer means first slicing it into a polytope — how large is the error and how does it converge? This page gives the full answer.

about 40 min3 theorems11 proof steps4 self-test questions

One-line takeaway

The friction cone is a convex second-order cone; the area error of an inscribed m-gon approximation vanishes at rate 1/m², and the tangential-force underestimate in the worst direction is cos(π/m): 4 wedges underestimate by 29%, 8 wedges by only 7.6%.

Grasp optimization, slip criteria, and the friction models of simulation engines deal with this cone every day. Why engineering practice often uses LP (linear programming), why SOCP (second-order cone programming) is more accurate, and what discretization costs — this page answers all three questions in one pass.

Why this tool exists

Push a box: push along the ground with 100 N and it moves; press down at an angle with 1000 N and the box does not budge. The force got bigger and the effect got worse — because slipping never compares force magnitudes. It compares the ratio of the tangential component to the normal component. The ceiling of that ratio is the friction coefficient μ.

Amontons' two experimental laws (1699): friction is independent of contact area, and friction is proportional to the normal load. Both are counterintuitive — why doesn't a larger contact area bring more friction? Because the true contact area is only on the order of one thousandth of the nominal area (asperity contact); increasing the nominal area just spreads the same true area out flatter. Coulomb later added the directional content: maximum static friction acts along the tangent of the contact surface and is proportional to the normal force.

For an optimizer, |f_t| ≤ μ·f_n is a nuisance: it is not linear, it is a «cone». In the world of optimization, the cheapest constraints are linear (LP), next cheapest are second-order cones (SOCP), and general nonlinear is pricier still. So every grasp optimization must answer the same question: approximating the cone with a polytope — how many faces are enough? This page derives the shape of the answer in full.

Definitions

Definition 2.1 Friction cone

The set of admissible forces at a contact point C_μ = {(f_n, f_t) : f_n ≥ 0, |f_t| ≤ μ·f_n}. This is a cone in the (f_n, f_t) plane with its apex at the origin, the normal axis as its centerline, and half-angle arctan(μ). The normal component can only push, not pull (f_n ≥ 0 is a unilateral constraint).

Why define it this wayWhat «cone» means: the set is closed under positive scaling — scale a force by λ and it stays admissible. The friction constraint restricts direction only, not magnitude; this is the essential difference from constraints of the form «force ≤ limit».

Definition 2.2 Second-order cone (SOC)

The n-dimensional second-order cone Q^n = {(x, t) : ||x||₂ ≤ t}. The friction cone is an affine image of the two-dimensional second-order cone: |f_t| ≤ μf_n ⟺ (f_t, μ·f_n) ∈ Q².

Why define it this wayThe second-order cone is convex (proof in Theorem 2.1), and there is a class of dedicated efficient solvers for it (SOCP). Recognizing the friction cone as an affine image of an SOC means knowing it can be handled rigorously by convex optimization — no approximation needed.

Definition 2.3 Extreme-ray linearization

Take m rays r₁, …, r_m inside the cone (direction angles spread uniformly over [−arctan μ, +arctan μ]) and replace the cone with their positive combinations Σαᵢrᵢ (αᵢ ≥ 0). Geometrically, this replaces the cone with an inscribed polyhedron with apex at the origin.

Why define it this wayPositive combinations turn «a cone constraint» into «α ≥ 0 plus an equality» — purely linear. The price is that the inscribed cone is smaller than the true cone: some physically admissible force directions get excluded. That loss is exactly what this page quantifies.

Theorems and proofs

Theorem 2.1 The friction cone is a convex cone

C_μ is closed under nonnegative scaling and convex combinations.

Proof

Closed under scaling: if f_n ≥ 0, |f_t| ≤ μf_n, and λ ≥ 0, then λf_n ≥ 0 and |λf_t| = λ|f_t| ≤ λμf_n = μ(λf_n).

Closed under convex combinations: let f, g ∈ C_μ and w = θf + (1−θ)g with θ∈[0,1]. Normal: w_n = θf_n + (1−θ)g_n ≥ 0. Tangential: |w_t| = |θf_t + (1−θ)g_t| ≤ θ|f_t| + (1−θ)|g_t| ≤ θμf_n + (1−θ)μg_n = μw_n. The first step uses the triangle inequality (convexity of the absolute value); the second uses that f and g are each in the cone. ∎

NoteThe triangle inequality is the only «substantive» step. This proof is short enough to memorize, but notice that it is the mechanical incarnation of the fact that the absolute value function is convex — the convexity of the friction cone is no coincidence; it is the direct consequence of Coulomb's law being linear in f_n.

Theorem 2.2 Area of the inscribed m-gon and its error order

The area of a regular m-gon inscribed in the unit circle is A_m = (m/2)·sin(2π/m), and its difference from π vanishes at rate 1/m².

Proof

Slice the m-gon into m isosceles triangles, each with vertex angle 2π/m and legs of length 1. One triangle's area = (1/2)·sin(2π/m) (the sine area formula for two sides of length 1 with included angle 2π/m), so the total is A_m = (m/2)sin(2π/m).

Error: π − A_m = m·[sector − triangle] = m·[(1/2)(2π/m) − (1/2)sin(2π/m)] = m·[π/m − (1/2)sin(2π/m)].

Use the Taylor expansion of sin x: sin x = x − x³/6 + O(x⁵), so (1/2)sin(2π/m) = π/m − (2π/m)³/12 + O(m⁻⁵). Substituting:

π − A_m = m·(2π/m)³/12 + O(m⁻⁴) = (2/3)π³/m² + O(m⁻⁴)

The leading coefficient is (2/3)π³ ≈ 20.67. ∎

NoteThis is where «error × m² ≈ constant» comes from — in the actual run below, this product drifts slowly between 19.6 and 20.7, which is the O(m⁻⁴) correction term at work. Archimedes approached π with the same geometry; we pay for friction linearization with the same geometry.

Theorem 2.3 The underestimate factor in the worst direction

Under the inscribed m-gon constraint (rays at uniform angles), the admissible tangential force is underestimated most in the direction whose angle to the nearest ray is π/m (midway between two rays); the maximum admissible magnitude is cos(π/m) times the true value.

Proof

Restate the problem: for a polygon inscribed in the unit circle, the ray from the origin at direction angle θ meets the polygon's boundary at radius r(θ). Each edge of the polygon is the chord joining two adjacent vertices (on the unit circle).

Let the direction in question fall between vertex angles v₁ and v₂; midway means θ = (v₁+v₂)/2, where θ makes an angle of π/m with both vertices. The boundary point lies on the chord v₁v₂.

The point on a chord closest to the circle's center is its midpoint (the foot of the perpendicular), at distance cos(π/m) from the center (right triangle: hypotenuse 1, angle π/m, adjacent side cos(π/m)). The direction in question points exactly at that foot, so r(θ) = cos(π/m).

For any direction, r(θ) ≥ cos(π/m) (the perpendicular foot is the closest point of the chord to the center; other directions intersect farther out), with equality exactly at the midway direction. So the maximal underestimate factor is cos(π/m), and the worst direction is midway between two rays. ∎

NoteNumerical comparison: m=4 → cos(45°)=0.707 (29.3% underestimate); m=8 → cos(22.5°)=0.924 (7.6%); m=16 → 0.981 (1.9%). Double the number of rays and the error drops fourfold — consistent with the 1/m² of Theorem 2.2.

Work it by hand

Fingertip normal force 1 N, μ = 0.4. Using 4-wedge and 8-wedge linearizations respectively, what is the admissible tangential force in the worst direction, and how much is underestimated?

  1. True value: |f_t| ≤ μf_n = 0.4 N (in any tangential direction).
  2. 4 wedges: admissible in the worst direction = 0.4×cos(π/4) = 0.4×0.7071 = 0.283 N, an underestimate of 0.117 N (29.3%).
  3. 8 wedges: 0.4×cos(π/8) = 0.4×0.9239 = 0.370 N, an underestimate of 0.030 N (7.6%).
  4. Engineering judgment: 4 wedges throw away nearly thirty percent of the tangential capability — if the grasp-optimization conclusion is sensitive to a 30% tangential margin (it usually is, since the margin exists precisely to absorb disturbances), you must go to 8 wedges or use SOCP directly.
Answer

4 wedges: 0.283 N (29.3% underestimate); 8 wedges: 0.370 N (7.6%). Pattern: error ∝ 1/m² — doubling the number of rays cuts the error to one quarter.

Run the numbers

Where the numbers come from

Area error of the inscribed regular m-gon and directional underestimate (fixed μ=0.4, normal force 1 N):

  • Area error (absolute) ×m²: m=6→6.23, m=12→6.49, m=24→6.56, m=48→6.57 — approaching the constant (2/3)π³≈20.67, whose percentage-of-π form is 657.97; second-order convergence holds
  • Note: absolute error = π − (m/2)sin(2π/m); ×m² tends to (2/3)π³ ≈ 20.67; the table prints percentages, and multiplying by m² corresponds to (200/3)π² ≈ 657.97
  • Directional underestimate: m=4 → 29.29%, m=8 → 7.61%, m=12 → 3.41%, m=16 → 1.92%
  • 100,000 random directions: the inscribed polygon's boundary is everywhere ≤ 1, so every direction is underestimated; the worst direction is exactly midway between two rays (Theorem 2.3)

Where this shows up

The object hereThe counterpart in RLUsed in
Convexity and linearizationThe trade-off between LP (wedge cone) and SOCP (true cone) in grasp optimizationused in →
Worst-direction underestimate cos(π/m)Why slip criteria are conservative: after linearization, the admissible tangential force is smaller than the true oneused in →
Unilateral constraint f_n ≥ 0The mathematical form of «fingers can only push, not pull», and the existence of grasp force solutionsused in →
Affine images of conesAfter summing multiple contacts, the wrench set is still a convex cone (prerequisite for Theorem 3.1 in T13)used in →

Boundaries: where this article stops

Where this page stops: rolling friction (related to rolling-resistance torque, a different mechanism), anisotropic friction (fiber direction), contact-area friction of soft fingers (Pearson-type, area-dependent), and stick-slip dynamics (covered in T15) — none of these. Honest disclaimer: Amontons' laws are statistical laws on true contact areas a thousandth the size of nominal ones, not fundamental forces — on rubber and viscoelastic materials they fail systematically.

Self-test: 4 questions

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

01Why is the friction cone a cone rather than a «ball» or a «box»?

Because the constraint restricts direction (the tangential-to-normal ratio) but not magnitude: scale a force by λ and the constraint's form is unchanged — that is closure under scaling, the definition of a cone. A box (|f_n|≤a, |f_t|≤b) would wrongly cap the normal force.

02After linearizing the cone into m wedges, where does the conservativeness of the optimization result show up? Could it ever be «non-conservative»?

The inscribed cone ⊂ the true cone, so the optimizer only accesses a smaller force set: the result is conservative (grasp capability underestimated), with a worst directional factor of cos(π/m). A circumscribed polygon would be non-conservative (overestimating capability, possibly yielding grasps that actually slip) — in engineering, prefer inscribed over circumscribed.

03μ = 0.4, 8 wedges, normal force 2 N: what is the admissible tangential force in the worst direction?

2×0.4×cos(π/8) = 0.739 N. First compute the true cap 0.8 N, then multiply by the underestimate factor 0.9239.

04Why doesn't «the true contact area is only a thousandth of the nominal area» overturn Amontons' second law (F ∝ N)?

Under asperity contact, increasing the normal load deforms the asperities themselves, and the true area grows with N — friction ∝ true area ∝ N. The nominal area only decides how flat the contact is spread, not the total true area. The two laws are consistent under the microscopic picture.

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