Arc length measures how far we travel along a curve. Curvature measures how quickly our direction changes as we travel. A tight bend has greater curvature than a gentle bend, regardless of how fast we move through it.
Turning per unit distance
Let r(t) be a regular curve: its velocity r′(t) is nonzero throughout the interval we are considering. Assume that r is twice continuously differentiable, so its first and second derivatives exist and are continuous.
The unit tangent vector is
T(t)=∥r′(t)∥r′(t).
It records the direction of motion without the speed. Since ∥T(t)∥=1, any change in T is a change of direction.
How the tangent direction changes
The arrows represent the unit tangent T. Curvature measures how quickly its direction changes per unit distance along the curve, not per unit time.
The quantity ∥T′(t)∥ measures how fast that direction changes per unit of the parametert. That still depends on how quickly we move along the curve. To measure the bending of the curve itself, we need change per unit of arc length instead.
Write s for arc length and κ (read “kappa”) for curvature. The definition is
κ=dsdT.
If the curve is already parametrized by arc length, we can differentiate its unit tangent directly with respect to s. Curvature is nonnegative and has units of inverse length.
Using an arbitrary parameter
Finding an explicit arc-length parametrization can be difficult. We can avoid that step by using the parameter we already have. Recall that the arc-length function satisfies
s(t)=∫t0t∥r′(u)∥du,dtds=∥r′(t)∥,
where t0 is a fixed starting parameter and u is a dummy integration variable. The chain rule connects the two rates of change:
dtdT=dsdTdtds=dsdT∥r′(t)∥.
Taking norms and dividing by the positive speed gives
κ(t)=∥r′(t)∥∥T′(t)∥.
Here κ(t) means the curvature at the point reached at parameter t. Dividing by speed removes the effect of moving through the curve faster or slower. The denominator is the length of the velocity, not the length of the position vector.
Circles and lines
A circle of radius a
For a fixed a>0, consider
r(t)=⟨acost,asint⟩.
This traces a circle of radius a centered at the origin.
A circle has the same curvature everywhere
For a circle of radius a>0, curvature is κ=a1. Smaller circles bend more sharply.
First find the velocity and its length:
r′(t)=⟨−asint,acost⟩,∥r′(t)∥=a2sin2t+a2cos2t=a.
Dividing by a gives the unit tangent. Differentiating that vector gives
T(t)=⟨−sint,cost⟩,T′(t)=⟨−cost,−sint⟩.
Its length is ∥T′(t)∥=cos2t+sin2t=1. Therefore
κ(t)=a1.
The curvature is the same everywhere on the circle. A smaller circle bends more sharply, so its curvature is larger.
A straight line
Let
r(t)=⟨x0,y0,z0⟩+t⟨a,b,c⟩,a2+b2+c2>0.
The point (x0,y0,z0) lies on the line, and the nonzero vector ⟨a,b,c⟩ gives its direction. Then
r′(t)=⟨a,b,c⟩,T(t)=a2+b2+c2⟨a,b,c⟩.
The unit tangent is constant, so T′(t)=0 and
κ(t)=0.
A line does not turn. Notice that zero curvature is allowed even though zero velocity is not allowed in the curvature formula.
Computing curvature with a cross product
Finding T and then differentiating it can be cumbersome, especially when speed is not constant. For a regular curve in R3, an equivalent formula uses the first two derivatives directly:
κ(t)=∥r′(t)∥3∥r′(t)×r′′(t)∥.
The symbol × denotes the cross product. Take the norm of that cross product in the numerator, and cube the speed in the denominator.
Example: a polynomial space curve
Take
r(t)=⟨t,t2,t3⟩.
Differentiate twice:
r′(t)=⟨1,2t,3t2⟩,r′′(t)=⟨0,2,6t⟩.
Using the coordinate unit vectors i, j, and k, the determinant expansion gives
This is valid for every real t: the first component of r′(t) is always 1, so the velocity never vanishes.
Curvature of a graph
For a graph y=f(x), we can use x itself as the parameter. Assume that f is twice continuously differentiable. View the graph in the z=0 plane:
r(x)=⟨x,f(x),0⟩.
Then
r′(x)=⟨1,f′(x),0⟩,r′′(x)=⟨0,f′′(x),0⟩,
and
∥r′(x)∥=1+(f′(x))2.
The cross product has only a third component:
r′(x)×r′′(x)=i10jf′(x)f′′(x)k00=f′′(x)k.
Its norm is ∣f′′(x)∣, giving
κ(x)=(1+(f′(x))2)3/2∣f′′(x)∣.
Keep the absolute value: curvature measures the amount of bending and cannot be negative. The denominator is always positive. This formula applies where the curve is represented by a twice differentiable graph y=f(x); a general plane curve need not have one such representation everywhere.
Example: a parabola
For y=x2, the derivatives are f′(x)=2x and f′′(x)=2. Therefore
κ(x)=(1+4x2)3/22.
The denominator is smallest at x=0, so curvature is greatest at the vertex:
κ(0)=2,x→±∞limκ(x)=0.
The parabola bends most sharply at its vertex
The curve shown is y=x2, not a graph of curvature. At its vertex, κ(0)=2; farther along either branch, its direction changes less per unit arc length.
Farther from the vertex, the tangent line becomes steeper, but its direction changes less per unit distance traveled. A large slope does not necessarily mean large curvature.
The TNB frame
The unit tangent tells us the direction of travel. When that direction is changing, we can also identify the direction in which it turns.
Since T has constant length,
T(t)⋅T(t)=1.
Differentiate using the dot-product rule:
2T(t)⋅T′(t)=0.
So T′(t) is perpendicular to T(t). Wherever T′(t)=0, normalize it to get the principal unit normal, also called the unit normal:
N(t)=∥T′(t)∥T′(t).
The binormal is
B(t)=T(t)×N(t).
These three vectors form the TNB frame:
T points along the curve in the direction of motion.
N points in the direction in which the unit tangent is changing.
B is perpendicular to both, with its direction fixed by the cross-product order T×N.
They are orthonormal: mutually perpendicular and each of length 1. The length of B is 1 because it is the cross product of two perpendicular unit vectors. The ordered frame is right-handed.
The nonzero condition matters. If T′(t)=0, then curvature is zero and the formulas above do not define N or B at that point. In particular, a straight line has a unit tangent and zero curvature, but its principal normal and binormal are not defined by these formulas.
Example: the TNB frame of a helix
Consider the circular helix
r(t)=⟨cost,sint,t⟩,t∈R.
Its velocity and speed are
r′(t)=⟨−sint,cost,1⟩,∥r′(t)∥=2.
The speed never vanishes, so T is defined for every real t:
T(t)=21⟨−sint,cost,1⟩.
Differentiate and normalize:
T′(t)=21⟨−cost,−sint,0⟩,∥T′(t)∥=21,
so T′(t) also never vanishes and
N(t)=⟨−cost,−sint,0⟩.
This normal vector points horizontally inward toward the z-axis. The binormal is
B(t)=T(t)×N(t)=21⟨sint,−cost,1⟩.
The three vectors are mutually perpendicular unit vectors, and the cross-product order makes the frame right-handed.
The TNB frame and acceleration on a helix
Velocity, speed, and acceleration
For motion along a regular curve, the vector quantities and scalar quantities answer different questions:
Quantity
Type
Meaning
v(t)=r′(t)
vector
velocity: speed and direction
a(t)=r′′(t)
vector
acceleration: change in velocity
s(t)=∫t0t∥r′(u)∥du
scalar
arc length measured from t0
s′(t)=∥r′(t)∥
scalar
speed
s′′(t)
scalar
signed rate of change of speed
In general,
s′′(t)=∥r′′(t)∥.
The left side records only how fast the length of the velocity vector changes. The right side also detects a change in the velocity's direction. The left side can even be negative when the object slows down, while the right side is a norm and cannot be negative.
Tangential and normal acceleration
Since velocity is speed times direction,
r′(t)=s′(t)T(t).
Differentiate with the product rule:
a(t)=r′′(t)=s′′(t)T(t)+s′(t)T′(t).
Where T′(t)=0, the definition of N gives
T′(t)=∥T′(t)∥N(t).
The curvature formula κ=∥T′∥/s′ then gives
∥T′(t)∥=κ(t)s′(t).
Putting these together produces the Frenet decomposition of acceleration:
a(t)=s′′(t)T(t)+κ(t)(s′(t))2N(t).
The first term is tangential acceleration. It changes the speed. The second is normal acceleration. It changes the direction and points toward the local bend. There is no binormal component, so the acceleration lies in the plane spanned by T and N. Because these two unit vectors are perpendicular,
∥a(t)∥2=(s′′(t))2+κ(t)2(s′(t))4.
This decomposition requires the Frenet quantities to be defined: r′(t)=0 for T, and T′(t)=0 for the principal normal term.
Constant speed can still mean nonzero acceleration