Directional Derivatives: Measuring Change in a Chosen Direction

Connect partial derivatives, unit vectors, and gradients through a concrete worked example.

A directional derivative measures how quickly a function changes at a point as you move in a specified direction. For a differentiable function, calculate it by combining its partial derivatives with the components of a unit direction vector. The result is a scalar: the instantaneous change in the function per unit distance traveled in that direction. [1], [3]

In this article

1. From axis-by-axis change to any direction
2. What the definition measures
3. The gradient makes the calculation practical
4. Worked example: one point, several possible changes
5. Reading the result without overinterpreting it

1. From axis-by-axis change to any direction

Choose a direction: partial derivatives describe movement along coordinate axes; directional derivatives also describe movement in directions between those axes. [1], [3]
Use a unit vector: this makes the answer a rate per unit distance, rather than a rate that also depends on the length of your chosen vector. [1], [3]
Combine the partial derivatives: for a differentiable function, weight each partial derivative by the corresponding direction component and add. [3]

For a function f(x,y), the partial derivative fₓ measures change as x varies while y stays fixed. Similarly, fᵧ measures change as y varies while x stays fixed. But a direction need not follow either axis: both coordinates can increase together, or one can increase while the other decreases. A directional derivative answers the more general question: at this point, how quickly does f change if I move this way? Specifying the point alone is not enough; the direction is part of the question. [1]

2. What the definition measures

Let the starting point be P = (x₀,y₀), and let u = (a,b) be a unit vector, so a² + b² = 1. Moving a signed distance h along this line takes you to (x₀ + ah, y₀ + bh). The directional derivative is Dᵤf(P) = limₕ→₀ [f(x₀ + ah, y₀ + bh) − f(x₀,y₀)]/h, when this limit exists. The numerator measures the change in the function; the denominator measures the signed step along the chosen direction. Taking the limit turns an average rate into an instantaneous one. [1], [3]

If a direction is supplied as a nonzero vector v, normalize it first: u = v/‖v‖. For example, v = (3,4) has length 5, so its unit vector is (3/5,4/5). If the direction instead comes as an angle θ measured counterclockwise from the positive x-axis, use u = (cos θ, sin θ). These descriptions serve the same purpose: they specify the direction while fixing the distance scale. The zero vector cannot be normalized and does not specify a direction. [1], [3]

3. The gradient makes the calculation practical

For a differentiable function, there is no need to evaluate the limit directly. With u = (a,b), the formula is Dᵤf(P) = fₓ(P)a + fᵧ(P)b. Collect the partial derivatives into the gradient vector, ∇f(P) = (fₓ(P),fᵧ(P)), and this becomes the dot product Dᵤf(P) = ∇f(P) · u. The gradient is a vector containing the coordinate rates of change; the directional derivative is the scalar obtained by combining those rates for one selected direction. [3]

The underlying mechanism is a reduction to one variable. Define g(t) = f(x₀ + at, y₀ + bt), which records the function along the line through P in direction u. Then Dᵤf(P) = g′(0). Applying the chain rule gives fₓ(P)a + fᵧ(P)b. This also explains why partial derivatives are special cases: choosing u = (1,0) leaves only fₓ, while choosing u = (0,1) leaves only fᵧ. Differentiability is the condition behind this shortcut; the limit remains the defining test. [1], [3]

4. Worked example: one point, several possible changes

For an original mathematical example, let f(x,y) = x² + y². Find its directional derivative at P = (1,2) in the direction v = (3,4). First normalize: u = (3/5,4/5). Next differentiate: fₓ = 2x and fᵧ = 2y, so ∇f(1,2) = (2,4). Therefore Dᵤf(1,2) = 2(3/5) + 4(4/5) = 22/5 = 4.4. The function is increasing instantaneously at 4.4 function units per unit distance in this direction. [3]

We can check the answer directly. Along this line, g(h) = f(1 + 3h/5, 2 + 4h/5) = 5 + (22/5)h + h². Thus [g(h) − g(0)]/h = 22/5 + h, whose limit is 22/5. This expansion also exposes an important distinction: for a small displacement h, the linear estimate is Δf ≈ 4.4h, but the exact change here includes h². A directional derivative describes local, first-order change, not an exact prediction for every finite step. [1], [3]

The following comparison keeps the function and starting point fixed. Only the unit direction changes, showing why there is no single slope for this two-variable function at P. [1], [3]

Unit direction at (1,2)	Directional derivative	Local interpretation
(1,0)	2	Increase along the positive x-axis [1], [3]
(0,1)	4	Increase along the positive y-axis [1], [3]
(3/5,4/5)	22/5	Increase in the selected diagonal direction [3]
(−3/5,−4/5)	−22/5	Decrease in the opposite direction [3]
(−2/√5,1/√5)	0	No instantaneous first-order change [3]

5. Reading the result without overinterpreting it

A positive directional derivative indicates local increase; a negative one indicates local decrease. Zero means the instantaneous rate vanishes, not that the function stays constant along the entire line. In our example, moving along u = (−2/√5,1/√5) gives f(P + hu) = 5 + h². The directional derivative is zero at h = 0, yet every nonzero step raises the function value. The first-order change cancels, while the quadratic change remains. [1], [3]

The dot-product formula also identifies the steepest direction. For a nonzero gradient, write Dᵤf = ‖∇f‖ cos φ, where φ is the angle between the gradient and the unit direction. This is greatest when u points along the gradient and least when it points opposite. Here ∇f = (2,4), so the steepest-increase unit direction is (1/√5,2/√5), with rate 2√5. The selected direction (3/5,4/5) produces 4.4, slightly below that maximum. These are statements about the starting point, not a guarantee of the same rate farther along a path. [3], [1]

Frequently asked questions

Why must I normalize the direction vector?

The unit-vector convention measures change per unit distance. In the worked example, using (3,4) directly in the dot product gives 22 rather than 22/5. That larger value corresponds to a parameterization moving five distance units per parameter unit. Normalizing separates the direction from that extra scale factor. [1], [3]

Is the gradient the same as a directional derivative?

No. The gradient collects the partial derivatives into a vector. A directional derivative is a scalar rate for one chosen unit direction. Taking the dot product of the gradient with that unit vector converts the coordinate rates into the requested directional rate. [3]

Does a zero directional derivative mean the function is constant?

No. It means only that the instantaneous rate in that direction is zero at the specified point. In the example above, the perpendicular direction gives f(P + hu) = 5 + h². Its derivative at zero vanishes even though its value changes for every nonzero h. [1], [3]

What happens when I reverse direction?

For a differentiable function, reversing u to −u reverses the sign: ∇f · (−u) = −∇f · u. Thus an instantaneous increase in one direction becomes an equally sized instantaneous decrease in the opposite direction at the same point. [3]

When should I use the limit instead of the gradient formula?

Use the gradient formula when differentiability is established. If that condition is uncertain, return to the directional limit and examine the chosen direction directly. The existence of partial derivatives alone should not be treated as satisfying the differentiability assumption stated in the theorem. [1], [3]

Conclusion

A directional derivative turns a multivariable question into a precise local rate: choose the point, normalize the direction, and—when the function is differentiable—take the gradient’s dot product with that unit vector. Its sign indicates increase or decrease, and its magnitude measures the instantaneous rate per unit distance. The worked example also shows the essential boundary: a derivative captures first-order behavior at a point, not the complete change over a finite journey. [1], [3]

Explore next

Sources

Open a source or expand its cited passages.

Relevance scores are model estimates of how well each source addresses this question, not probabilities or fact-check ratings. Reasoned guidance is identified on hover.