Calculus reference
What kind of answer is requested?
Section titled “What kind of answer is requested?”| Expression | What it returns | For f(x) = x² |
|---|---|---|
| An output value | 9 | |
| A function describing the rate of change | ||
| The rate at one input | 6 | |
| A predicted output change for a small input change h | ||
| A predicted new output | ||
| Accumulated signed area over an interval | ||
| A family of antiderivatives on an interval |
Derivative rules
Section titled “Derivative rules”Assume u and v are differentiable where the rule is applied, and c is constant.
| Operation | Rule | Example |
|---|---|---|
| Sum | ||
| Constant multiple | ||
| Power | ||
| Product | ||
| Composition | ||
| Quotient |
The quotient requires . The power rule applies to real n on ; particular powers have larger domains. The derivative of is for , not at zero.
Other useful derivatives are , for , and for . The formulas and use radians.
A zero derivative is a candidate
Section titled “A zero derivative is a candidate”An interior local maximum or minimum of a differentiable function has . The converse does not hold: has a zero derivative at zero and keeps increasing.
At a stationary point, a positive second derivative gives a local minimum, a negative one gives a local maximum, and zero gives no verdict. Also check boundaries and points where the derivative does not exist. For a global claim, compare the whole stated domain.
Integrals
Section titled “Integrals”For continuous f on and an antiderivative F with ,
| Integrand | An antiderivative | Condition |
|---|---|---|
| , on an interval where the power rule applies | ||
| Interval does not cross zero | ||
| All real x | ||
| c | cx | c constant |
A definite integral counts area below the axis negatively. The total geometric area is when .
The unweighted average height is . An expectation instead weights by a probability distribution; for a continuous X, when the expectation exists.
Partial derivatives and gradients
Section titled “Partial derivatives and gradients”For , hold the other input fixed:
At , the gradient is . A small step gives the linear prediction .
For steps of the same Euclidean length, the nonzero gradient gives the direction of greatest first-order increase. A sufficiently small step against it decreases a differentiable function:
Here θ is the vector of parameters, L the loss, and η a positive learning rate. A large step can increase the loss. A zero gradient alone does not prove a minimum.
Three notation questions
Section titled “Three notation questions”Why can h approach zero after we divided by h?
The quotient is evaluated for nonzero h. Its limit describes the value those quotients approach. For at 3, the quotient is and its limit is 6. We never evaluate .
What does dx mean?
In the integral notation used here, dx identifies the integration variable. The finite rectangle widths are Δx; the integral is their limiting sum. In df/dx, the notation identifies a derivative. Differential notation has further uses, but treating dx as literally zero does not justify these operations.
What is the difference between d and ∂?
A partial derivative holds other independent inputs fixed. A total derivative follows how the input actually changes. For , the partial derivative in x is y. If instead y also changes as , then and the derivative along that path is . State what is fixed before differentiating.
Practice these distinctions or apply a gradient to a fitted line.