Differentiating Products and Compositions Using the Product and Chain Rules

First: identify the structure of the function

A derivative measures local sensitivity. If a scalar function $f(x)$ has derivative $f^{\prime }(x)$, then for a small input change $\Delta x$,

So $f^{\prime }(x)$ is the local multiplier that converts a small change in input into an approximate change in output.

For now, both the input and output are single numbers. The input might be a conventional variable $x$, or it could be one scalar model parameter such as a logistic-regression coefficient $w$.

Before differentiating, inspect the outermost operation in the formula.

For the scope of this lesson, the important categories are:

The most useful preliminary derivative facts are:

Also, derivatives distribute over sums:

The rules in this lesson extend these basic facts to expressions built from multiple pieces.

A small but valuable habit is to simplify only when it genuinely helps. For example,

so differentiating the simplified expression is easier:

But you cannot simplify $(x^2+1)(x-4)$ in a way that avoids reasoning about how both factors change, unless you first expand it accurately.

The product rule: both factors can change

Suppose

is the product of two functions. The product rule is

In words:

Differentiate the first factor and leave the second unchanged, then add the first unchanged factor times the derivative of the second.

A frequent error is to write

This is false. It misses the fact that both factors change as $x$ changes.

Why does a sum appear?

Think of $u(x)v(x)$ as the area of a rectangle with side lengths $u(x)$ and $v(x)$. A small input change alters both sides. The resulting area change has:

• a thin strip due to the change in $u$, with approximate area $v\Delta u$;
• a thin strip due to the change in $v$, with approximate area $u\Delta v$;
• a tiny corner $\Delta u\Delta v$, which becomes negligible compared with $\Delta x$ in the derivative limit.

That leaves the two main contributions:

Dividing by $\Delta x$ and taking the limiting rate of change gives the product rule.

Example: a product involving a power and sine

Let

Define the factors:

Their derivatives are:

Apply the product rule:

Notice that the answer has two terms. The first records how the changing $x^2$ factor affects the product; the second records how the changing sine factor affects it.

A compact mnemonic sometimes used is “left derivative right, plus left right derivative.” The formula is more reliable than the mnemonic:

Keep the undifferentiated factor exactly as it was, evaluated at the original input.

The chain rule: changes pass through nested functions

A composite function has a function inside another function. For example,

has two stages:

1. Calculate the inner value:

2. Raise that value to the fourth power:

Thus,

The chain rule states:

In standard verbal form:

Differentiate the outside function while keeping the inside unchanged; then multiply by the derivative of the inside.

The derivative of the outside function measures how the final output reacts to a change in the intermediate value $h$. The derivative of the inner function measures how that intermediate value reacts to a change in $x$. Their product gives the final sensitivity to $x$.

A practical template: powers of functions

The most common chain-rule pattern at this stage is:

Then

The final factor $u^{\prime }(x)$ is essential. It is the derivative of the inner expression.

Example: a nested power

Return to

Start with the outer operation: “raise something to the fourth power.” Its derivative is

Keep the inner expression in place:

Now multiply by the derivative of the inner expression:

Therefore,

or, after collecting constants,

The incorrect answer

has differentiated the outer layer but ignored the fact that the inner value changes at rate $6x$.

A relevant example: sigmoid of a linear score

In the previous lesson, the sigmoid function was

Suppose a one-feature model has score

and probability

This is a composition: the outer function is $\sigma$, and the inner function is $2x-1$.

A useful sigmoid identity is:

Because the inner score has derivative

the chain rule gives:

The multiplier $2$ is not optional: it tells us that the score itself changes twice as fast as $x$.

When product and chain rules appear together

Real model expressions often contain both structures. Consider

At the top level, this is a cube of an inner expression, so start with the chain rule:

The derivative left inside is a product, so use the product rule:

Substituting it back gives a clean factored answer:

There is no need to expand this unless a later calculation specifically benefits from expansion. The factored form exposes the computational structure: outer cube, then inner product.

There can sometimes be more than one valid route. In this example, exponent laws allow

After that rewrite, one could use the product rule first and then the chain rule on ${\sin }^3(x)$. Both routes lead to the same derivative. The best route is generally the one that preserves a clear structure and minimizes algebraic opportunities for error.

A dependable procedure is:

1. Locate the outermost operation.
2. Apply that rule once.
3. Do not alter the inner expression prematurely.
4. Differentiate the remaining inner pieces recursively.
5. Simplify only after the derivative structure is correct.

From scalar derivatives to a model-loss derivative

The chain rule is the core mechanism by which learning algorithms connect a loss back to the parameters that produced it.

Consider one labeled observation, with:

• feature value $x$,
• observed binary outcome $y$,
• scalar weight $w$,
• fixed intercept $b$.

A logistic model proceeds in three stages:

Here $L$ is a squared-error loss for a single observation. It is not the usual preferred loss for logistic regression, but it makes the chain-rule structure especially visible.

When differentiating with respect to $w$, treat $x$, $y$, and $b$ as fixed. Trace the calculation from the loss backward:

By applying the chain rule across all three stages,

Thus,

This product has a direct interpretation:

• $2(p-y)$ measures the local error direction and magnitude;
• $p(1-p)$ measures the sigmoid’s local sensitivity;
• $x$ measures how strongly changing $w$ changes the score.

The full model computation goes forward from $w$ to $z$, then $p$, then $L$. The derivative calculation carries sensitivity backward from $L$ toward $w$. Later, this same idea will be extended from one parameter to many parameters; that extension is the gradient and, in neural networks, backpropagation.

A numerical sanity check in Python

For a complicated derivative, a central finite-difference approximation can help detect an algebra mistake:

import numpy as np

x = 1.5
y = 1.0
b = -0.3

def sigmoid(z):
    return 1 / (1 + np.exp(-z))

def loss(w):
    p = sigmoid(w * x + b)
    return (y - p) ** 2

def analytic_derivative(w):
    p = sigmoid(w * x + b)
    return 2 * (p - y) * p * (1 - p) * x

w = 0.7
h = 1e-6

finite_difference = (loss(w + h) - loss(w - h)) / (2 * h)

print("Analytic derivative: ", analytic_derivative(w))
print("Finite difference:  ", finite_difference)

The two printed values should be very close. A finite difference is an approximation, not a proof, but it is an effective implementation-level check that your product and chain-rule derivation is plausible.

Key takeaways

The product rule applies when two changing functions are multiplied:

The chain rule applies when one function is nested inside another:

For powers of an inner function:

The main procedural skill is not memorizing a large list of rules. It is accurately reading an expression’s structure: identify its outer operation, differentiate one layer, and continue inward until every changing component has been accounted for.

Next, you will extend this scalar reasoning to losses with several parameters by computing gradients from partial derivatives.