Chain Rule Partial Derivatives Multiple Variables

You're staring at a function like z = f(x, y)* where x = g(t)* and y = h(t), and you need dz/dt. You nodded along. So naturally, your professor derived it in three lines. Then you got to the homework and realized you have no idea which derivative goes where.

Sound familiar?

The chain rule for multiple variables isn't harder than the single-variable version. More places to lose track of what depends on what. Because of that, more moving parts. It's just... This leads to messier. And most textbooks explain it by throwing a tree diagram at you and calling it a day.

Let's actually walk through it. No tree diagrams required — unless you want one.

What Is the Chain Rule for Partial Derivatives

At its core, the chain rule answers one question: if I wiggle the input, how does the output change — when the input affects the output through intermediate variables?*

Single variable: y = f(u), u = g(x). Also, then dy/dx = (dy/du)(du/dx). Think about it: you learned this in Calc I. Day to day, it's a chain. One link after another.

Multiple variables: z = f(x, y)* where x = g(s, t)* and y = h(s, t). Now z depends on s and t through two intermediate variables. Even so, wiggle s — x changes, y changes, both affect z. The total effect is the sum of both pathways Turns out it matters..

That's it. That's the whole intuition. Every formula you'll see is just bookkeeping for "sum over all pathways.

The Two Main Flavors

You'll encounter two setups constantly:

Case 1: One independent variable
z = f(x, y), x = g(t), y = h(t)* → find dz/dt*
This is the direct generalization of the single-variable chain rule Which is the point..

Case 2: Two (or more) independent variables
z = f(x, y), x = g(s, t), y = h(s, t)* → find ∂z/∂s and ∂z/∂t
This is where partial derivatives enter the picture. You're holding one independent variable fixed while differentiating with respect to the other.

There's also Case 3: Mixed dependencies — z = f(x, y), x = g(s, t), y = h(t)* where y doesn't depend on s. The rule still works; you just have fewer pathways No workaround needed..

Why It Matters / Why People Care

You might wonder: when would I ever actually use this?*

Short answer: constantly. Any time a quantity depends on other quantities that themselves depend on something else — which is basically every real-world model Worth keeping that in mind. Still holds up..

Physics: temperature T at a point (x, y, z) in a room, but you're moving along a path (x(t), y(t), z(t)). How fast does the temperature you feel change? dT/dt = ∂T/∂x dx/dt + ∂T/∂y dy/dt + ∂T/∂z dz/dt*. That's the chain rule.

Economics: profit P depends on labor L and capital K. But L and K depend on time t and interest rate r. You need ∂P/∂r to understand how policy changes affect the bottom line Not complicated — just consistent..

Machine learning: backpropagation is the chain rule for partial derivatives, applied to a computational graph with millions of variables. Every gradient calculation in a neural network is a chain rule computation.

Implicit differentiation? Day to day, chain rule. Related rates? Chain rule. Which means change of variables in integrals? Chain rule (the Jacobian determinant is the multivariable chain rule in disguise).

If you're shaky on this, everything downstream gets shakier.

How It Works

Let's derive the formulas from scratch. Not because you need to re-derive them on exams — you don't — but because seeing where they come from makes them impossible to forget.

Case 1: One Independent Variable (t)

z = f(x, y)*
x = g(t)*
y = h(t)*

Assume f, g, h* are differentiable. z is ultimately a function of t alone. By definition of the derivative:

dz/dt = lim_{Δt→0} [f(x(t+Δt), y(t+Δt)) - f(x(t), y(t))] / Δt*

Add and subtract f(x(t+Δt), y(t))* in the numerator:

= lim [f(x(t+Δt), y(t+Δt)) - f(x(t+Δt), y(t))] / Δt + lim [f(x(t+Δt), y(t)) - f(x(t), y(t))] / Δt

First term: y changes, x held at x(t+Δt). As Δt→0, x(t+Δt)→x(t). This becomes ∂f/∂y · dy/dt Easy to understand, harder to ignore. Which is the point..

Second term: x changes, y held fixed at y(t)*. Becomes ∂f/∂x · dx/dt.

Result:
dz/dt = ∂f/∂x · dx/dt + ∂f/∂y · dy/dt*

Notice the pattern: ∂f/∂(intermediate) × d(intermediate)/dt, summed over all intermediates.

Case 2: Two Independent Variables (s, t*)

z = f(x, y)*
x = g(s, t)*
y = h(s, t)*

Now z is a function of s and t. To find ∂z/∂s, treat t as constant and differentiate with respect to s. Exact same logic as Case 1, but with partials:

∂z/∂s = ∂f/∂x · ∂x/∂s + ∂f/∂y · ∂y/∂s

Similarly:

∂z/∂t = ∂f/∂x · ∂x/∂t + ∂f/∂y · ∂y/∂t

The General Pattern (Memorize This, Not the Formulas)

For z = f(u₁, u₂, ..., uₙ)* where each uᵢ = uᵢ(v₁, v₂, ..., vₘ)*:

∂z/∂vⱼ = Σᵢ (∂f/∂uᵢ) · (∂uᵢ/∂vⱼ)

Sum over all intermediate variables uᵢ. Each term: derivative of outer function with respect to that intermediate, times derivative of that intermediate with respect to the variable you're differentiating The details matter here..

That's the only formula you need. Everything else is a special case.

A Concrete Example

z = x²y + sin(y)*
x = s + t*
y = st*

Find ∂z/∂s and ∂z/∂t.

First, compute the partials of z with respect to its

direct inputs:

∂z/∂x = 2xy
∂z/∂y = x² + cos(y)

Next, determine how x and y respond to changes in s and t:

∂x/∂s = 1
∂x/∂t = 1
∂y/∂s = t
∂y/∂t = s

Now apply the general pattern. For ∂z/∂s:

∂z/∂s = (∂z/∂x)(∂x/∂s) + (∂z/∂y)(∂y/∂s)
= (2xy)(1) + (x² + cos(y))(t)
= 2xy + t(x² + cos(y))

Substituting back x = s + t and y = st:

∂z/∂s = 2(s + t)(st) + t[(s + t)² + cos(st)]

For ∂z/∂t:

∂z/∂t = (∂z/∂x)(∂x/∂t) + (∂z/∂y)(∂y/∂t)
= (2xy)(1) + (x² + cos(y))(s)
= 2xy + s(x² + cos(y))

Substituting the same expressions:

∂z/∂t = 2(s + t)(st) + s[(s + t)² + cos(st)]

Why This Matters

The chain rule isn't just calculus—it's the backbone of modern computational thinking. In machine learning, when you adjust weights in a neural network, you're implicitly computing chain rule derivatives through layers of transformations. In economics, when interest rates shift and ripple through financial models, you're seeing the chain rule in action.

The pattern is always the same: identify your intermediate variables, compute their rates of change, multiply by how your final output depends on them, then sum across all paths. Master this, and you'll recognize it everywhere—from thermodynamics to computer graphics to financial modeling Small thing, real impact..

The chain rule teaches us that complex systems rarely change for just one reason. Which means understanding those multiple influences requires breaking down composite relationships into their constituent derivatives. That decomposition—the ability to see how changes propagate through interconnected systems—is what makes the chain rule indispensable.

Visualizing the Flow: Tree Diagrams

When the number of variables grows, a tree diagram keeps the accounting straight. Here's the thing — branch downward to the intermediate variables ($x, y$). Branch again to the independent variables ($s, t$). Draw the dependent variable $z$ at the top. Every path from $z$ to the target variable represents a term in the sum.

          z
         / \
        x   y
       / \ / \
      s   t s  t

To find $\partial z/\partial s$, trace every path from $z$ to $s$:

1. $z \to x \to s$ $\implies \frac{\partial z}{\partial x} \frac{\partial x}{\partial s}$
2. $z \to y \to s$ $\implies \frac{\partial z}{\partial y} \frac{\partial y}{\partial s}$

Sum them. The diagram makes the general pattern $\partial z/\partial v_j = \sum_i (\partial f/\partial u_i)(\partial u_i/\partial v_j)$ visually obvious: sum over all paths connecting the output to the input.

Case 3: Nested Dependencies (The "Real World" Scenario)

Often, intermediate variables depend on each other*, not just on the root inputs. Suppose: $z = f(x, y), \quad x = g(s, t), \quad y = h(x, t)$ Here $y$ depends on $x$ explicitly. The tree grows a cross-branch:

          z
         / \
        x   y
       / \  ↑
      s   t └── (y depends on x too)

Now $\partial z/\partial s$ requires the chain rule inside* the chain rule. Start with the general pattern: $\frac{\partial z}{\partial s} = \frac{\partial z}{\partial x}\frac{\partial x}{\partial s} + \frac{\partial z}{\partial y}\frac{\partial y}{\partial s}$

But $\partial y/\partial s$ isn't a primitive—it must be expanded because $y = h(x, t)$ and $x = g(s, t)$: $\frac{\partial y}{\partial s} = \frac{\partial h}{\partial x}\frac{\partial x}{\partial s} + \frac{\partial h}{\partial t}\underbrace{\frac{\partial t}{\partial s}}_{=0} = \frac{\partial y}{\partial x}\frac{\partial x}{\partial s}$

It sounds simple, but the gap is usually here Worth keeping that in mind..

Substitute back: $\frac{\partial z}{\partial s} = \frac{\partial z}{\partial x}\frac{\partial x}{\partial s} + \frac{\partial z}{\partial y}\left(\frac{\partial y}{\partial x}\frac{\partial x}{\partial s}\right)$ Factor out $\partial x/\partial s$: $\frac{\partial z}{\partial s} = \left( \frac{\partial z}{\partial x} + \frac{\partial z}{\partial y}\frac{\partial y}{\partial x} \right) \frac{\partial x}{\partial s}$

The lesson: The tree diagram still works. Just follow every* directed path from $z$ to $s$. The cross-branch $y \leftarrow x$ creates a second path: $z \to y \to x \to s$. The algebraic sum of paths handles the nesting automatically.

Common Pitfalls (And How to Avoid Them)

1. Confusing $\partial z/\partial x$ with $dz/dx$ In Case 2, $z = f(x,y)$ and $x=g(s,t)$. The symbol $\partial z/\partial x$ means "hold $y$ constant, vary $x$." It is a property of $f$ alone. It does not mean "total derivative of $z$ with respect to $x$ accounting for how $y$ changes with $x$." Label your derivatives carefully: $f_x$, $g_s$, $h_t$ removes ambiguity.

2. Forgetting to Evaluate at the Correct Point If asked for $\partial z/\partial s$ at $(s,t)=(1,2)$, you must compute $

the partial derivatives for every function in the chain first, and then* plug in the values. A common mistake is to plug in the values too early, especially if the variables are linked through intermediate functions. Always maintain the functional form until the final summation step.

3. Misinterpreting the "Constant" Variable When calculating $\partial z/\partial x$ in a multivariable context, you are implicitly assuming all other variables (like $y$) are constant. Still, if $y$ is actually a function of $x$, the partial* derivative $\partial z/\partial x$ ignores that connection, whereas the total* derivative accounts for it. If you are traversing a computational graph, you must distinguish between "local" partial derivatives (the edges in your diagram) and the "global" total derivative (the sum of the paths) The details matter here..

Summary Table: The Computational Perspective

To solidify your intuition, view the chain rule through the lens of a computer scientist or a machine learning engineer. In modern neural networks, this is known as Backpropagation.

Conclusion

The chain rule is often taught as a rigid formula to be memorized, but it is actually a geometric principle. Whether you are dealing with simple direct dependencies or complex, nested architectures where variables influence one another through multiple channels, the rule remains the same: The total change in an output is the sum of the changes propagated through every possible path.

People argue about this. Here's where I land on it Which is the point..

By visualizing dependencies as a directed graph, you transform a daunting algebraic problem into a simple task of pathfinding. Once you master the ability to trace these paths and correctly distinguish between local partial derivatives and global total derivatives, you possess the fundamental tool required for multivariable calculus, thermodynamics, and the training of deep neural networks.