Circuit Sensitivity Analysis: ∂H/∂Component

Every design has tolerances — resistors at ±1%, capacitors at ±10% or worse, transistor betas that vary 3:1 across process corners. Sensitivity analysis answers the question that matters after the ideal transfer function is derived: how much does the circuit's behavior actually move when a component drifts? It's the bridge between a nominal SPICE result and a design that survives manufacturing, temperature, and aging.

Defining Sensitivity

For a transfer function H depending on a component value x (resistance, capacitance, gain, etc.), the raw sensitivity is the partial derivative:

S = ∂H/∂x

This has units (e.g., V/Ω), which makes it awkward to compare across component types. Circuit designers almost always use the normalized (relative) sensitivity instead:

S(H,x) = (∂H/∂x) · (x/H) = ∂(ln H)/∂(ln x)

This is dimensionless and reads directly as: a 1% change in x produces S% change in H. A sensitivity of −1 means H drops 1% for every 1% rise in x. A sensitivity of 0 means H is locally immune to that component — the ideal case for a bias network or reference.

Why It's Not Just "Take the Derivative"

For simple functions this is one line of calculus, but two complications drive most of the design work:

• Interacting components. Real circuits have many parameters at once. The total variation in H combines every component's contribution, and if components are uncorrelated, worst-case addition is pessimistic — statistical (RSS) combination is usually the realistic figure.
• Nonlinear and non-monotonic dependence. Resonant circuits, feedback loops near instability, and multi-pole filters can have sensitivity that spikes sharply near critical points (e.g., near a pole pair with ζ → 0), so the sensitivity computed at nominal values may not represent behavior across the full tolerance range — a small-signal (linearized) result only holds locally.

Combining Multiple Component Tolerances

If H = H(x₁, x₂, ..., xₙ) and each xᵢ has independent relative tolerance Δxᵢ/xᵢ, the first-order relative variation in H is:

ΔH/H ≈ Σ S(H,xᵢ) · (Δxᵢ/xᵢ)          (worst-case, linear sum of magnitudes)

ΔH/H(rss) ≈ √[ Σ (S(H,xᵢ) · Δxᵢ/xᵢ)² ]    (statistical, root-sum-square)

Worst-case sums every contribution as if all components drifted to their limits simultaneously and in the worst direction — realistic for guaranteeing a hard spec but often overly conservative. RSS assumes independent, random (typically normal) distributions and predicts the standard deviation of H — the right model for yield estimation.

Worked Example: RC Low-Pass Corner Frequency

Take a first-order RC low-pass filter: ω₀ = 1/(RC).

Step 1 — sensitivity to R:

∂ω₀/∂R = −1/(R²C)
S(ω₀,R) = (∂ω₀/∂R)·(R/ω₀) = [−1/(R²C)]·R·(RC) = −1

Step 2 — sensitivity to C: by symmetry, S(ω₀,C) = −1 as well.

Step 3 — interpretation: both sensitivities are exactly −1, meaning ω₀ tracks R and C directly and equally: a +1% R or +1% C each produces a −1% shift in ω₀. This is expected — ω₀ is a simple inverse product, so log(ω₀) = −log(R) − log(C), and the partial log-derivatives are trivially −1.

Step 4 — combine tolerances. Suppose R is a 1% resistor and C is a 10% capacitor (typical for ceramic/film without tight binning):

Worst-case: |ΔR/R| + |ΔC/C| = 1% + 10% = 11%
RSS:        √(1%² + 10%²) = √(1 + 100)% ≈ 10.05%

Check: the capacitor dominates almost completely in both models, as expected since its tolerance is 10× the resistor's. Worst-case (11%) and RSS (~10.05%) nearly coincide because one term dwarfs the other — confirms the arithmetic is consistent with intuition, and shows tightening R further would barely help; the leverage is in the capacitor spec.

Where Sensitivity Becomes the Design Driver

Practical Workflow

1. Derive H(s) symbolically in terms of component values (or extract it from a linearized SPICE small-signal model).
2. Differentiate or perturb — either compute ∂H/∂x analytically for simple cases, or run SPICE .STEP sweeps (perturb each component ±1% one at a time) for anything with more than two or three reactive elements.
3. Normalize each result to relative sensitivity so component types can be compared on equal footing.
4. Combine using worst-case for hard guarantees (safety-critical, contractual specs) or RSS for realistic yield/tolerance-stack predictions.
5. Rank contributors — usually one or two components dominate (as in the RC example above); spend tightening budget there, not uniformly across every part.

Key Takeaways

• Relative sensitivity S(H,x) = (∂H/∂x)·(x/H) gives a dimensionless, comparable measure of how much a transfer function moves per percent change in a component.
• For simple inverse/product relationships (like ω₀ = 1/RC), sensitivities are often exactly ±1; more complex topologies (active filters, oscillators) can have sensitivities well above 1, especially near high-Q or near-instability conditions.
• Combine multiple component tolerances with worst-case (linear sum) for guaranteed bounds, or RSS for realistic statistical variation — RSS is usually the better predictor of actual yield.
• The dominant tolerance term (often the widest-tolerance component, e.g. a ceramic capacitor) usually sets the overall variation — identify it before spending effort tightening others.
• Sensitivity analysis directly drives topology choice in precision references, active filters, and oscillators, where minimizing ∂H/∂x for critical components is a primary design goal, not an afterthought.