The Transfer Function

Why the transfer function matters

Before Laplace methods, analyzing a controlled machine meant solving its differential equations by hand for every input — a fresh integration each time the forcing changed. The transfer function collapses that work into a single algebraic object. Because the Laplace transform maps differentiation to multiplication by s, a differential equation of order n becomes an n-th degree polynomial equation, and the system's entire dynamic personality — how fast it responds, whether it overshoots, whether it is even stable — is read directly from where the roots of that polynomial sit in the complex plane.

• One object, all inputs. Multiply G(s) by any input transform X(s) to get the output transform Y(s); invert to recover the time response.
• Diagrams become arithmetic. Series blocks multiply, parallel blocks add, and feedback loops close with G/(1+GH) — no calculus required to reduce a plant model.
• Stability at a glance. The sign of the real part of every pole answers the single most important question in control: will the loop settle or run away?
• The gateway to design tools. Root locus, Bode plots, Nyquist criterion, gain and phase margins, and lead-lag compensator design all operate on the transfer function.
• Universal reach. The same G(s) machinery describes a servo motor, an aircraft pitch loop, an op-amp filter, a chemical reactor's temperature, and a suspension damper.

From differential equation to G(s), step by step

The transfer function is defined only for LTI systems with zero initial conditions. Under those assumptions the Laplace transform of the n-th derivative reduces to sⁿ·(transform), so the recipe is mechanical:

1. Write the governing ODE relating output y(t) to input x(t).
2. Laplace-transform both sides with all initial conditions zero, replacing each d^k/dt^k with s^k.
3. Collect Y(s) and X(s) and form the ratio G(s) = Y(s)/X(s).
4. Factor numerator and denominator to expose zeros and poles.

The canonical second-order mechanical example is a mass–spring–damper driven by force f(t):

m·y″ + c·y′ + k·y = f(t)  ⟶  (m·s² + c·s + k)·Y(s) = F(s)  ⟶  G(s) = Y(s)/F(s) = 1 / (m·s² + c·s + k)

where m is mass (kg), c is the viscous damping coefficient (N·s/m), k is stiffness (N/m), y is displacement (m), and f is force (N). Written in the standard normalized form,

G(s) = ω_n² / (s² + 2ζω_n·s + ω_n²), with ω_n = √(k/m) and ζ = c / (2√(k·m)),

where ω_n is the undamped natural frequency (rad/s) and ζ is the dimensionless damping ratio. The two poles sit at s = −ζω_n ± ω_n√(ζ²−1). For 0 < ζ < 1 they form a complex conjugate pair −ζω_n ± jω_d with damped frequency ω_d = ω_n√(1−ζ²), producing a decaying oscillation — the ringing you feel in a car suspension after a pothole.

Poles, zeros, and what the response looks like

Every pole contributes one natural mode to the time response. A real pole at s = −a contributes a decaying exponential e^−at with time constant τ = 1/a; the larger a is (the farther left the pole), the faster that mode dies out. A complex conjugate pole pair contributes a damped sinusoid whose envelope decays as e^−σt (σ = |real part|) and whose ringing frequency is the imaginary part ω_d. Zeros add no modes of their own but reweight how strongly each pole's mode appears; a zero in the right-half plane produces the counterintuitive initial undershoot of a non-minimum-phase system — the output first moves the wrong way, as an aircraft briefly loses altitude when the elevator commands a climb.

Pole location on the s-plane and the resulting time-domain behavior

Pole location	Real part Re(s)	Time response	Stability
Left-half plane, real	< 0	Decaying exponential e−at	Stable
Left-half plane, complex pair	< 0	Damped oscillation e−σt·sin(ωdt)	Stable
On imaginary axis, simple	= 0	Sustained oscillation sin(ωt)	Marginally stable
On imaginary axis, repeated	= 0	Growing t·sin(ωt)	Unstable
Right-half plane	> 0	Growing exponential e+at	Unstable

Stability, the characteristic equation, and feedback

Stability is decided entirely by the denominator. Set the denominator of the closed-loop transfer function to zero and you have the characteristic equation, D(s) = 0, whose roots are the closed-loop poles. A continuous-time LTI system is bounded-input bounded-output (BIBO) stable if and only if every root of D(s) has a strictly negative real part. In practice you rarely factor high-order polynomials by hand: the Routh–Hurwitz criterion counts right-half-plane roots directly from the polynomial coefficients, flagging instability without ever solving for the poles.

Feedback is what makes the characteristic equation interesting. For a forward path G(s) wrapped in a feedback path H(s), block-diagram algebra gives the closed-loop transfer function

T(s) = G(s) / (1 + G(s)·H(s))  (negative feedback),

so the closed-loop characteristic equation is 1 + G(s)H(s) = 0, or equivalently 1 + L(s) = 0 where L(s) = G(s)H(s) is the open-loop transfer function. This single equation is the object that the root locus sweeps as loop gain varies, and that the Nyquist criterion encircles in the complex plane. The block-diagram reduction rules are worth memorizing:

Block-diagram algebra: reducing interconnected blocks to one G(s)

Interconnection	Equivalent transfer function	Note
Series (cascade)	G1(s)·G2(s)	Blocks multiply; order irrelevant for LTI
Parallel (summed)	G1(s) + G2(s)	Same input, outputs added at a summing junction
Negative feedback	G(s) / (1 + G(s)·H(s))	Standard closed loop
Positive feedback	G(s) / (1 − G(s)·H(s))	Sign flips; often destabilizing
Unity feedback	G(s) / (1 + G(s))	H(s) = 1; sensor is ideal

Worked example: closing a loop around an integrator plant

Suppose a plant behaves like a motor whose speed integrates its input with a lag, G(s) = K / [s(s + 4)], and we close a unity-feedback loop (H = 1). The closed-loop transfer function is

T(s) = G / (1 + G) = K / [s² + 4s + K].

The characteristic equation is s² + 4s + K = 0. Comparing with s² + 2ζω_ns + ω_n² gives ω_n = √K and 2ζω_n = 4, so ζ = 2/√K. Pick the loop gain K to place the poles where you want them:

• K = 4: ζ = 1, poles both at s = −2 — critically damped, fastest response with no overshoot.
• K = 16: ζ = 0.5, poles at s = −2 ± j2√3 — about 16% overshoot, ω_d ≈ 3.46 rad/s of ringing.
• K = 100: ζ = 0.2, poles at s = −2 ± j9.8 — lightly damped, ~52% overshoot; the loop rings for many cycles.

Note that for any positive K both poles keep a real part of −2, so this particular loop stays stable no matter how hard you push the gain — but the damping ratio, overshoot, and settling behavior swing wildly. That is the everyday design trade the transfer function exposes: raise gain for speed, pay in overshoot and oscillation. Settling time here is roughly t_s ≈ 4/(ζω_n) = 4/2 = 2 s (the 2% criterion), independent of K in this special case.

Common misconceptions and failure modes

• "It works with initial conditions." The transfer function is defined only for zero initial conditions. Nonzero states contribute extra terms; use state-space if you need them.
• "Zeros affect stability." Only poles set stability. Zeros reshape the response and can cause overshoot or undershoot, but a stable system stays stable regardless of zero location.
• "Cancel a bad pole with a zero." Pole-zero cancellation is fragile — a right-half-plane pole cancelled on paper still exists physically as a hidden unstable mode. Never cancel across the imaginary axis; state-space controllability/observability catches what G(s) hides.
• "Any input works." Transfer functions describe LTI systems only. Saturation, backlash, and other nonlinearities break the ratio; you must linearize about an operating point first.
• "More gain is always faster." Higher loop gain speeds response but pushes closed-loop poles toward — and eventually across — the imaginary axis, driving overshoot then outright instability.
• "Marginally stable is fine." Poles exactly on the jω-axis give undamped oscillation that any real disturbance or modeling error can tip into instability; treat marginal stability as a design failure.