Frequency response function

1 Introduction to Frequency Response Functions

A frequency response function (FRF) characterizes how a linear dynamical system transforms an input into an output at each frequency. Unlike time-domain descriptions that track signals as they evolve, FRFs express the response in the frequency domain, enabling direct inspection of resonance behavior, phase delays, and frequency-dependent gains.

1.1 Motivation and common use cases

Engineers use FRFs to connect experimentally measurable quantities—such as applied force to measured displacement, or driving voltage to resulting current—to a system’s underlying dynamics. Typical motivations include:

identifying resonant frequencies and their damping trends,
comparing an experimentally observed dynamic model to a simulated one,
diagnosing changes in stiffness or mass by monitoring shifts in FRF features,
supporting control and design workflows that depend on frequency-domain behavior.

FRFs are particularly valuable when the system is approximately linear over the excitation level and when frequency-dependent effects dominate the behavior.

1.2 Relationship to transfer functions and dynamic models

For linear time-invariant (LTI) systems, FRFs are closely related to transfer functions. A transfer function expresses the output-to-input ratio in the frequency domain, usually as a complex-valued function of frequency (or Laplace variable). When measurements are performed under sinusoidal steady-state conditions or broadband excitations that allow frequency-domain estimation, the resulting FRF can be interpreted as an experimental counterpart to the theoretical transfer function of the same input-output pair.

In dynamic models (e.g., modal, state-space, or differential-equation based), FRFs serve as a bridge between parameters in the model and measurable frequency-dependent responses, facilitating model validation.

1.3 Scope: linear time-invariant vs. linear time-varying contexts

The classical FRF framework assumes linearity, and most standard definitions presume time invariance so that the response at a given frequency is consistent across time. In linear time-varying (LTV) situations, a strict single FRF may no longer exist; nevertheless, frequency-domain descriptions can still be used locally or via time–frequency methods. In that setting, “FRF” terminology may refer to approximate or averaged frequency-domain transfer relationships computed over selected time intervals.

2 Mathematical Foundations

FRFs rely on representing signals and systems in the frequency domain, typically using Fourier analysis. The core objects are complex-valued frequency-dependent ratios or relationships among spectral quantities.

2.1 System representation in the frequency domain

2.1.1 Input–output relations (transfer-function form)

For an LTI system with input \(u(t)\) and output \(y(t)\), the frequency-domain relationship can be written as \[ Y(\omega) = H(\omega)U(\omega), \] where \(H(\omega)\) is the frequency response (transfer-function) and \(\omega\) is angular frequency. The FRF for a given input-output pairing is often identified with this complex ratio (with appropriate scaling and conventions).

In experimental settings, \(U(\omega)\) and \(Y(\omega)\) are replaced by estimated spectra from measured records, and the FRF is computed from those estimates (see Sections 4.3 and 4.2).

2.1.2 Impulse response and frequency-domain equivalence

For a stable LTI system, the impulse response \(h(t)\) and the frequency response \(H(\omega)\) are Fourier transform pairs: \[ H(\omega) = \int_{-\infty}^{\infty} h(t)e^{-j\omega t}\,dt. \] This provides an alternative view: FRFs can be derived either by solving the system in the frequency domain or by transforming its impulse response. The impulse response perspective is also useful for interpreting causality and stability constraints.

2.1.3 Fourier transform conventions and sign/phase conventions

FRFs are complex quantities and therefore depend on Fourier transform sign conventions. A change from \(e^{-j\omega t}\) to \(e^{+j\omega t}\), or different definitions of phase reference, effectively conjugates or shifts the phase behavior. Consequently, reported phase in FRFs must be interpreted relative to the adopted convention and the chosen measurement reference (e.g., which channel leads or lags).

Practical workflows often include phase unwrapping and reference alignment so that phase trends across frequency can be meaningfully compared.

2.2 FRF definitions for different measurement types

2.2.1 SISO FRFs (single-input single-output)

A single-input single-output (SISO) FRF relates one input channel to one output channel. Under LTI assumptions, the FRF is a complex ratio representing both amplitude scaling and phase shift between the output and input at each frequency.

2.2.2 MIMO FRFs (multiple-input multiple-output)

In multiple-input multiple-output (MIMO) systems, multiple inputs can influence multiple outputs. A generalization uses a matrix of FRFs, with each element representing the frequency response from a specific input to a specific output while other inputs may be treated through measurement design. When multiple excitations are used, care is required to ensure the estimated transfer relationships correspond to the intended input-output mapping.

2.2.3 Cross-power and auto-power formulations

Measured FRFs are frequently computed using spectral densities:

the cross-spectrum between output and input, and
the auto-spectrum of the input.

In a common formulation for complex FRF estimation, the FRF can be expressed as a ratio involving the output–input cross-spectrum divided by the input auto-spectrum. This approach naturally incorporates measurement noise assumptions and enables direct computation of complex magnitude and phase from spectral estimates.

2.3 Magnitude and phase interpretation

FRFs are complex-valued, so they are decomposed into magnitude and phase to reveal system behavior in intuitive plots.

2.3.1 Bode magnitude and phase concepts

Bode-style plots depict magnitude (often in decibels or as a linear gain) versus frequency, alongside phase (in degrees or radians). Resonant features appear as peaks in magnitude, while phase often exhibits characteristic transitions near resonances and anti-resonances. The slope and curvature of magnitude reflect underlying poles and zeros.

2.3.2 Nyquist-style visualization (overview)

Nyquist-style plots represent the complex FRF as a curve in the complex plane, typically parameterized by frequency. The geometry of the curve can reveal resonant ordering, phase evolution, and the influence of zeros and poles. While less commonly used than Bode plots in basic maintenance workflows, Nyquist plots can be informative in stability and frequency-response verification.

3 FRF Properties and System Behavior

The usefulness of FRFs depends on properties of the underlying system and on the assumptions made during identification.

3.1 Linearity and superposition assumptions

FRFs are fundamentally grounded in linear system theory. When the system exhibits nonlinearity, the response may depend on excitation amplitude or waveform shape, and the FRF estimated from measured spectra can become excitation-dependent. Many nonlinearities manifest as distorted FRF shapes, frequency-dependent discrepancies, or inconsistent phase behavior across repeated tests.

3.2 Causality, stability, and realizability implications

For physically realizable systems, FRFs must be consistent with causality and stability. In frequency-domain language, these constraints influence how the magnitude and phase behave across frequency and limit which combinations of poles and zeros are permissible in stable models. In practice, model validation often checks whether a fitted FRF corresponds to a realizable dynamical system rather than an arbitrary complex curve.

3.3 Resonances, anti-resonances, and bandwidth

Resonances correspond to frequencies where the system stores and releases energy in a way that amplifies the response. In FRFs, they appear as magnitude peaks and phase changes that track the resonant dynamics. Anti-resonances (sometimes called notches) arise near frequencies associated with zeros, producing dips in magnitude and phase behavior that differs from resonance crossings.

Bandwidth is often identified using amplitude thresholds relative to a baseline gain, but more rigorous definitions relate to the range over which the system responds significantly or where uncertainty stays within acceptable bounds.

3.4 Non-minimum phase considerations (general)

Some systems exhibit non-minimum-phase behavior, characterized by phase patterns that do not correspond to the magnitude profile expected from stable minimum-phase models. In an FRF context, non-minimum-phase behavior is linked to the presence and placement of zeros in the complex plane. General inspection of phase lag relative to magnitude can suggest such behavior, though definitive classification typically requires system identification with modeling assumptions.

3.5 Damping effects on FRF shapes

Damping reduces the sharpness of resonant peaks and alters how quickly phase transitions occur around resonant frequencies. Increasing damping generally flattens magnitude peaks and broadens resonant features. In modal terms, damping determines the width of each resonance and can also affect the depth and placement of anti-resonances when zeros interact with modal dynamics.

4 Estimation from Measured Data

FRFs are often computed from experimental time series using signal processing techniques. The key steps involve designing excitations, estimating spectra reliably, and forming FRF estimates from spectral ratios.

4.1 Measurement setup and excitation strategies

4.1.1 Single-sine sweeps (practical considerations)

A single-sine sweep applies a sinusoid at a sequence of frequencies, measuring steady-state response at each point. This method can provide good signal quality at each frequency and clear phase referencing, but it may be time-consuming and can struggle with systems that drift during long tests. It is also sensitive to synchronization between reference and measurement channels.

4.1.2 Chirp excitations and broadband tests

Chirp signals vary frequency over time, allowing broadband data acquisition in a single run. They can improve efficiency and reduce sensitivity to slow drifts. Estimation accuracy depends on signal bandwidth coverage and the ability to ensure that the FRF remains approximately valid during the sweep duration.

4.1.3 Random noise excitation and averaging

Random excitation methods use broadband inputs with spectral content across the frequency range of interest. By repeating experiments and averaging, estimators can reduce variance. These approaches are widely used because they can provide high-quality FRFs with manageable test durations, provided the random process is sufficiently rich and excitation is well measured.

4.2 Spectral estimation methods

4.2.1 FFT-based frequency-domain estimation

Frequency-domain estimators frequently rely on the fast Fourier transform (FFT). Time records are segmented, transformed, and converted into spectral estimates. The FRF is then computed at discrete frequency bins, corresponding to those FFT outputs.

4.2.2 Windowing and leakage control

Finite record lengths lead to spectral leakage if a signal does not align with an integer number of periods within each segment. Windowing mitigates this leakage by tapering the record, at the cost of reduced resolution or altered spectral characteristics. Choice of window depends on the intended trade-off between leakage suppression and frequency selectivity.

4.2.3 Averaging, coherence, and variance reduction

Averaging across segments reduces estimator variance. However, averaging does not guarantee correctness if excitation is insufficient at certain frequencies or if noise dominates. Coherence metrics help distinguish frequencies where the output is consistently linearly related to the input versus frequencies dominated by noise or unrelated dynamics.

4.3 Computing FRFs from spectra

4.3.1 FRF via cross-spectrum/auto-spectrum

A common complex FRF estimator uses the relationship between cross-spectrum \(G_{yu}(\omega)\) (output–input) and auto-spectrum \(G_{uu}(\omega)\) (input–input). In the idealized case, their ratio yields the complex FRF: \[ \hat{H}(\omega) = \frac{G_{yu}(\omega)}{G_{uu}(\omega)}. \] This approach accounts for noise in a way that is consistent with linear stochastic assumptions.

4.3.2 Complex FRF estimation (magnitude + phase)

Once the complex FRF is estimated, magnitude and phase follow directly from the complex value. Magnitude is computed as the modulus, and phase as the argument. For frequency response inspection, the phase is typically unwrapped so that continuous trends across frequency are visible despite discontinuities at \(\pm \pi\).

4.3.3 Handling scaling and units

Cross-spectral and auto-spectral calculations are affected by the physical scaling of measured signals and any sensor transfer functions. Proper calibration ensures that the FRF corresponds to the intended physical relationship, including the correct gain scaling between actuator input units and measured output units.

4.4 Coherence and quality metrics

4.4.1 Interpreting coherence values

Coherence is a measure of the linear correlation between input and output at each frequency, based on spectral estimates. Values near unity suggest that the output is strongly and consistently driven by the input at that frequency, while low coherence indicates poor linear relationship due to noise, unmodeled dynamics, or insufficient excitation.

4.4.2 Error sources and model mismatch

Even with high coherence, discrepancies can occur due to assumptions violations such as weak nonlinearity, time variance, incorrect sensor placement, or unaccounted delays in reference signals. FRF comparisons to models can reveal whether mismatches are localized to specific frequency regions or consistent across the band.

4.4.3 Outlier rejection and preprocessing

Preprocessing steps—such as detrending, removing transients, and filtering—can improve the stationarity of the data used for spectral estimation. Outlier rejection may be necessary when occasional measurement disturbances contaminate the spectral averages. Robust preprocessing helps ensure that the estimated FRF reflects the underlying system rather than transient artifacts.

5 FRF-Based System Identification

System identification uses FRFs to infer model structure and parameters. The goal is not only to reproduce the measured response but also to produce a model that behaves correctly outside the measured points.

5.1 Parameter fitting approaches

5.1.1 Rational function and transfer-function fitting (overview)

A widely used strategy fits a rational function in frequency that approximates the measured FRF. Rational models reflect the pole–zero structure of many physical systems. The fitting process chooses parameters to minimize an error metric between measured and modeled FRFs in magnitude and phase, often with weighting to emphasize frequencies with reliable coherence.

Physical systems often admit modal descriptions, where dynamics are represented by a set of modes with associated natural frequencies, damping ratios, and mode shapes. From FRFs, modal parameters can be extracted by identifying peaks, fitting local regions around resonances, or using global complex fitting methods that estimate modal contributions directly. Modal extraction supports interpretable models for vibration analysis and structural diagnostics.

5.2 Validation and comparison to time-domain models

Validation typically includes checking that the identified model reproduces the measured frequency response within uncertainty and that time-domain simulations driven by the same inputs match observed outputs. Because FRFs are derived under linear assumptions, time-domain validation helps detect issues such as unmodeled delays, nonlinearities, or incorrect input interpretation.

5.3 Regularization and noise-robust fitting (general)

Noise in FRF estimates can destabilize fitting procedures, especially when phase data is noisy or coherence is low. Regularization techniques constrain parameter growth or promote smoothness and can improve generalization. Weighting by coherence or confidence bands is often used to reduce the influence of unreliable frequency points.

6 FRFs in Control and Electrical Engineering Applications

FRFs are used across engineering disciplines, particularly where frequency-dependent behavior determines performance.

In control engineering, FRFs help interpret loop behavior by examining how gain and phase evolve with frequency. While stability analysis depends on system structure and feedback configuration, frequency-response methods provide insight into margins, resonance interactions, and the risk of undesirable oscillations when specific phase relationships occur.

6.2 FRFs for impedance and admittance analogs (general)

In electrical engineering, frequency-domain ratios such as impedance and admittance represent how voltage and current relate across frequency. These quantities behave similarly to FRFs in that they encode both amplitude and phase information. In electromechanical systems, FRF-style relationships also connect mechanical motion to electrical signals via transduction effects.

6.3 Frequency-domain design and verification (overview)

Design workflows often use frequency responses to verify that a system meets requirements such as attenuation, bandwidth, or transient sensitivity. FRFs can serve as design targets and as verification evidence, especially when measurements confirm whether implemented components match modeled dynamics.

7 Practical Implementation Details

Implementing FRF estimation and interpretation requires careful handling of sampling, sensor behavior, and phase alignment.

7.1 Sampling, record length, and frequency resolution

The sampling rate sets the maximum representable frequency, while record length determines frequency resolution in FFT-based estimation. Insufficient record length leads to coarse frequency bins and imprecise resonance localization. Choosing appropriate segment duration and overlap affects estimator variance and computational load.

7.2 Anti-aliasing and sensor dynamics (overview)

Anti-aliasing filters prevent high-frequency content from folding into the measurement band. Additionally, sensors and actuators may have their own dynamics, so the measured FRF may include sensor transfer effects unless explicitly compensated. Awareness of bandwidth limits and response time is essential for accurate modeling.

7.3 Phase unwrapping and reference alignment

Phase is inherently wrapped modulo \(2\pi\). Phase unwrapping reconstructs a continuous curve across frequency by adding or subtracting multiples of \(2\pi\) to remove jumps. Reference alignment ensures that the phase corresponds to the chosen input-output relationship and accounts for any time delay between channels.

7.4 Unit consistency and calibration

Because FRFs mix signals with physical units, calibration determines whether the resulting FRF has interpretable dimensions (e.g., displacement per unit force). Consistent unit handling across acquisition hardware, sensor scaling factors, and any applied normalization is required for meaningful magnitude comparisons.

8 Visualization and Reporting

Clear visualization supports interpretation, comparison, and decision-making based on FRFs.

8.1 Plotting conventions (Bode, Nyquist, polar)

Bode plots are common for presenting magnitude and phase separately. Polar plots show complex FRF values in angle–radius form, while Nyquist plots display the trajectory of the complex response across frequency. Consistent axes labeling, frequency units, and phase conventions are important for avoiding misinterpretation.

8.2 Frequency grid selection

The frequency grid determines which points are shown and affects apparent smoothness and resonance localization. A denser grid can improve visualization but depends on actual estimator resolution from FFT parameters. Interpolation may be used for presentation, but interpretation should remain tied to the underlying frequency-bin data.

8.3 Uncertainty reporting (confidence bands, repeatability)

FRF estimates depend on finite data and noise. Reporting uncertainty can include confidence intervals derived from repeated measurements, variability across segments, or statistical bounds from spectral estimation. Transparency about which uncertainties are considered helps readers judge reliability across the frequency band.

9 Common Pitfalls and Best Practices

Most FRF issues stem from mismatched assumptions, measurement limitations, or estimation artifacts.

9.1 Excitation coverage and SNR management

If the input does not sufficiently excite a frequency region, the FRF estimate there may be dominated by noise. Best practices include ensuring broadband excitation (or adequate frequency sweep density), monitoring signal-to-noise ratio across the band, and selecting analysis windows that avoid transient contamination.

9.2 Nonlinearities and their signatures in FRFs

Nonlinear behavior can appear as amplitude-dependent FRF changes, additional harmonics, or inconsistent phase trends. In practice, one can detect nonlinear signatures by repeating tests at different excitation levels or comparing FRFs computed from subsets of data. Significant deviations indicate that a linear FRF may not fully represent the system.

9.3 Aliasing, time synchronization, and detrending

Aliasing arises when sampling and anti-alias filtering are inadequate. Time synchronization errors between input and output channels can distort phase. Detrending and removal of DC offsets reduce spurious spectral content, improving estimator stability.

9.4 Coherence failures and mitigation strategies

Low coherence can indicate insufficient excitation, strong measurement noise, or uncorrelated disturbances. Mitigation strategies include improving sensor quality, increasing excitation amplitude within linear limits, using better averaging, refining preprocessing steps, and revisiting reference channel measurement.

10 Summary and Further Reading

FRFs provide a compact frequency-domain description of linear system behavior, enabling efficient characterization, identification, and validation across mechanical and electrical domains. They integrate magnitude and phase information, support model fitting, and allow quality assessment through spectral coherence and uncertainty measures.

10.1 Key takeaways

An FRF is a complex frequency-dependent relationship that captures both gain and phase.
For LTI systems, FRFs align with transfer functions; for time-varying systems they become approximate or local.
Reliable FRF estimation depends on excitation design, spectral estimation choices, and calibration.
Coherence and uncertainty metrics help determine which frequency regions are trustworthy.
FRF-based identification can yield physically interpretable modal parameters or rational transfer models, which should be validated against time-domain behavior.

10.2 Suggested references and standards (general)

Further study typically covers Fourier analysis and spectral estimation, system identification methods, and engineering practice guides for frequency-domain testing. References may include textbooks on vibration testing and signal processing, as well as general standards describing measurement uncertainty reporting, frequency response testing, and related instrumentation practices.