Fractional Delay Filter

Standard digital delays shift signals by integer multiples of the sampling period T_s. A fractional delay filter synthesizes a delay τ = (N + α)T_s, where N is an integer and α ∈ [0, 1) is the fractional component. This is achieved by approximating an ideal linear-phase all-pass filter with frequency response H(e^{jω}) = e^{-jωτ}, which corresponds to a shifted sinc impulse response in the time domain. The filter reconstructs inter-sample values that were never physically sampled, effectively acting as a continuous-time interpolator.

The Farrow structure is the dominant hardware-efficient architecture for variable fractional delay filters. It decomposes the filter into M+1 parallel fixed FIR sub-filters C_m(z) whose outputs are weighted by powers of the fractional delay parameter α and summed. This allows the delay to be adjusted continuously by changing a single parameter without reloading filter coefficients, making it ideal for real-time tracking of time-varying loop delays in adaptive DPD systems.

Lagrange interpolation is a maximally flat approximation of the ideal fractional delay at DC. The filter coefficients are computed directly as polynomials in α, making them simple to generate on-the-fly. Key properties include:

• Maximally flat magnitude response at ω = 0
• Exact delay at DC, with increasing error near Nyquist
• Coefficients derived from the closed-form Lagrange polynomial basis
• Odd-order filters (N=3,5,7) are preferred for better wideband performance Lagrange filters are widely used in DPD time alignment due to their computational simplicity and adequate performance for moderate fractional bandwidths.

For wideband signals approaching the Nyquist frequency, windowed sinc and least-squares designs offer superior performance. A windowed sinc filter truncates and tapers the ideal infinite impulse response using windows like Kaiser or Chebyshev to control passband ripple and stopband attenuation. Least-squares designs minimize the integrated squared error between the achieved and desired frequency response over a specified bandwidth. These methods trade increased filter length for flatter in-band delay response and are essential for linearizing wideband 5G signals where Lagrange filters exhibit unacceptable delay dispersion.

All practical fractional delay filters deviate from the ideal all-pass response, introducing magnitude distortion and delay dispersion that vary with frequency and the fractional delay value α. The achievable bandwidth is limited by the filter order N and the design method. Key performance metrics include:

• Passband ripple: deviation from unity gain
• Delay flatness: variation of group delay across frequency
• Nyquist edge performance: worst-case error occurs near f_s/2 In DPD applications, magnitude distortion in the delay filter directly corrupts the error signal used for coefficient estimation, making filter design a critical step in system calibration.

In a closed-loop DPD system, the transmitted reference signal and the feedback receiver signal must be aligned to sub-sample accuracy before computing the error signal e(n) = y_{desired}(n) - y_{observed}(n). A misalignment of even a fraction of a sample introduces decorrelation that degrades the coefficient estimation, increasing residual EVM. The fractional delay filter is placed in the reference path to compensate for the non-integer component of the loop delay after integer-sample coarse alignment via cross-correlation. Adaptive algorithms can dynamically tune α to track thermal drift in analog components.

The process of precisely synchronizing the transmitted reference signal with the observed feedback signal in the digital domain. Fractional delay filters are the primary tool for achieving the sub-sample alignment accuracy required for meaningful error computation. Without precise time alignment, the adaptive algorithm cannot correctly correlate the input and output, leading to coefficient divergence or poor linearization performance.

The total propagation latency through the transmission chain and feedback observation path, including:

• Analog group delay through the PA and coupler
• PCB trace propagation delays
• ADC/DAC pipeline latency
• Digital interface buffering

This delay must be accurately estimated and compensated before fractional delay filtering can align signals for coefficient estimation.

An efficient hardware implementation architecture for variable fractional delay filters that uses a bank of fixed FIR sub-filters and polynomial interpolation. The Farrow structure allows the fractional delay value to be changed dynamically by adjusting a single parameter without recomputing filter coefficients, making it ideal for real-time tracking of time-varying loop delays in adaptive DPD systems.

A classical polynomial interpolation method widely used to design fractional delay filters. For an Nth-order Lagrange interpolator, the filter coefficients are computed directly from the desired fractional delay value. Key characteristics:

• Maximally flat magnitude response at DC
• Excellent low-frequency phase accuracy
• Increasing magnitude distortion near the Nyquist frequency for higher orders

The instantaneous difference between the desired linear output and the actual observed PA output. This signal drives the adaptive coefficient update loop. Any residual timing misalignment between the reference and feedback paths directly contaminates the error signal with uncorrelated noise, degrading the convergence rate and steady-state performance of algorithms like LMS or RLS.

A dedicated observation receiver chain that down-converts and digitizes a coupled sample of the PA output. The feedback receiver's anti-aliasing filter and ADC sampling clock must be carefully designed to preserve the signal fidelity required for sub-sample alignment. Any phase non-linearity in the feedback path introduces systematic errors that fractional delay filters cannot correct.