Nonlinear Order

In differential and push-pull power amplifier architectures, odd-order nonlinearities (3rd, 5th, 7th) dominate the distortion spectrum. Even-order products tend to cancel in balanced topologies, making odd-order Volterra terms the primary focus for digital predistortion. Third-order intermodulation products fall closest to the carrier and are typically the most problematic for adjacent channel interference.

• 3rd order: Produces IM3 products near the carrier, primary source of spectral regrowth
• 5th order: Generates IM5 products, significant for wideband signals
• 7th order: Captures deep compression behavior in GaN amplifiers
• Even orders: Often negligible in differential designs but critical for single-ended stages

The number of Volterra coefficients grows combinatorially with nonlinear order. A model with order K and memory depth M requires approximately M^K coefficients in its full form. This exponential growth makes high-order full Volterra models computationally prohibitive, motivating simplified structures like memory polynomials that retain only diagonal kernel terms.

• 3rd order, M=3: ~27 coefficients (manageable)
• 5th order, M=5: ~3,125 coefficients (demanding)
• 7th order, M=7: ~823,543 coefficients (intractable without pruning)
• Sparse regression techniques like LASSO become essential above 5th order

Each nonlinear order contributes distinctively to AM-AM distortion (gain compression/expansion) and AM-PM distortion (phase shift vs. amplitude). Odd-order terms generate both in-band distortion and spectral regrowth, while the relative phase of Volterra kernel coefficients determines whether the amplifier exhibits gain compression or expansion at saturation.

• 3rd order: Typically causes gain compression near the 1 dB compression point
• 5th order: Can create gain expansion before deep compression in Doherty amplifiers
• AM-PM conversion: Phase distortion increases with order, critical for high-order QAM signals
• Complex baseband Volterra models capture both amplitude and phase nonlinearities simultaneously

Selecting the appropriate nonlinear order involves balancing model fidelity against numerical stability. The Akaike Information Criterion (AIC) provides a statistical framework for order selection by penalizing over-parameterization. In practice, the required order depends on the amplifier's operating point relative to its saturation power.

• Backed-off amplifiers: 3rd order often sufficient
• Near-saturation operation: 5th to 7th order required
• Deep compression: 9th order or higher may be necessary
• Cross-validation on held-out test data prevents overfitting to measurement noise
• Condition number of the regression matrix degrades rapidly with excessive order

The frequency spacing of intermodulation products is directly determined by nonlinear order. IM3 products appear at 2f1-f2 and 2f2-f1, falling immediately adjacent to the carrier frequencies. Higher-order products spread further into adjacent and alternate channels, making order selection critical for meeting ACLR specifications.

• IM3: Closest to carrier, primary ACLR contributor
• IM5: Wider spacing, affects alternate channel leakage
• IM7: Even wider, relevant for wideband 5G signals (100 MHz+)
• The required order scales with signal bandwidth relative to channel spacing

Tensor decomposition techniques like Canonical Polyadic Decomposition (CPD) exploit the inherent low-rank structure of Volterra kernels to dramatically reduce parameter count without sacrificing modeling accuracy. A K-th order kernel tensor can be factorized into a sum of rank-1 components, converting exponential scaling to linear scaling in the tensor rank.

• CP-Volterra: Reduces O(M^K) to O(R·K·M) where R is tensor rank
• Typical rank R: 2-5 for weakly nonlinear PAs
• Enables practical 7th-order models on FPGA hardware
• Dynamic deviation reduction is an alternative approach that separates static and dynamic nonlinearities

In differential and push-pull power amplifier architectures, odd-order nonlinearities (3rd, 5th, 7th) dominate the distortion landscape. Even-order products are suppressed by circuit symmetry, making odd-order Volterra kernels the primary targets for digital predistortion.

• 3rd-order: Causes spectral regrowth into immediate adjacent channels
• 5th-order: Extends distortion into alternate channels
• 7th-order and above: Becomes significant at deep saturation

The nonlinear mapping between input signal amplitude and output signal amplitude. As nonlinear order increases, the AM-AM curve deviates more sharply from linear gain, exhibiting gain compression at saturation.

• 1st-order: Ideal linear gain (slope = 1)
• 3rd-order: Introduces cubic compression (gain expansion or compression)
• 5th-order: Adds quintic curvature for modeling hard saturation knees

The nonlinear relationship between input amplitude and output phase shift. Higher nonlinear orders capture the phase rotation that occurs as the amplifier approaches compression, critical for modulation accuracy.

• Phase distortion degrades EVM and constellation integrity
• Odd-order terms model the amplitude-dependent phase shift
• Essential for correcting QAM and OFDM signals

When multi-tone signals pass through a nonlinear system, intermodulation products appear at sum and difference frequencies. The order of these products directly corresponds to the nonlinear order generating them.

• 3rd-order IMD: 2f₁ ± f₂, 2f₂ ± f₁ — falls in-band
• 5th-order IMD: 3f₁ ± 2f₂ — wider spectral spread
• Higher orders produce progressively weaker but wider-band interference

Choosing the correct nonlinear order involves balancing modeling accuracy against computational complexity. The Akaike Information Criterion (AIC) and cross-validation are standard tools for this tradeoff.

• Too low an order: residual nonlinearity and poor ACLR
• Too high an order: overfitting to noise, numerical instability
• LASSO regression can automatically prune unnecessary high-order terms

Nonlinear amplification causes a band-limited signal to spread into adjacent frequency channels. The extent of regrowth is directly proportional to the nonlinear order and signal PAPR.

• 3rd-order regrowth: 3× original bandwidth
• 5th-order regrowth: 5× original bandwidth
• Regulatory ACLR limits drive the required DPD linearization order