Probabilistic Shaping

Unlike standard uniform QAM where every constellation point has an equal probability of transmission, probabilistic shaping assigns a Maxwell-Boltzmann distribution to the symbols. Inner points, which require less energy, are transmitted with high probability, while outer, high-energy points are used sparingly. This creates a Gaussian-like distribution over the constellation, which is the optimal input distribution for the additive white Gaussian noise (AWGN) channel. The result is a direct reduction in average transmit power for a fixed data rate.

The primary benefit of probabilistic shaping is the shaping gain, which can reach up to 1.53 dB in the high signal-to-noise ratio (SNR) regime. This gain represents the reduction in required SNR to achieve a target error rate compared to uniform signaling. Critically, this gain is achieved without increasing the number of constellation points, meaning the peak-to-average power ratio (PAPR) and implementation complexity remain manageable. It directly translates to extended reach in optical fiber or higher throughput in wireless links.

To map uniformly distributed data bits to the desired non-uniform symbol sequence, a Constant Composition Distribution Matcher (CCDM) is employed. The CCDM is an invertible, fixed-length mapping algorithm that transforms a sequence of independent, uniformly distributed bits into a sequence of symbols with a specific empirical distribution. This process is fixed-to-fixed length, meaning a fixed number of input bits maps to a fixed number of output symbols, which is essential for maintaining a constant framing structure and data rate in practical transceivers.

A key practical advantage of probabilistic shaping over geometric shaping is its fine-grained rate adaptivity. By simply changing the target distribution's entropy—controlled by a single shaping parameter—the data rate can be adjusted in very small increments without altering the underlying constellation or forward error correction (FEC) code. This allows a system to dynamically adapt to changing channel conditions, maximizing throughput in real-time. The shaping rate is defined as the entropy of the symbol distribution in bits per symbol.

Probabilistic shaping is typically implemented using a reverse concatenation architecture. In this scheme, the distribution matcher is placed inside the FEC encoder loop, meaning the shaping operation occurs before FEC encoding at the transmitter. This is crucial because it ensures that the non-uniform symbol distribution is preserved after decoding. If FEC were applied first, the parity bits added by a systematic encoder would be uniformly distributed, destroying the shaping gain. Reverse concatenation decouples the shaping and coding gains.

The most common architecture for implementing probabilistic shaping is Probabilistic Amplitude Shaping (PAS). PAS operates by separating the constellation into amplitude rings and sign bits. The distribution matcher shapes only the amplitudes, while the sign bits remain uniformly distributed. This elegantly leverages the symmetry of the QAM constellation and allows the use of a standard, off-the-shelf binary FEC code to protect the uniformly distributed sign and parity bits, drastically simplifying the system design.

The complementary approach to probabilistic shaping that optimizes the physical locations of constellation points in the continuous IQ plane rather than their probability of occurrence. Unlike probabilistic shaping, which uses a regular QAM lattice, geometric shaping moves points to arbitrary positions to maximize mutual information for a specific channel model.

• Key difference: Alters point locations, not frequencies
• Result: Non-uniform, often circular or Gaussian-like constellations
• Combined use: Can be merged with probabilistic shaping for ultimate performance

The core algorithmic engine that transforms a sequence of uniformly distributed input bits into a sequence of constellation symbols with a target non-uniform probability distribution. A CCDM operates by maintaining a fixed composition—a specific number of occurrences of each symbol—within each output block.

• Function: Maps bits to shaped symbols with fixed empirical distribution
• Rate loss: Incurred due to finite block length; vanishes as block size increases
• Implementation: Typically realized via arithmetic coding or enumerative sphere shaping

The theoretical maximum rate at which information can be reliably transmitted over a communication channel, as defined by Claude Shannon in 1948. For the additive white Gaussian noise (AWGN) channel, capacity is achieved only when the transmitted signal has a Gaussian distribution.

• Gap to capacity: Uniform QAM operates 1.53 dB from the limit at high SNR
• Probabilistic shaping benefit: Recovers up to 1.2 dB of this gap
• Formula: C = B log₂(1 + SNR) for the continuous Gaussian input case

The specific non-uniform probability distribution applied to constellation points in probabilistic shaping. Named after the statistical mechanics distribution of particle energies, it assigns exponentially decreasing probabilities to points with higher energy, minimizing the average transmitted power for a given entropy rate.

• Inner points: Transmitted with high probability (low energy)
• Outer points: Transmitted with low probability (high energy)
• Parameter ν: Controls the shaping strength; ν=0 recovers uniform distribution

The architectural challenge of combining probabilistic shaping with FEC coding. In the probabilistic amplitude shaping (PAS) architecture, shaping operates on the amplitude bits while FEC encodes the sign bits and shaped amplitude bits, enabling seamless integration without altering the FEC code design.

• PAS architecture: Standardized approach for bit-interleaved coded modulation
• Sign bits: Carry uniform information and are FEC-protected
• Rate adaptation: Achieved by adjusting shaping distribution, not FEC code rate

The achievable information rate metric used to evaluate probabilistically shaped systems. GMI quantifies the number of bits per symbol that can be reliably transmitted when using a suboptimal bit-metric decoder, making it the practical benchmark for real systems rather than the idealized mutual information.

• Practical metric: Accounts for bit-wise decoding constraints
• Optimization target: Shaping distribution is designed to maximize GMI
• Comparison: Uniform QAM typically achieves lower GMI at the same SNR