A radio that sends very short packets, a few hundred channel uses carrying a few dozen bits, faces an awkward problem. At these lengths, most errors come from failing to notice that a packet has arrived, not from failing to decode it. The usual fix is a known preamble, and the preamble eats much of an already tiny budget. Neuroscience has studied a phenomenon that looks like a detector built for exactly this situation. In cortex, and in chaotic recurrent neural networks, ongoing activity fluctuates wildly from one trial to the next. When a stimulus arrives, those fluctuations collapse, an effect called variability quenching. We propose building a receiver from a recurrent network that sits in its chaotic, fluctuating state while the channel carries only noise. The arrival of a packet that contains a weak, structured bias pins the network's trajectory, and that collapse is the signal that a packet has started.
The proposal rests on two pieces. The first is a transmitter-side observation that costs nothing. Information theory says the detection-friendly packet should have its input distribution skewed toward one symbol. The direction of that skew can be rotated onto any known ±1 sign pattern without changing the fundamental limits, so it can be aimed at the frequency where the network quenches most easily. The second is a receiver-side test that can be measured on a single received stream: run two copies of the same chaotic network from different starting states on that stream, and declare a packet when the two copies converge. The open question is whether this does anything that a narrowband energy detector or a learned non-chaotic detector cannot do. The experiments are built to answer it, including the possibility that it does not.
Core Claim
On Joint Detection and Decoding in Short-Packet Communications gives information-theoretic bounds for a binary-input additive white Gaussian noise (AWGN) channel in which the receiver must both detect and decode. It shows that below roughly 190 channel uses at 0 dB signal-to-noise ratio, detection is the bottleneck. It also shows that the best input skews toward one symbol, with the probability of that symbol $p \to 1$ as the blocklength shrinks. The paper does not propose a receiver that achieves these bounds, and it assumes the packet boundaries are synchronous.
Separately, Stimulus-dependent suppression of chaos in recurrent neural networks and its companion Stimulus-dependent suppression of intrinsic variability in recurrent neural networks show a sharp transition in chaotic rate networks. Once a periodic input exceeds a critical amplitude $I_c$, chaos disappears. $I_c$ is smallest at a resonant frequency that is not zero, even though the network's spontaneous spectrum peaks at zero. Those papers drive the network only with clean sinusoids. These two results are in tension, because the skew that best aids detection is a constant (DC) offset, which is not where quenching is cheapest.
Our claim has two parts. First, the tension disappears at no information-theoretic cost. Write the transmitted symbols as $x_i = s_i b_i$, where $b_i \in \{\pm1\}$ are the skewed coded symbols of the bounds above and $s$ is a fixed, known ±1 sequence whose dominant frequency is the receiver network's resonance. Multiplying by a diagonal ±1 matrix is an orthogonal map that preserves the binary alphabet and leaves white Gaussian noise unchanged, so every detection and decoding bound for $b$ holds exactly for $x$. Second, a receiver built from a chaotic network can turn this pattern-shaped bias into a sharp, phase-insensitive detection event, and that event can be measured within a single transmission. Let the network follow
$$\tau \dot{h}(t) = -h(t) + g J \phi(h(t)) + w\, y(t),$$
where $h$ is the $N$-dimensional state, $J$ is a random coupling matrix scaled by gain $g$ (chaotic for $g>1$), $\phi$ is a saturating nonlinearity, $w$ is a fixed input vector, and $y(t)$ is the received stream. Two replicas $h^{(1)}, h^{(2)}$ receive the same $y$ but start from different initial states. The detection statistic is their normalized squared distance $D(t) = \|h^{(1)}(t)-h^{(2)}(t)\|^2/N$. This is literally an across-trial variance computed inside one realization. Noise alone should leave $D$ large. A packet whose bias sits at resonance should drive $D$ toward zero. This is the same input-slaving of transients that Neural Heterogeneity Enhances Reliable Neural information Processing links to reliable processing.
Research Program
The work has three strands. The first is analysis. Rajan, Abbott and Sompolinsky computed their phase boundary with dynamic mean-field theory, a method that replaces the network with self-consistent statistics of a single unit. Their analysis assumed each neuron received an independent random phase. Our input is a single scalar stream spread through $w$, which is a different and more coherent drive. We will rederive, or at least numerically map, $I_c$ as a function of frequency for a drive made of a sign pattern plus white noise. The key quantity is the gap between the critical amplitude for a resonant pattern and for white noise of equal power. Without that gap, quenching is just energy detection.
The second strand turns this into detection error rates. We will measure false-alarm and misdetection probabilities for thresholding $D$, and the detection delay in channel uses. That delay matters because Neural Variability in Premotor Cortex Provides a Signature of Motor Preparation and Task-evoked activity quenches neural correlations and variability across cortical areas show that quenching takes finite time, governed in the latter's models by a characteristic timescale. Every symbol spent settling is a symbol lost from a blocklength of a few hundred. We will tune $\tau$ and $g$ so that settling fits within the bias-bearing part of the packet.
The third strand is a full receiver. After detection, the quenched state of one replica, read by a linear or small learned readout, provides soft information to a standard short-code decoder. The comparison to the information-theoretic bounds then covers the whole chain. We do not claim the recurrent network should decode better than the strong codes in Deep Learning Based Near-Orthogonal Superposition Code for Short Message Transmission, which uses a deep-network encoder with a MAP and CRC-aided tree-search decoder. We hold the payload code fixed and vary only the detection front end.
Baselines
The naive reference is a preamble followed by a matched-filter threshold, with the preamble length optimized for each blocklength. We also report a genie receiver that knows the arrival time exactly, which separates detection loss from decoding loss. On Joint Detection and Decoding in Short-Packet Communications shows that preamble schemes reach zero rate at short blocklengths where joint detection still achieves positive rate. Beating this baseline is therefore expected, and beating it alone would prove little.
The strongest competitor in pure AWGN is a coherent correlator on the known pattern $s$ combined with the same skewed coding. This is essentially the optimal composite test the bounds point toward. Under unknown arrival time and carrier frequency offset, the correlator becomes a generalized likelihood-ratio test searched over offsets. We also run a narrowband power detector at the frequency of $s$, which is cheap and insensitive to phase. That narrowband detector is the direct mechanistic rival of quenching. All of these get the same received streams, the same false-alarm targets, and thresholds tuned on the same held-out noise-only data. The search-based test is reported both at full offset resolution and at an operation count matched to the network.
For multipath fading, the strongest published blind method is BMOCZ-ACPC from MOCZ for Blind Short-Packet Communication: Some Practical Aspects. It reports a block error rate of $10^{-1}$ at about 6 dB SNR with 4 receive antennas on the Quadriga Berlin urban-macro non-line-of-sight channel. Matching or beating that point at equal antennas and payload counts as a win in that regime. For the joint detection bounds themselves there is no reported numeric achievable-rate result on Lacuna for a concrete receiver, only the qualitative 190-symbol detection-limited regime. We will compute the bounds ourselves and report the gap to them.
The key ablation keeps everything identical but sets $g<1$, so the network is stable rather than chaotic. A second variant uses a trained gated recurrent unit detector of equal parameter count. If either matches the chaotic receiver, the quenching mechanism is decorative.
Why not just…?
Why not just use a preamble and a matched filter? At the blocklengths of interest, the cited joint-detection analysis shows preambles waste enough channel uses to drive the rate to zero where preamble-free schemes still work. Our first experiment reproduces this, but we treat it as a sanity floor rather than evidence for the proposal.
Why not just correlate against the bias pattern, or measure narrowband power at its frequency? This is the real threat. In synchronous AWGN, the coherent correlator should be at least as good as any nonlinear network, and we expect the chaotic receiver to lose there. Its possible advantage appears only with unknown arrival time and frequency offset. In that setting the correlator needs a search, and a narrowband power detector cannot tell the structured pattern from in-band noise or interference of equal power. The quenching receiver helps only if the measured critical-amplitude gap between the pattern and equal-power noise yields better false-alarm and misdetection curves than the narrowband detector at matched cost. If it does not, we will report that the mechanism reduces to filtered energy detection and keep only the transmitter-side result.
Why not just train a feedforward or attention-based detector? AttentionCode: Ultra-Reliable Feedback Codes for Short-Packet Communications reaches a block error rate of $10^{-7}$ at 0 dB in AWGN. It assumes frame synchronization, however, and its memory grows with the square of the window length when run on a stream. A sliding-window convolutional or recurrent detector trained end to end is the fair version of this alternative, and it is our gated-recurrent-unit baseline. If that detector equals the replica-collapse receiver at equal parameters, the trained model wins on simplicity and we will say so.
Proposed Experiments
The first experiment maps $I_c$ against frequency in $N=1000$ to $10{,}000$-unit networks under scalar sign-pattern drive with and without white noise. The decisive quantity is the critical-amplitude ratio between equal-power white noise and the pattern. A ratio near one ends the detection claim early.
The second experiment uses binary-input AWGN at blocklengths from 64 to 500 and 0 dB SNR, with arrival time drawn uniformly over a long noise stream. We compare the replica-collapse receiver with every baseline using false-alarm and misdetection curves, overall error at a fixed false-alarm rate, detection delay, and achievable rate against the computed joint-detection bounds. Each condition is run with and without the free rotation from DC skew to resonant-pattern skew, which isolates the transmitter-side contribution.
The third experiment adds carrier frequency offset and timing offset. The decisive check is whether the chaotic receiver's gain over the narrowband and search-based detectors grows with the size of the offset. That is the only regime where the mechanism has a principled reason to help.
The fourth experiment runs the full chain on the Quadriga urban-macro non-line-of-sight channel with 4 receive antennas, against the BMOCZ-ACPC reference point. The $g<1$ and gated-recurrent-unit ablations are reported in every experiment.
References
- On Joint Detection and Decoding in Short-Packet Communications
- Stimulus-dependent suppression of chaos in recurrent neural networks
- Stimulus-dependent suppression of intrinsic variability in recurrent neural networks
- Neural heterogeneity enhances reliable neural information processing: Local sensitivity and globally input-slaved transient dynamics
- Neural Variability in Premotor Cortex Provides a Signature of Motor Preparation
- Task-evoked activity quenches neural correlations and variability across cortical areas
- Deep Learning Based Near-Orthogonal Superposition Code for Short Message Transmission
- MOCZ for Blind Short-Packet Communication: Some Practical Aspects
- AttentionCode: Ultra-Reliable Feedback Codes for Short-Packet Communications