Industry Insights

Why block processing beats per-symbol sync in LPWAN PHYs

Per-symbol synchronisation, not the radio, caps LPWAN sensitivity. What a block transmission framework changes, and what the simulations show.

A LoRa receiver at −20 dB SNR is not limited by its radio. It is limited by its synchroniser. Every demodulation decision in a conventional LPWAN receiver waits on a symbol-timing estimate, and below about −20 dB that estimate is the first thing to fall apart: the tracking loop cannot tell signal from noise, so it loses lock, and the energy in the rest of the packet is wasted.

That observation is the starting point of a paper published by our own R&D team with Sun Yat-sen University: Unified Block Signal Processing Framework for LPWANs: Sequence Index Modulation Spreading (arXiv:2511.20364). This article explains what the framework proposes in engineering terms, and what it changes for anyone specifying a sub-GHz link.

Processing gain against spreading factor: 15 dB at SF5 rising to 36 dB at SF12

The per-symbol problem

Classical receivers recover carrier phase and symbol timing per symbol: a phase-locked loop or Costas loop for the carrier, Gardner or Mueller–Müller detectors for timing. These work well in Wi-Fi, LTE and 5G NR because the SNR is high and pilots are cheap. In LPWAN the assumptions invert:

  • SNR can sit below −20 dB, where a closed-loop tracker sees mostly noise and either drifts or drops lock.
  • Power budgets forbid long pilot sequences and high-rate tracking loops.
  • Traffic is sporadic, so every packet pays the acquisition cost again, from cold.
  • Per-symbol detection cannot accumulate energy coherently across symbols, which is exactly the capability low-SNR reception needs.

The paper puts a number on the gap: coherent accumulation across a LoRa-style spread block can deliver 20–30 dB of SNR improvement relative to per-symbol detection, with roughly 3 dB per doubling of the spreading factor — up to about 21 dB of coherent despreading gain at SF7.

What block transmission changes

Instead of modulating symbol by symbol, the transmitter takes a block of N_b information bits at a time and maps it to one codeword of N_SF = 2^SF chips, then sends the whole codeword as a unit. The framework parameterises that mapping with three objects:

Component

What it is

What it buys

Signal block vector

The N_b-bit block mapped to a codeword index

One decision per block instead of one per symbol

Intra-block structure generator Λ(k, n)

A function of subcarrier index k and time index n that shapes the waveform inside the block

One formula covers FSK (linear), chirp (quadratic) and arbitrary spread patterns

Signal basis matrix W

The K × N_SF matrix of fundamental waveforms spanning the codebook

Codebook construction becomes a linear-algebra problem, not an ad-hoc waveform design

The general chip value is s(k,n) = A(k)·exp(j·2π/K·f₀·Λ(k,n)). Set Λ linearly and you get frequency-shift keying; set Λ = ϑn² and you get chirp spread spectrum. That is the useful part of the framework: LoRa's CSS and Sigfox's FSK are not separate designs, they are two points in the same parameter space, which means the space between them can be explored.

The proposed generalisation, sequence index modulation spreading (SIMS), replaces the fixed waveform with cyclically shifted versions of a root sequence chosen by the codeword index. Each block therefore carries log₂(N_SF) bits in the shift index while keeping a constant-envelope waveform — the same bits per block as CSS, but with codebooks that can be redesigned per deployment by changing the root sequence.

The gain, and where it comes from

Spreading N_b bits over N_SF chips yields approximately 10·log₁₀(N_SF) dB of processing gain:

Spreading factor

N_SF (chips per block)

Processing gain

SF5

32

15.1 dB

SF6

64

18.1 dB

SF7

128

21.1 dB

SF8

256

24.1 dB

SF9

512

27.1 dB

SF10

1,024

30.1 dB

SF11

2,048

33.1 dB

SF12

4,096

36.1 dB

Nothing in that table is new to anyone who has sized a LoRa link. What is new is the claim that the rest of the receive chain can be built to exploit it without a symbol-tracking loop: block synchronisation on cyclic pilots, and one block-correlation decision instead of thousands of symbol decisions.

Multi-user separation without scheduling

The same structure solves a second problem. In a codebook built from cyclic shifts of a root sequence, any two codewords are quasi-orthogonal rather than strictly orthogonal, and the paper proves the cross-correlation between them concentrates: the probability that it exceeds a threshold falls exponentially in the product K·N_SF (the number of subcarriers times the spreading factor). For constant-modulus modulation such as PSK the bound simplifies to a single exponential in K·N_SF.

Practically: assign each device a codebook derived from its equipment serial number, and devices can transmit asynchronously — no scheduling, no time-slot map — while the receiver separates them by correlation. The interference is not eliminated, it is bounded with high probability, and the bound is wide enough to plan with.

That is the second article in this pair; the details, including the Welch-bound coherence figures and what happens at 1,000 users, are here.

The receiver: correlation as an FFT

A brute-force matched filter compares the received block against every codeword: O(N_SF²) operations per block. Because the codebook of a cyclically shifted root sequence is a circulant matrix, the correlation is a convolution, and a convolution is a pointwise product in the frequency domain:

z = IFFT( FFT(q)* ⊙ FFT(y) )

That drops the cost to O(N_SF log N_SF), and in the multi-user case to O(K·N_SF log N_SF) for all M users — the M codebooks are processed in parallel, with no successive interference cancellation stage.

Direct codebook correlation versus FFT-based matched filtering, log scale

At N_SF = 4,096 the difference is 16.8 million operations against 49 thousand, a factor of 341. That is the difference between a receiver that needs an FPGA and one that fits a modest MCU — which is the reason this kind of architecture is interesting for battery-powered endpoints rather than only for gateways.

How the synchronisers compare

The paper is blunt about the alternatives:

Scheme

How it synchronises

Consequence

LoRa

Time-domain correlation between received and reference chirps

Strong noise immunity, but serial processing and tied to chirp waveforms

Sigfox

Coarse time-of-arrival on ultra-narrowband signals, centralised

Cheap endpoints, long detection windows, weaker under mobility and fading

NB-IoT

OFDM PSS/SSS correlation

Robust, but power and complexity cost is out of reach for coin-cell nodes

SIMS

Block-level correlation on cyclic pilots, FFT-parallel, blind

Decentralised, not tied to one waveform family; complexity O(N_SF log N_SF)

"Blind" here means the receiver does not need an explicit preamble pattern matched to a fixed waveform — the block structure itself carries the alignment information, which is what allows the framework to accept different root sequences without redesigning the sync stage.

What the simulations show

All results are baseband Monte Carlo with 2 × 10⁶ trials, equal energy per bit across schemes:

  • AWGN: at small spreading factors SIMS is slightly worse than CSS and M-FSK, because quasi-orthogonal sequences are not strictly orthogonal. As SF grows the curves converge, and each SF step buys roughly 3 dB.
  • Frequency-selective channel, no small-scale fading: at BER 10⁻⁴, CSS beats M-FSK by about 2 dB, and SIMS adds another 3.5 dB over CSS at the same SF.
  • Frequency-selective with small-scale fading: SIMS gains about 12 dB over CSS at BER 10⁻⁴ and SF7.
  • Multi-user, asynchronous: per-user BER approaches the single-user bound as SF grows, and at SF10–11 the single-user and multi-user curves are nearly identical.

The frequency-selective result is the one worth remembering, because it is the case that matters in a meter room. Orthogonal codes lose their orthogonality after multipath convolution; quasi-orthogonal sequences are designed so that the cross-correlation stays low under time shifts, which is exactly what a delayed multipath copy presents.

What this means for a real link budget

Three practical consequences if you are specifying a sub-GHz product:

  1. Sensitivity is a system property, not a radio property. A module with −137 dBm sensitivity is worth nothing if the synchroniser needs −10 dB SNR to lock. Ask suppliers for packet error rate at the SNR you will actually operate at, not for a sensitivity number measured with a perfect timing reference.
  2. SF is a throughput dial with a 3 dB step. Every step up buys 3 dB and costs half the data rate. Combined with duty-cycle limits, that is the whole capacity calculation — as worked through in our LoRa module airtime breakdown.
  3. Antenna and height still dominate. 3 dB of processing gain is one SF step; 3 dB of cable loss is a cheap connector done badly. Fix the passive side first.

Where block processing does not help

It does not turn a 76.8 kbps channel into a video link, it does not remove duty-cycle limits, and it does not help a link that is already limited by transmit power rather than by receiver SNR. It is a way to spend receiver computation to buy sensitivity and multi-user robustness — useful precisely because receiver computation is the cheapest resource in an LPWAN endpoint.

For the technology background behind our own sub-GHz products, see the band and device overview and the product catalogue. The terminal line (TP5500, TP5600) and the modules (TP1107, TP5301) use the same CPFSK-with-spreading lineage described in the team's earlier IEEE IoT-J paper, which this framework generalises.

More in Industry Insights