How HARQ Soft Combining Actually Works — Chase vs Incremental Redundancy

"HARQ uses soft combining" — every textbook says it. But what does that actually mean at the bit level? When does the receiver use Chase vs IR? And why does it matter to a working RF engineer?

The Soft Buffer

When a UE receives a PDSCH and fails to decode (CRC mismatch), it doesn't throw away the data. It keeps the soft information — the per-bit log-likelihood ratios from the LDPC decoder — in the soft buffer keyed by HARQ Process ID.

For each bit in the codeword, the receiver computes an LLR: positive values lean toward 0, negative toward 1, magnitude = confidence. These LLRs are what the LDPC belief-propagation decoder iterates on. When decoding fails, the LLRs are kept; when retransmission arrives, they get combined with the new LLRs.

Chase Combining

The simplest form: gNB retransmits the exact same coded bits (same RV). Receiver simply averages the LLRs across the two transmissions (technically, sums them — averaging is equivalent in log-domain).

Why averaging works: each transmission is hit by independent noise, so the noise term in the LLR has zero mean across transmissions. Averaging reduces noise variance, increasing the effective SNR.

Result: 3 dB SNR gain after one retransmission, ~5 dB after two.

Chase combining is simple but inefficient. The same parity bits are retransmitted; no new information is added.

Incremental Redundancy (IR)

The smarter form: gNB retransmits a different RV — different starting position in the rate-matched bit sequence. The new RV reveals different parity bits that the original transmission didn't include.

Now the receiver has a richer codeword: original bits + new parity bits. The effective code rate drops (more parity, fewer info bits per coded bit), making decoding easier.

Result: Up to 6 dB SNR gain after one retransmission with the optimal RV — far better than Chase's 3 dB.

RV Selection

Per TS 38.212, 4 RVs are defined: 0, 1, 2, 3. RV 0 starts at the systematic bits (best for first transmission). RV 2 starts at parity (best for retransmission). RV 1 and 3 are between.

Standard pattern: gNB sends RV 0, 2, 3, 1 across HARQ retransmissions. This sequence maximizes IR gain at each step.

For very high code rates (e.g., 0.93 mother rate), the RVs may overlap — you can't add infinite parity bits. In this regime, IR degenerates into Chase: same coded bits retransmitted.

The Soft Buffer Size Constraint

UE soft buffer is finite — typically 5-20 megabits per HARQ Process. If the codeword is very long (e.g., a 10 Gbps single-layer transmission would produce ~70 kbits per code-block), the soft buffer can overflow. This is "limited buffer rate matching" (TS 38.212 §5.4.2).

When the soft buffer is too small, only a subset of LLRs are stored. The retransmission then only adds soft info for those bits; the others are decoded fresh each time. Effective IR gain is reduced.

This is why UE category specifies a minimum soft buffer size. UEs that claim 1024-QAM at high MIMO need correspondingly larger soft buffers.

When Chase vs IR is Used in Practice

Most modern gNB schedulers default to IR — it gives more gain. Chase combining shows up in:

• Very high code rates where RVs overlap
• Buffer-limited UEs
• Spec-compliance fallback when scheduler can't determine optimal RV

What This Means in Logs

When you see HARQ retransmissions in a UE log:

• If RV stays 0 across retransmissions → Chase combining (or scheduler bug)
• If RV cycles 0 → 2 → 3 → 1 → IR with optimal sequence
• If 2nd retransmission fails too → likely propagation issue (block fading) or extreme interference; soft combining isn't fixing this

Code-Block-Group HARQ (Rel-16)

Rel-16 introduced CBG-level HARQ: instead of acknowledging the whole transport block, the UE acknowledges per code-block-group. The gNB retransmits only the failing CBGs, saving spectrum.

For very large TBs (e.g., 4 layers × 256-QAM × 100 MHz), this is a 5-10× efficiency gain on HARQ retransmissions. Highly recommended for high-throughput UEs.

The full course covers HARQ end-to-end in Lesson 2-1-9 with Tanner-graph soft-combining visualizations. $29 lifetime.