DMRS — Demodulation Reference Signals in 5G NR

1. Why Reference Signals Exist, and Why Each Channel Has Its Own

A radio channel rotates and scales everything that passes through it. The transmitter sends a QAM constellation point; the receiver sees that point multiplied by a complex number it does not know, and buried in noise. Demodulation is the process of dividing that unknown complex number back out, and to divide it out you have to know it. The only way to know it is to have sent something whose value the receiver already knew, observed what came out, and taken the ratio. That known signal is a reference signal, and demodulation reference signal is exactly what the name says: the pilot whose only job is to make demodulation of the channel it rides on possible.

Everything else about DMRS follows from three questions. How densely do you have to sample the channel in frequency, given that it varies across the band? How densely in time, given that it varies as the UE moves? And how do you separate the pilots of several spatial layers that arrive on the same resource elements at the same instant? Those three questions produce the configuration type, the additional positions, and the CDM groups respectively, and the answer to each costs resource elements the transport block would otherwise have had.

1.1 The LTE arrangement, and why NR could not keep it

LTE broadcast a cell-specific reference signal, CRS, on a fixed lattice of resource elements in every subframe, on every antenna port, whether or not anyone was being scheduled. Every UE in the cell used the same CRS. It was elegant: one pilot amortised across the entire cell, always present, always available for measurement as well as demodulation.

That arrangement depends on one assumption which NR breaks. CRS is only a valid channel estimate for the data if the data traverses the same channel as the CRS. In LTE that was nearly true, because the eNB transmitted CRS and PDSCH from the same wide sector beam. In NR the gNB forms a narrow beam per UE, chosen from a codebook and updated as the UE moves, and it may apply a different precoder to every layer of every grant. A cell-wide pilot transmitted on a wide beam tells a UE nothing whatsoever about the narrow beam its own PDSCH arrived on. The pilot must traverse the same beam as the data. That is the single most important conceptual point in this document, and it is why NR has no CRS at all.

The consequences are worth listing because they shape the rest of the specification:

• DMRS is only present when the channel is present. No transmission, no pilots. A UE cannot use DMRS to measure a cell it is not being scheduled in, which is why NR needs SSB and CSI-RS as separate, independently transmitted measurement references (§17).
• DMRS is precoded with the data. The UE estimates the effective channel -- physical channel times precoder -- and never needs to know the precoder. This is what "non-codebook-transparent" precoding means in practice, and it is why the gNB can change its precoder between grants without telling the UE anything.
• DMRS density is a per-UE decision. A stationary UE and a UE on a train can be scheduled in adjacent slots of the same cell with completely different pilot densities. LTE could not do this at all.
• Every DMRS resource element is a data resource element you did not send, charged to the transport block of the UE that needs it rather than spread across the cell (§18).

Table 1. Why the mechanism changed. The "beam / precoder" row is the cause; every other row is a consequence. The companion 02 Radio Frame Structure document owns the resource grid both columns are drawn on.

1.2 What triggers DMRS

Nothing, in the sense of a procedure. DMRS is not a procedure and there is no DMRS trigger, no DMRS request and no DMRS timer. It is a mapping rule: given a grant -- its mapping type, its start symbol, its length, its layer count and the UE's configured DMRS parameters -- the rule says which resource elements of that grant carry pilots instead of data. Both ends run the same rule and neither signals the result. That is precisely why the failure modes in §20 are so quiet: when the two ends run the rule with different inputs, there is nothing in the protocol that notices.

2. Front-Loading, and the Price of It

LTE put its pilots in the middle of the subframe as well as at the start, and there was no reason not to: the receiver had to wait for the whole subframe to arrive before it could do anything with it. NR made a different choice, deliberately. The first DMRS symbol sits at the start of the allocation -- symbol 2 or 3 of the slot for mapping type A, the first symbol of the allocation for mapping type B TS 38.211 cl. 7.4.1.1.2. Nothing about channel estimation requires this. Interpolation from pilots spread evenly through the allocation gives a better estimate than extrapolation from pilots at one end.

The reason is pipelining. A receiver that has the channel estimate in hand after the second or third symbol can start equalising, demodulating and feeding the LDPC decoder while the remaining symbols are still arriving. A receiver that has to wait for a mid-allocation pilot cannot begin until the pilot arrives, and the whole of its processing latency then stacks on top of the transmission duration rather than overlapping it. On a 14-symbol slot at 30 kHz that is the difference between finishing 100 µs after the last symbol and finishing 500 µs after it, and the HARQ round trip is built from exactly those numbers -- see the companion 05 HARQ document for how k1 and the N1 processing time consume the budget front-loading protects.

Two smaller consequences of front-loading show up constantly in logs:

• A mapping type A allocation must contain its own DMRS symbol. This is why the start symbol S is restricted to {0, 1, 2, 3} for PDSCH type A: the DMRS is at slot symbol 2 or 3, and an allocation starting at symbol 5 would begin after its own reference signal. The companion 32 SLIV and Time Domain Allocation document derives the S restriction from this constraint; §9 here closes the loop from the other side.
• A very short mapping type B allocation spends a large fraction of itself on pilot. A two-symbol type B PDSCH carries DMRS in one of its two symbols. Half the allocation is reference signal. Front-loading makes the mini-slot possible and simultaneously makes it expensive, which is why mini-slots are a latency tool and never a throughput tool.

3. Configuration Type 1: Comb-2 and Two CDM Groups

There are exactly two DMRS configuration types for PDSCH and PUSCH, selected by dmrs-Type, and they differ only in how the DMRS resource elements are laid out in frequency within a PRB. Everything else -- symbol positions, sequence generation, front-loading -- is common. Type 1 is the default and by a wide margin the more common.

Type 1 places DMRS on every second subcarrier, a comb-2 pattern, with two interleaved combs TS 38.211 cl. 7.4.1.1.1:

Configuration type 1, frequency mapping:

    k = 4n + 2k' + delta        n = 0, 1, 2, ...
                                k' in {0, 1}
                                delta in {0, 1}

  delta = 0  ->  CDM group 0  ->  k = 0, 2, 4, 6, 8, 10   within a PRB
  delta = 1  ->  CDM group 1  ->  k = 1, 3, 5, 7, 9, 11   within a PRB

  k' selects which of the two REs of a group the frequency-domain
  orthogonal cover code element w_f(k') is applied to: the pair is
  (4n, 4n+2) for group 0 and (4n+1, 4n+3) for group 1.

  6 DMRS REs per CDM group, per PRB, per DMRS symbol.

Table 2. Configuration type 1 at a glance. The delay-spread row is the one people forget: CDM only works if the channel is approximately constant across the resource elements the code spans, and type 1 spans twice the bandwidth type 2 does.

4. Configuration Type 2: Adjacent Pairs and Three CDM Groups

Type 2 exists for one reason: more orthogonal ports per DMRS symbol. It gets them by cutting the PRB into three groups of four resource elements instead of two groups of six, and by pairing adjacent subcarriers rather than subcarriers two apart:

Configuration type 2, frequency mapping:

    k = 6n + k' + delta         n = 0, 1, 2, ...
                                k' in {0, 1}
                                delta in {0, 2, 4}

  delta = 0  ->  CDM group 0  ->  k = 0, 1,  6, 7    within a PRB
  delta = 2  ->  CDM group 1  ->  k = 2, 3,  8, 9    within a PRB
  delta = 4  ->  CDM group 2  ->  k = 4, 5, 10, 11   within a PRB

  k' selects which RE of the adjacent pair (6n+delta, 6n+delta+1) the
  frequency-domain OCC element w_f(k') applies to.

  4 DMRS REs per CDM group, per PRB, per DMRS symbol.

Table 3. Type 1 against type 2. Note the two cost rows: at maximum port count the two types cost exactly the same -- the whole DMRS symbol -- and type 2's advantage is entirely in ports per symbol, not in overhead. Type 2 is cheaper only when few CDM groups are declared without data.

5. CDM, Orthogonal Cover Codes, and CDM Groups Without Data

A CDM group is a set of resource elements shared by several antenna ports. Not divided between them -- shared. Every port of the group transmits on every resource element of the group, at the same time, on the same subcarriers, and the receiver separates them by code. That is code division multiplexing, and it is the mechanism that lets four ports fit in a symbol that has room for only two orthogonal frequency combs.

5.1 The frequency-domain cover code

Within a CDM group the resource elements come in pairs: (4n, 4n+2) for type 1, (6n+delta, 6n+delta+1) for type 2. Two ports share the pair, and they are distinguished by a length-2 orthogonal cover code applied across the two resource elements:

Frequency-domain OCC, w_f(k'), k' in {0, 1}:

    port A:   w_f = [ +1, +1 ]        both REs get the same value
    port B:   w_f = [ +1, -1 ]        the second RE is negated

The receiver recovers each port by combining the two REs:

    h_A  =  ( y(k0) + y(k1) ) / 2
    h_B  =  ( y(k0) - y(k1) ) / 2

which is exact if and only if the channel is the same at k0 and k1.
If it is not, energy from port A leaks into the estimate of port B.
This leakage is the practical limit on CDM, and it grows with delay
spread and with the frequency separation of the pair.

That assumption -- flat channel across the code span -- is the reason type 2 pairs adjacent subcarriers. Two subcarriers 30 kHz apart are much more likely to see the same channel than two 60 kHz apart. Type 1 spans twice the bandwidth and so leaks sooner, and in a channel with a long delay spread the practical port isolation of type 1 degrades before type 2's does. Nothing in the configuration warns you about this; it shows up as a MIMO throughput ceiling in dispersive channels and nowhere else.

5.2 The time-domain cover code

With maxLength set to len2 the DMRS occupies two consecutive symbols, and a second length-2 cover code becomes available across those two symbols: w_t = [+1, +1] and [+1, −1]. That doubles the port count again. Section 7 works through it. The important structural point is that the two symbols are one DMRS instance, not two DMRS positions -- a double-symbol DMRS at l0 is a single front-loaded pilot that happens to be two symbols wide, and it is counted once by dmrs-AdditionalPosition.

5.3 CDM groups without data

Here is the part that appears in every DCI dump and confuses everybody. The gNB does not have to use all the CDM groups. A single-layer transmission needs one port, which lives in one CDM group, and the resource elements of the other group or groups are then free -- the gNB can put PDSCH data in them. Or it can leave them empty, which it will do when it is co-scheduling another UE whose DMRS occupies that group.

The UE cannot work out which of those two things happened. It has to be told, and the field that tells it is "Number of DMRS CDM groups without data", encoded jointly with the port list in the DCI's antenna port field TS 38.212 cl. 7.3.1.2.2:

Table 4. The three values and their resource-element consequence. This is not a statement about how many layers this UE has -- it is a statement about which resource elements this UE must not treat as data. In MU-MIMO the two are deliberately different.

The rate-matching mechanism is behind several of the failure modes in §20. The UE builds an ordered list of the resource elements available for data, frequency-first, and maps the coded bits onto it. Disagree about how many resource elements the DMRS symbol holds and every subsequent bit lands in the wrong place -- no gradual degradation, no partial success, just a failed transport block.

6. Antenna Ports, Layers, and the DCI Antenna Port Field

An antenna port in NR is not a physical antenna. It is defined by its reference signal: two signals are on the same antenna port if the channel over which one is conveyed can be inferred from the channel over which the other is conveyed TS 38.211 cl. 4.4. In practice, a DMRS port is a spatial layer, and the port numbering tells you which physical channel the reference signal belongs to:

Table 5. Antenna port numbering, TS 38.211 cl. 4.4. The 1000-series collision between PDSCH DMRS and SRS is real and is resolved only by direction, which is a genuine source of confusion in logs that print port numbers without a direction tag.

6.1 The antenna port field, and how wide it is

The DCI field named Antenna port(s) carries the CDM-groups-without-data value and the port list as a single joint code point. Its width depends on the configured dmrs-Type and maxLength, which is why the DMRS configuration changes the DCI size:

Table 6. Antenna port field widths for DCI format 1_1, one codeword. Two codewords use separate tables. Changing dmrs-Type or maxLength in an RRCReconfiguration therefore changes the total DCI 1_1 payload size -- and until both ends apply the change at the same slot, the UE is blind-decoding for the wrong length and finds nothing at all. The companion 31 PDCCH Blind Decoding document explains why that produces silence rather than an error.

6.2 The type 1, single-symbol table

This is the one you will actually decode, because type 1 with maxLength absent is the default and covers most traffic. Ports are written relative to the base -- port "0" here means 1000 for PDSCH:

Table 7. TS 38.212 Table 7.3.1.2.2-1, one codeword, dmrs-Type type1 and maxLength len1. Note values 2 and 11: both give two layers, and they differ only in whether the two ports share a CDM group. Value 2 costs 6 pilot REs per PRB, value 11 costs 12.

The layer count is not signalled anywhere as a number. It is the length of the port list in the antenna port field, and that is the only place it appears. A log printing "rank 4" is reporting its own interpretation of this field -- so when the port list and the reported rank disagree, trust the port list.

7. Double-Symbol DMRS and the Port Doubling

maxLength = len2 tells the UE that the DMRS may be two symbols long, and the antenna port field then says whether a given grant actually uses one symbol or two. It is a capability switch, not a commitment: a len2-configured UE still receives single-symbol DMRS on most of its grants, and the cost of the configuration is paid in the wider antenna port field (§6.1) rather than in pilots.

Table 8. maxLength len1 against len2. The Doppler row is the counter-intuitive one: len2 with pos1 puts four symbols of pilot in the slot but samples the channel at only two points in time, so it is worse for a fast-moving UE than len1 with pos2, which costs less and samples three times.

8. Configuration, Field by Field

The DMRS configuration is small -- six fields plus an optional PT-RS substructure -- and every one of them has a default that is expressed by absence rather than by a value. That single design choice is responsible for a disproportionate share of the field problems in §20, so it is worth going through the fields one at a time.

Table 9. DMRS-DownlinkConfig, field by field. Four of the six fields encode their most common value as absence. A configuration dump that shows an empty DMRS-DownlinkConfig is not unconfigured -- it is type 1, pos2, len1, cell-ID scrambling, no PT-RS.

8.1 Two configurations, one per mapping type

The DMRS configuration is not attached to the bandwidth part or to the cell. It is attached to the mapping type, and there are two separate instances inside PDSCH-Config:

PDSCH-Config
 |
 +-- dmrs-DownlinkForPDSCH-MappingTypeA   SetupRelease { DMRS-DownlinkConfig }
 |     applies to every grant whose TDRA row says mappingType typeA
 |     typically: dmrs-AdditionalPosition pos1 or pos2 (long allocations)
 |
 +-- dmrs-DownlinkForPDSCH-MappingTypeB   SetupRelease { DMRS-DownlinkConfig }
       applies to every grant whose TDRA row says mappingType typeB
       typically: dmrs-AdditionalPosition pos0 (short allocations)

PUSCH-Config carries the same pair:
   dmrs-UplinkForPUSCH-MappingTypeA / -MappingTypeB, both DMRS-UplinkConfig

-- Either may be absent. An absent structure for a mapping type that the
-- TDRA list actually uses means the UE applies the all-defaults case:
-- type 1, pos2, len1, cell-ID scrambling.

This separation is the right design and it is also a trap. A cell running both slot-based eMBB and mini-slot URLLC has two DMRS configurations in flight for the same UE at the same time, and a grant switches between them purely by its TDRA row's mappingType field. When a log shows a UE decoding type A grants correctly and type B grants not at all -- or the reverse -- the mapping-type-specific configuration is the first place to look, not the radio.

8.2 dmrs-TypeA-Position: the field that arrives before RRC does

One DMRS parameter does not live in DMRS-DownlinkConfig at all. dmrs-TypeA-Position, which selects whether the mapping type A DMRS is at slot symbol 2 or symbol 3, is broadcast in the MIB TS 38.331 `MIB` and repeated in ServingCellConfigCommon:

MIB ::= SEQUENCE {
    systemFrameNumber              BIT STRING (SIZE (6)),
    subCarrierSpacingCommon        ENUMERATED {scs15or60, scs30or120},
    ssb-SubcarrierOffset           INTEGER (0..15),
    dmrs-TypeA-Position            ENUMERATED {pos2, pos3},
    pdcch-ConfigSIB1               PDCCH-ConfigSIB1,
    cellBarred                     ENUMERATED {barred, notBarred},
    intraFreqReselection           ENUMERATED {allowed, notAllowed},
    spare                          BIT STRING (SIZE (1))
}

-- 1 bit of the 23-bit MIB payload. It has to be here, and not in any
-- dedicated configuration, because the UE must receive SIB1 -- a
-- mapping type A PDSCH -- before any dedicated configuration exists.

The reason it is in the MIB is causal, not historical. SIB1 is carried on a mapping type A PDSCH. To receive that PDSCH the UE must know where its DMRS is. So the DMRS position has to be known before SIB1 is read, which means it has to be in the MIB. The companion 18 MIB and SIB1 IEs document owns the MIB payload in full and the companion 02 Radio Frame Structure document owns how the MIB is carried on PBCH.

9. The First DMRS Symbol: Type A, Type B, and dmrs-TypeA-Position

The position of the first DMRS symbol, written l0 throughout TS 38.211, has exactly two possible definitions and the mapping type chooses between them TS 38.211 cl. 7.4.1.1.2:

Table 10. l0 and ld for the two mapping types. The ld row is the one that produces wrong answers: for mapping type A, ld is not the allocation length. A grant with S = 2 and L = 12 has ld = 14, and it is row 14 of the position table that applies, not row 12.

9.1 Why S is capped for mapping type A

The chain of reasoning is short and it runs in the direction people usually do not expect. The DMRS symbol is fixed at slot symbol 2 or 3. The allocation must contain its own DMRS. Therefore S cannot exceed l0. The (S, L) validity table in TS 38.214 encodes that as S in {0, 1, 2, 3} for PDSCH mapping type A, with S = 3 usable only when dmrs-TypeA-Position = pos3. The restriction is not about control regions or about the CORESET -- it is entirely about the DMRS. The companion 32 SLIV and Time Domain Allocation document works the same constraint from the allocation side.

For PUSCH mapping type A the cap is tighter still: S = 0, full stop. There is no uplink control region to skip, so a nonzero S buys nothing, and the spec simply forbids it.

9.2 Mapping type B and the mini-slot

Mapping type B moves the reference point to the allocation, and that is the whole of what mapping type B is. Everything else about the mini-slot -- the ability to start at any symbol, the two-symbol PDSCH, the sub-slot PUSCH -- follows from the DMRS travelling with the grant. Rel-15 PDSCH type B permits L in {2, 4, 7} only; PUSCH type B permits L from 1 to 14.

Table 11. Type B DMRS positions, relative to the start of the allocation, from TS 38.211 Table 7.4.1.1.2-3 (PDSCH) and Table 6.4.1.1.3-3 (PUSCH). Positions are given as offsets from S because l0 = 0 for type B. Below ld = 7 there is nowhere to put an additional position, so pos1, pos2 and pos3 all collapse to pos0.

10. Additional Positions: the TS 38.211 Tables

dmrs-AdditionalPosition says how many DMRS symbols the allocation gets beyond the front-loaded one, and the spec then decides where they go as a function of ld. It is not a free choice of positions -- the UE and the gNB both read the same table and there is no signalling of the result. The table is TS 38.211 Table 7.4.1.1.2-3, "PDSCH DM-RS positions l-bar for single-symbol DM-RS"; the PUSCH equivalent is Table 6.4.1.1.3-3 and has the same structure.

Table 12. TS 38.211 Table 7.4.1.1.2-3, PDSCH mapping type A, single-symbol DMRS, normal CP -- the rows for ld = 3 to 14. Values other than l0 are absolute symbol indices within the slot, not offsets. For mapping type A, ld = S + L (§9). Rows below ld = 3 and the mapping type B columns are omitted; the type B columns are in §9.2.

Three structural properties of this table are worth having in your head before you use it:

• Below ld = 8, every value of dmrs-AdditionalPosition gives the same answer. There is no room for an additional position in a seven-symbol allocation, so pos3 behaves exactly like pos0. Configuring pos3 on a cell that only schedules short allocations achieves nothing and costs nothing.
• Only the entries after l0 are absolute. "l0, 7, 11" means the configured l0 -- 2 or 3 -- followed by symbols 7 and 11 of the slot. The second and third positions do not move when l0 moves from 2 to 3, which is why pos2 with l0 = 3 gives {3, 7, 11}: a 4-symbol gap followed by a 4-symbol gap, slightly more even than {2, 7, 11}.
• The additional positions are not evenly spaced. pos2 at ld = 14 gives gaps of 5 and 4 symbols; pos3 gives 3, 3 and 3. The designers chose positions that keep the last pilot near the end of the allocation, because extrapolation past the final pilot is the worst part of the estimate.

10.1 The double-symbol table

With maxLength = len2, a different and much shorter table applies (TS 38.211 Table 7.4.1.1.2-4), and only pos0 and pos1 exist:

Table 13. TS 38.211 Table 7.4.1.1.2-4, double-symbol DMRS, condensed. The exact index of the second pair depends on ld -- check the table for your ld rather than assuming, because the second pair moves between the ld = 12 and ld = 13 rows. What is certain is that there are at most two double-symbol positions in a slot, and each of them occupies two symbols.

11. Doppler, Coherence Time, and Why pos3 Exists

dmrs-AdditionalPosition is the only DMRS parameter whose correct value depends on something the configuration cannot see: how fast the UE is moving. This section does the arithmetic that turns a speed into a parameter value, because without it the four enumeration members are just four numbers.

A moving UE sees a Doppler shift, and the spread of Doppler shifts across the multipath components is what makes the channel change with time. The maximum Doppler frequency is:

    f_d  =  v * f_c / c

      v    UE speed, metres per second
      f_c  carrier frequency, hertz
      c    3 x 10^8 m/s

The coherence time -- the interval over which the channel stays
correlated enough to reuse an estimate -- is inversely proportional:

    T_c  ~  0.423 / f_d       (50% correlation, Clarke model)
    T_c  ~  0.179 / f_d       (a stricter 90% criterion)

The 0.423 figure is used below. It is an approximation and the exact
constant depends on the scattering model; treat the results as an
order of magnitude, not a specification.

Two conclusions fall out. First, what matters is the ratio of coherence time to symbol duration, and because a higher SCS shortens the symbol almost as fast as a higher carrier raises the Doppler, FR2 at 120 kHz is not eight times worse than FR1 at 30 kHz -- compare cases 2 and 4 and the symbol counts are 12 and 6. Second, the parameter is genuinely per-UE: cases 1 and 2 are the same cell and the same numerology, and they want different values.

Table 14. The parameter as a speed limit. "T_c in symbols" is the coherence time divided by the symbol duration at that SCS. The rule of thumb the table encodes: pick a value whose worst pilot spacing is comfortably below T_c in symbols, and no denser. Approximate throughout -- the Clarke constant is model-dependent.

The opposite error is easier to spot. Under-configuring produces BLER that rises with allocation length and with UE speed, and correlates with nothing else. A UE whose short grants succeed and whose 14-symbol grants fail, at good SINR, is a UE whose channel estimate has gone stale by the end of the allocation. Outer-loop link adaptation will respond by cutting the MCS, which fixes the BLER and hides the cause -- see the companion 34 Link Adaptation document for how that feedback loop converts a DMRS density problem into an apparent coverage problem.

12. Choosing a Configuration Before the First Grant

The gNB has to commit to a DMRS configuration in an RRCReconfiguration, before it has scheduled anything, and it cannot change it per grant -- only the number of ports actually used and the CDM groups without data are dynamic, through the antenna port field. So the configuration has to cover the whole envelope of what the scheduler might want to do with this UE.

The decision order matters, because the port requirement is a hard constraint and the Doppler requirement is a trade: