I think one of the most confusing concept in LTE physical layer is the concept of 'Antenna port'.  Part of the trouble is the name. The word antenna points at hardware, so the reader expects to be told which piece of metal a port is. The specification never answers that, because the term is not a hardware term. Counting the antennas on a base station tells you nothing about how many ports it is using.

The rest of the trouble is the shape of the definition. Most definitions say what a thing is made of. This one says what a receiver is allowed to assume once it has measured the thing. It defines the port by its consequence rather than by its substance, so you finish the first reading without the answer you were looking for. The page works through it in that order, starting with the specification text itself.

The official definition of Antenna port goes as follows. (To be honest, this official definition does not make any clear sense to me)

Followings are the topics :

• What does the specification say?
• Simply put
• Where does the antenna port appear in the processing chain?
• What do the resource grids look like?
• What does the definition actually promise?
• What quasi co-location adds
• What the port numbers tell you
• What the current specification adds
• Reference

What does the specification say?

The specification defines the antenna port twice, once for the uplink resource grid and once for the downlink, and both definitions open with the same sentence. Both are quoted below in full, together with the tables they point at.

36.211 5.2.1 Resource grid (Uplink) says :

An antenna port is defined such that the channel over which a symbol on the antenna port is conveyed can be inferred from the channel over which another symbol on the same antenna port is conveyed. There is one resource grid per antenna port. The antenna ports used for transmission of a physical channel or signal depends on the number of antenna ports configured for the physical channel or signal as shown in Table 5.2.1-1.

< 36.211 - Table 5.2.1-1: Antenna ports used for different physical channels and signals >

The uplink port number is not just an index. It also states how many ports were configured, which is why the same physical channel appears three times across the columns.

• PUSCH and SRS share one numbering : port 10 for a single configured port, 20 and 21 for two, and 40 to 43 for four.
• PUCCH is numbered apart : 100 for one port and 200 and 201 for two, and it has no four port column at all.
• A dash means the combination does not exist : index 2 and 3 appear only in the four port column.

36.211 6.2.1 Resource grid (Downlink) says :

An antenna port is defined such that the channel over which a symbol on the antenna port is conveyed can be inferred from the channel over which another symbol on the same antenna port is conveyed. For MBSFN reference signals, positioning reference signals, UE-specific reference signals associated with PDSCH and demodulation reference signals associated with EPDCCH, there are limits given below within which the channel can be inferred from one symbol to another symbol on the same antenna port. There is one resource grid per antenna port. The set of antenna ports supported depends on the reference signal configuration in the cell:

NOTE : As you see here, there are several different port combinations for a specific refernece signal type. Which of the combination is used is determined by a specific antenna configuration (i.e, Transmission Mode). For further details on the antenna port combination and each transmission mode and reference signal, refer to Transmission Mode page and Reference Signal (Downlink) page.

Simply put

Those two quotations and their tables contain a great deal of detail. Three statements are left once you take the detail away, and the rest of this page keeps returning to them.

Simply put,

• Antenna port is logical concept, not a physical concept (meaning 'Antenna port' is not the same as 'Physical Antenna')
• Each Antenna port represents a specific channel model
• The channel that is transmitted by a specific antenna port can be done by using the reference signal assinged fort the port (This is why each antenna port has its own reference signal)

The first of those three is the hardest to accept, so it is worth a moment. A logical port does not map one to one onto a radiating element, and it fails to do so in both directions at once.

• Several elements can produce one port : a precoder spreads one port across every element in the array, and the UE still sees a single channel.
• One element can produce several ports : a cross-polarised element carries two, and transmission mode 9 puts eight ports through the same radio hardware. Figure 3 below is that case.
• The count is configuration, not inventory : the UE learns how many ports there are from signalling. Counting the antennas on the mast answers a different question.

The second statement follows from the first. Once a port stops being a piece of hardware, something else has to say what makes two ports different, and the answer is the channel. Two ports may reach the UE through the same physical antenna and remain separate ports, because the channel each one presents is its own. The third statement then supplies the mechanism, since the reference signal is what lets the UE measure that channel and tell one port from another.

Where does the antenna port appear in the processing chain?

To be honest, any of the verbal description of Antenna port was not so clear to me for a long time. I am kind of person who has huge difficulties on understanding things if I don't visualize it (have any form of visual image). Just to give you another angle of the concept of Antenna port, I will try to show you on exactly which point in physical layer processing the antenna port are introduced. As illustrated below, antenna port is introduced in Precoding process at first and each antenna port will generate its own resource grid.

Figure 1. Antenna ports appear at precoding and not before it. Everything to the left of that block is counted in layers, and everything to the right is counted in ports.

• Precoding is the boundary : the annotation says that all the layer data gets combined and redistributed to each antenna port data, which is the moment layers stop and ports begin.
• Each port gets its own resource element mapper : the chain splits at point B, and the two mappers drawn are per port rather than per code word.
• Each port gets its own resource grid : the annotation at point C says so directly, which is the same statement as the specification's one resource grid per antenna port.
• Point C is where the examples below are taken : the grids further down this page are what an observer at C would see.

What do the resource grids look like?

Now let's take a look at some of practical examples of how each of antenna port are associted each resource grid. These example shows the all the resource grid that can be observed at point (C) on the physical layer processing shown above. In these examples, I will draw the resource grid with only one RB just for simplicity.

Example 1 > 4x4 MIMO, Transmission Mode 3 or 4.

Figure 2. Four ports, four grids, one RB each. The reference signal of each port occupies a different set of resource elements, and the other three ports leave those elements empty.

• Each port carries its own reference signal : port 0 in red, port 1 in blue, and ports 2 and 3 in yellow, each at its own resource elements.
• The grey elements are the important part : where one port places its reference signal the other ports transmit nothing there, so the measurement is not corrupted by the neighbouring ports.
• PDSCH fills what is left : the green elements carry data on all four ports, because in these transmission modes the data is spread across all of them.
• The yellow column is control : the leftmost symbols of the subframe carry control rather than PDSCH, and that is the same on every port.

Example 2 > Transmission Mode 9, 4 Layer.

Figure 3. Transmission mode 9 needs eight grids, not four. Ports 0 to 3 still exist and still carry their reference signals, and none of them carries any data.

• The top row is unchanged and empty of data : ports 0 to 3 carry CRS exactly as before, but every PDSCH element is grey.
• The data moved to ports 7 to 10 : the bottom row carries PDSCH in green with its own reference signal in red, which is the UE-specific demodulation reference signal.
• Eight ports do not mean eight antennas : this is the clearest case on the page of ports being logical. The same radio hardware produces all eight grids.
• The two rows answer different questions : the CRS ports remain for cell-wide measurement and synchronisation, while the DMRS ports carry this UE's data.

That split is the reason the concept is worth the trouble. A UE in this mode demodulates from ports 7 to 10 and keeps measuring ports 0 to 3, and the specification needs the port idea to say that cleanly without mentioning a single antenna element.

What does the definition actually promise?

That definition reads as circular the first time through, and that is the usual complaint about it. It becomes clear once you read it as a promise made to the receiver rather than as a description of hardware.

The promise is this. If the UE measures the channel on one symbol of a port, it may assume the same channel applies to another symbol of that same port. Nothing else in the specification lets it do that. This is the reason each port carries its own reference signal. The reference signal is the symbol the UE measures, and the promise is what lets that result apply to the data symbols beside it.

Read that way, the definition is doing something quite specific. It is not saying what a port is made of. It is saying how far a channel estimate may be carried, and the answer is exactly as far as the port extends and no further.

The quoted text adds that for several reference signal types there are limits given below within which the channel can be inferred. Those limits are the part that makes the rule usable, and they differ by signal type. The table below collects them from 36.211 clause 6.2.1.

The pattern in that table is worth naming. The more a reference signal is tied to one UE, the narrower the window in which its channel estimate stays valid. CRS has no window at all, because it is transmitted continuously across the cell. A UE-specific reference signal is precoded for one UE in one allocation, so its estimate expires at the edge of that allocation.

• A port is a promise, not a piece of hardware : it says a channel estimate from one symbol applies to another symbol on the same port.
• That promise is why every port carries a reference signal : without one there would be nothing to measure the channel from.
• The window is set by the reference signal type : a PRB pair for EPDCCH DMRS, a subframe and a PRG for PDSCH DMRS, and no stated limit for CRS.
• Crossing the window is not allowed : an estimate from a different PRB pair or a different subframe is not covered by the definition, whatever the hardware happens to do.

What quasi co-location adds

The port definition deliberately says nothing about two different ports. That leaves a gap, because a real receiver has to combine information across ports all the time. It needs a frequency offset, a timing reference and a delay spread before it can even estimate a channel, and it will not get those from the port it is about to demodulate.

36.211 closes the gap with a second and weaker relationship. Two antenna ports are said to be quasi co-located if the large-scale properties of the channel over which a symbol on one antenna port is conveyed can be inferred from the channel over which a symbol on the other antenna port is conveyed. The large-scale properties include one or more of delay spread, Doppler spread, Doppler shift, average gain, and average delay.

Notice what is not on that list. The channel itself is not, and neither is anything the UE could use to demodulate directly. Quasi co-location hands over statistics, not coefficients.

Figure 4. The two relationships side by side. Within a port the estimate applies to the whole channel. Between two quasi co-located ports it applies to the five large-scale properties only.

• Within a port, across ports : the port definition works inside one port, and quasi co-location is the only thing that reaches between two.
• Five properties, and no more : delay spread, Doppler spread, Doppler shift, average gain and average delay.
• It is a weaker promise on purpose : statistics change slowly and stay valid when the precoder changes, while the channel itself does not.
• It is what makes CSI-RS usable : a UE can track timing and frequency on CRS and still apply that tracking when it measures a CSI-RS port.

What the port numbers tell you

Port numbers on this page look arbitrary at first, and they are not. They are allocated in disjoint ranges so that a number alone identifies which reference signal a port belongs to, and in the uplink they carry a second piece of information as well.

The uplink case is 36.211 Table 5.2.1-1, reproduced near the top of this page. Read down the columns rather than across the rows and the scheme appears. PUSCH and SRS use port 10 when one port is configured, 20 and 21 when two are, and 40 to 43 when four are. PUCCH uses 100 for one port and 200 and 201 for two.

So the uplink port number encodes the configured port count as well as the index. A log line naming port 41 is telling you two things at once : this is the second of four configured ports, and it belongs to PUSCH or SRS rather than PUCCH.

The downlink numbering follows the same idea without the count. Ports 0 to 3 are cell-specific reference signals and 4 is MBSFN. Ports 5 through 14 are the UE-specific and PDSCH demodulation ports, 15 upward are CSI reference signals, and the 107 to 110 range belongs to EPDCCH demodulation. The ranges never overlap, so a port number is unambiguous on its own.

• The number identifies the reference signal : the ranges are disjoint by design, so no context is needed to read one.
• In the uplink the number also gives the count : 10 means one port, 20 and 21 mean two, 40 to 43 mean four.
• PUCCH is numbered separately from PUSCH : the 100 and 200 series keeps it apart from the 10, 20 and 40 series.

What the current specification adds

The lists quoted on this page came from an earlier version of 36.211, and the set of antenna ports has grown since. Anyone comparing this page against a current specification will find rows that are not here, so it is worth naming what changed.

The current specification adds three reference signal types that the quoted list does not include. Demodulation reference signals associated with MPDCCH have their own ports, with the channel inferable only within the same set of consecutive subframes and at the same PRB index. Demodulation reference signals associated with SPDCCH have ports of their own as well. And UE-specific reference signals for PDSCH now split into a non-BL/CE case and a BL/CE case, each with a different inference window.

The CSI reference signal row has grown furthest. The list on this page stops at sixteen ports. Current 36.211 states that CSI reference signals support a configuration of 1, 2, 4, 8, 12, 16, 20, 24, 28, or 32 antenna ports. The concept did not change, and the count did, which is what full dimension MIMO required.

• The port concept is stable, the port list is not : the definition has survived unchanged while the set of signals using it kept growing.
• CSI-RS now reaches 32 ports : against the sixteen the list on this page ends at.
• MPDCCH, SPDCCH and BL/CE ports arrived later : each brought its own inference window rather than a new kind of port.

Reference

One specification contains everything quoted above. The transmission modes and reference signal layouts that decide which ports a cell actually uses are on the two pages this page already points to.

• 36.211 - E-UTRA; Physical channels and modulation, v19.3.0. Clause 5.2.1 is the uplink resource grid and Table 5.2.1-1, and clause 6.2.1 is the downlink resource grid, the per-signal inference limits and the definition of quasi co-location.
• 5G Antenna Ports - the same concept in NR, on this site.