Ferkans — Interactive Telecom Tutor

Why Reciprocity Is a Claim About Hardware

The entire TDD massive-MIMO program rests on a single claim: the downlink channel from the BS to a user equals the transpose of the uplink channel, so that estimating one lets us precode the other. That claim is true in the air — physics is reciprocal — but it is false at baseband, because the uplink observation includes the receive RF chain of the BS while the downlink observation depends on the transmit RF chain. Those two chains have different gains, different phase responses, and different temperature drifts. Without correcting for them, reciprocity is a statement about nothing the baseband ever sees.

Calibration is the procedure that measures those RF-chain differences and applies a diagonal correction so that the baseband's picture of the uplink can be reused as a picture of the downlink. The celebrated result of Rogalin, Bursalioglu, Papadopoulos, and Caire is that we do not need to measure the Tx and Rx chains separately — only their ratio — and that this relative calibration can be done entirely at the BS without cooperation from the user equipment. This section proves that result and then examines how long the calibration holds before temperature drift forces us to update it.

Definition:
TDD Reciprocity with RF-Chain Mismatch

Let $\mathbf{H}_{\rm air}$ denote the over-the-air channel between BS antenna $n$ and user $k$ . By reciprocity of physics $(\mathbf{H}_{\rm air})_{\rm UL} = (\mathbf{H}_{\rm air})_{\rm DL}^T$ . However, the baseband observes the channel through the RF chains. Let $\mathbf{D}_{\rm tx} = \mathrm{diag}(d_1^{\rm tx}, \ldots, d_{N_t}^{\rm tx})$ be the diagonal matrix of complex Tx-chain gains at the BS (each entry is a per-antenna gain and phase), and similarly $\mathbf{D}_{\rm rx}$ for the Rx chains. The user equipment contributes an additional scalar complex gain $u_k$ on uplink and $v_k$ on downlink. The baseband-observed channels are

$\mathbf{H}_{k}^{\rm UL} = \mathbf{D}_{\rm rx} \, \mathbf{H}_{{\rm air},k} \, u_k, \qquad \mathbf{H}_{k}^{\rm DL} = v_k \, \mathbf{H}_{{\rm air},k}^T \, \mathbf{D}_{\rm tx}.$

The per-antenna complex gains drift with temperature on a time scale of minutes to hours; the UE gains drift with gain-control updates. Calibration is the problem of estimating enough of $\mathbf{D}_{\rm tx}, \mathbf{D}_{\rm rx}$ to convert one observed channel into the other.

Theorem: Relative Calibration Suffices

For TDD massive MIMO reciprocity-based downlink precoding, it is sufficient to estimate the ratio $\mathbf{C} = \mathbf{D}_{\rm rx}^{-1}\mathbf{D}_{\rm tx}$ between the Tx and Rx chain responses at the BS, up to a global complex scalar. The UE-side factors $u_k, v_k$ do not need to be calibrated: they appear as a per-user complex gain that is absorbed into the downlink effective channel and tracked by the UE's own estimator.

Reciprocity is about the relative behavior of the Tx and Rx paths at each BS antenna, not their absolute values. A global scalar shift corresponds to an overall power scaling of the downlink beam, which the per-user equalizer absorbs. The UE-side gains don't matter at the BS because the UE receives a scalar channel plus noise and computes its own equalizer anyway.

Show Hint

Write the downlink channel as $\mathbf{H}_{k}^{\rm DL} = v_k \, u_k^{-1} \, (\mathbf{H}_{k}^{\rm UL})^T \, \mathbf{D}_{\rm rx}^{-1} \mathbf{D}_{\rm tx}$ .

Identify which factor depends on the BS only and which depends on the UE only.

The BS-only factor is $\mathbf{C} = \mathbf{D}_{\rm rx}^{-1}\mathbf{D}_{\rm tx}$ ; argue that the UE-only factor is a per-user scalar the UE can estimate.

Conclude that the BS only needs $\mathbf{C}$ , up to a global complex scalar that corresponds to a harmless overall gain.

Proof

Substitute the observation model

From the definitions, $\mathbf{H}_{k}^{\rm DL} = v_k \mathbf{H}_{{\rm air},k}^T \mathbf{D}_{\rm tx}$ and $\mathbf{H}_{k}^{\rm UL} = \mathbf{D}_{\rm rx} \mathbf{H}_{{\rm air},k} u_k$ , so $\mathbf{H}_{{\rm air},k}^T = \frac{1}{u_k} (\mathbf{H}_{k}^{\rm UL})^T \mathbf{D}_{\rm rx}^{-1}.$ Substituting gives $\mathbf{H}_{k}^{\rm DL} = \frac{v_k}{u_k} (\mathbf{H}_{k}^{\rm UL})^T \mathbf{D}_{\rm rx}^{-1}\mathbf{D}_{\rm tx} = \alpha_k (\mathbf{H}_{k}^{\rm UL})^T \mathbf{C},$ where $\alpha_k = v_k/u_k$ is a per-user complex scalar.

Absorb the per-user scalar

The per-user scalar $\alpha_k$ multiplies the effective downlink channel by a constant that depends only on the UE's own RF-chain gains. The UE receives $y_k = \alpha_k \mathbf{c}^T \mathbf{v}_k x_k + w_k$ on a reference pilot, estimates $\alpha_k \mathbf{c}^T \mathbf{v}_k$ as part of its DMRS-based per-layer channel estimate, and uses that combined quantity for coherent demodulation. The BS does not need to know $\alpha_k$ .

Reduce to estimating C

The BS is left with the task of estimating $\mathbf{C} = \mathbf{D}_{\rm rx}^{-1}\mathbf{D}_{\rm tx}$ . This is a purely BS-internal quantity — measurable by injecting pilots between BS antennas and observing which antennas serve as Tx and which as Rx — and can be estimated without any cooperation from the UE.

Global-scalar ambiguity

The estimator can only recover $\mathbf{C}$ up to a global complex scalar $\beta$ : multiplying $\mathbf{C}$ by $\beta$ and dividing all $\alpha_k$ by $\beta$ yields the same observations. The global scalar is a harmless overall downlink power and phase offset, absorbed by the UE's DMRS equalizer. $\blacksquare$

🎓CommIT Contribution(2014)

Hardware Impairments in Large TDD Arrays: Relative Calibration

R. Rogalin, O. Y. Bursalioglu, H. Papadopoulos, G. Caire, A. F. Molisch, A. Michaloliakos, V. Balan, K. Psounis — IEEE Transactions on Wireless Communications, vol. 13, no. 4

The Rogalin-Caire-Molisch paper was the first to formalize the claim that massive-MIMO reciprocity requires only relative calibration of the BS Tx and Rx chains, and to derive distributed algorithms that estimate this relative gain by injecting reference tones between the BS antennas themselves. The key insight — that the UE-side RF chains never need to enter the calibration procedure — removed what had until then been considered a major obstacle to practical TDD massive MIMO. The Argos / ArgosV3 testbeds adopted this framework immediately, and every subsequent massive-MIMO testbed has inherited the same structural separation between BS-internal calibration and UE-side equalization.

calibrationreciprocitycommittddView Paper →

Definition:
Argos Relative-Calibration Protocol

The Argos calibration protocol estimates $\mathbf{C} = \mathbf{D}_{\rm rx}^{-1}\mathbf{D}_{\rm tx}$ using only BS-internal pilot injections. The procedure:

Choose one BS antenna as the reference, indexed $0$ . The reference does not need to be special; any antenna works.
For each non-reference antenna $n = 1, \ldots, N_t-1$ :

a. Transmit a known pilot from antenna $n$ and receive on antenna $0$ . This yields $y_{0,n} = d_0^{\rm rx} h_{n,0} d_n^{\rm tx} s + w$ , where $h_{n,0}$ is the over-the-air coupling between the two antennas.

b. Transmit a pilot from antenna $0$ and receive on antenna $n$ . By reciprocity $h_{0,n} = h_{n,0}$ , so $y_{n,0} = d_n^{\rm rx} h_{n,0} d_0^{\rm tx} s + w$ .

c. Form the ratio $r_n = y_{0,n}/y_{n,0} = (d_0^{\rm rx} d_n^{\rm tx})/(d_n^{\rm rx} d_0^{\rm tx})$ . The air channel $h_{n,0}$ cancels exactly.
The calibration coefficients are $c_n = r_n \cdot c_0$ , where $c_0$ is a free global scalar conventionally set to 1. The calibration matrix is $\mathbf{C} = \mathrm{diag}(c_0, c_1, \ldots, c_{N_t-1})$ .

The protocol requires $\Theta(N_t)$ pilot exchanges, which consume a small fraction of a single slot in a dedicated calibration window.

Argos Calibration: Pseudocode

Complexity:

\mathcal{O}(N_t)

pilot exchanges per calibration cycle.

Input: Reference pilot

s

, array of

N_t

antennas.

Output: Calibration vector

\mathbf{c} \in \mathbb{C}^{N_t}

.

1.

c_0 \leftarrow 1

// anchor global phase

2. for

n = 1, \ldots, N_t-1

do

3.

\quad

Transmit pilot

s

from antenna

n

, receive

y_{0,n}

at antenna

0

4.

\quad

Transmit pilot

s

from antenna

0

, receive

y_{n,0}

at antenna

n

5.

\quad r_n \leftarrow y_{0,n} / y_{n,0}

6.

\quad c_n \leftarrow r_n \cdot c_0

7. end for

8. return

\mathbf{C} = \mathrm{diag}(c_0, \ldots, c_{N_t-1})

9. Downlink precoding: use

\mathbf{C} \, \mathbf{W}

instead of

\mathbf{W}

The calibration coefficients are applied once on the downlink precoder matrix; the runtime cost during operation is zero. ArgosV2 refined this to use multiple reference antennas and a least-squares combiner, which reduces the effect of noise on the reference antenna but has the same asymptotic structure.

Argos Reciprocity Calibration Flow

Animated walkthrough of the Argos relative-calibration exchange: a reference antenna and a generic antenna

n

trade pilots in both directions; forming the ratio

r_n

of the two baseband observations cancels the over-the-air coupling

h_{n,0}

exactly by electromagnetic reciprocity, leaving only the BS-internal Tx/Rx chain ratio that the calibration matrix

\mathbf{C}

must capture.

Two pilot exchanges per antenna pair, then a ratio that eliminates the air channel. The UE never participates.

Example: Why the Reference Antenna Does Not Need to Be Central

A common worry with the Argos protocol is that the reference antenna sits at one end of the array, far from the others, so pairwise pilot exchanges with distant antennas would be too noisy to calibrate reliably. Show that this intuition is wrong: the calibration ratio $r_n = y_{0,n}/y_{n,0}$ is independent of the coupling strength, so any pair with nonzero SNR works.

Solution

Write the measurement model

Let the over-the-air coupling between antennas $n$ and $0$ have complex amplitude $h_{n,0}$ . The two measurements are $y_{0,n} = d_0^{\rm rx} h_{n,0} d_n^{\rm tx} s + w_1$ , $y_{n,0} = d_n^{\rm rx} h_{0,n} d_0^{\rm tx} s + w_2$ .

Apply reciprocity of the air channel

Physics guarantees $h_{n,0} = h_{0,n}$ — this is the statement of electromagnetic reciprocity for passive media. Therefore $y_{0,n}/y_{n,0} = (d_0^{\rm rx} d_n^{\rm tx})/(d_n^{\rm rx} d_0^{\rm tx})$ , independent of $h_{n,0}$ .

Conclude

The only role of $h_{n,0}$ is to set the SNR of each individual measurement. A weakly coupled antenna pair needs longer integration to average out the noise, but the value of the ratio is unchanged. The reference antenna can be anywhere; in practice ArgosV3 uses a couple of rotating references to average over noise realizations and to verify self-consistency.

Definition:
Calibration Residual and Update Period

The calibration residual $\rho_{\rm cal}(t)$ is the RMS relative error between the true calibration matrix and the last stored estimate, expressed as a fraction of the signal amplitude. It has three contributions:

Estimation noise. Finite-SNR pilot exchanges produce an error of order $1/\sqrt{N_{\rm pilot} \gamma_{\rm cal}}$ , where $N_{\rm pilot}$ is the calibration pilot length and $\gamma_{\rm cal}$ is the calibration SNR.
Short-term drift. LO phase noise and analog thermal fluctuations add a stochastic error term that accumulates as $\sqrt{(t - t_0)/\tau_{\rm st}}$ , with short-term correlation time $\tau_{\rm st}$ of seconds.
Long-term drift. RF-chain temperature and aging effects add a slower error that accumulates linearly as $(t - t_0)/\tau_{\rm lt}$ , with long-term time constant $\tau_{\rm lt}$ of minutes to hours.

The calibration update period $T_{\rm cal}$ is the interval at which the BS reruns the Argos protocol and refreshes $\mathbf{C}$ . Shortening $T_{\rm cal}$ improves the residual but steals slots from user data.

Residual Calibration Error vs Update Period

Trace the end-of-cycle residual error $\rho_{\rm cal}$ as a function of the calibration update period $T_{\rm cal}$ , for a given calibration SNR. Fast updates eliminate the drift term but cost slot overhead; slow updates let the drift dominate.

Parameters

Calibration SNR (dB)25

\tau_{\rm lt}

(min)30

Long-term drift time constant

N_{\rm pilot}

64

🚨Critical Engineering Note

Temperature Drift and the Real-World Calibration Cadence

Measurements on LuMaMi and ArgosV3 reported calibration drifts of the order of 0.5 dB per degree Celsius in gain and a few degrees per degree Celsius in phase for commodity LNAs and PAs. In a rooftop or outdoor testbed, ambient temperature can swing 20 degrees over a day, which would destroy reciprocity long before the slot deadline if not re-calibrated. Operating cadences reported in the literature:

LuMaMi: recalibrate once per second in indoor labs, once every $\sim 200~\mathrm{ms}$ during outdoor trials.
ArgosV3: continuous background calibration using one slot per $\sim 500~\mathrm{ms}$ in the steady-state regime.
HHI OpenAirNet: calibration slot scheduled every 100 $\,\mathrm{ms}$ and validated against a parallel offline estimate on a second SDR.

The slot overhead is small ( $< 1\%$ ) but nontrivial, and it must be accounted for in link-budget calculations.

Practical Constraints

•
0.5 dB gain drift per °C is the observed rule of thumb for commodity RF front ends
•
Outdoor deployments need calibration every hundreds of milliseconds to track temperature
•
Indoor lab testbeds can use calibration periods of seconds to minutes
•
Massive Beams cell-free uses a per-AP calibration schedule driven by AP-local temperature sensors

📋 Ref: 3GPP Rel-18 calibration procedures (WG4 M-plane)

,

Common Mistake: Chasing Absolute Calibration

Mistake:

Attempting to calibrate the BS Tx and Rx chains absolutely, for example by connecting each chain to an external vector network analyzer. This is not only unnecessary but it also cannot track drift during operation.

Correction:

Relative calibration via the Argos protocol is both sufficient (by Theorem 26.3) and self-tracking: running it during operation captures drift in real time without external instrumentation. The right mental model is "measure the ratio, not the individual values."

Common Mistake: Over-Frequent Calibration

Mistake:

Running calibration every few slots on the assumption that more calibration is always better. In a mostly-stable system, the estimation noise in each calibration refresh can exceed the drift it would correct.

Correction:

The optimal $T_{\rm cal}$ is where the marginal drift error equals the marginal estimation error. Use the analysis of DCalibration Residual and Update Period: if the drift contribution over one period is much smaller than the per-cycle estimation noise, slow down. A well-tuned system typically calibrates on a sub-second cadence indoors and every 100 ms outdoors.

Historical Note: From Van Trees to Argos: A 40-Year Calibration Journey

1970s--2014

Array calibration is one of the oldest problems in coherent array processing. Van Trees (Optimum Array Processing, 2002) devotes a full chapter to it for beamforming receivers, building on work from the 1970s on adaptive calibration in radar and sonar arrays. Those early algorithms targeted fully offline calibration against a known reference signal — a setting where the calibration error is bounded by the SNR of the reference itself.

The TDD-reciprocity-for-massive-MIMO twist, formalized by Rogalin, Bursalioglu, Papadopoulos, and Caire in 2014, reframed calibration as a BS-internal operation with no external reference and no UE involvement. That reframing is what made online, drift-tracking calibration possible. It is, historically, one of the rare cases where a massive-MIMO-specific result looped back to improve the entire array-calibration literature.

Key Takeaway

Calibration recap. (1) Reciprocity is a statement about the air channel, not the baseband channel; (2) relative calibration of the BS Tx/Rx ratio is sufficient, by Rogalin-Caire; (3) the Argos protocol estimates that ratio with only BS-internal pilot exchanges; (4) drift and estimation noise set an optimal calibration period of hundreds of milliseconds outdoors; (5) absolute calibration is neither necessary nor helpful.

Quick Check

According to the Rogalin-Caire relative-calibration theorem, which of the following quantities is sufficient for the BS to estimate in order to support reciprocity-based downlink precoding?

Absolute Tx-chain gain $\mathbf{D}_{\rm tx}$ and absolute Rx-chain gain $\mathbf{D}_{\rm rx}$ separately

The ratio $\mathbf{C} = \mathbf{D}_{\rm rx}^{-1}\mathbf{D}_{\rm tx}$ up to a global complex scalar

The per-user downlink channel $\mathbf{H}_{k}^{\rm DL}$ from each BS antenna

The UE-side RF-chain gains $u_k$ and $v_k$

Correction:

The ratio

\mathbf{C} = \mathbf{D}_{\rm rx}^{-1}\mathbf{D}_{\rm tx}

up to a global complex scalar

Theorem 26.3 shows that only the relative gain matters: any global complex scalar is absorbed by the UE's per-user equalizer, and the UE-side RF-chain factors $u_k, v_k$ never need to be estimated at the BS.

Why This Matters: Calibration and the O-RAN M-plane

The calibration framework of this section has started to enter 3GPP and O-RAN standardization. 3GPP Rel-18 includes M-plane procedures for scheduled reciprocity calibration on distributed radio units; O-RAN WG4 is in the process of defining a calibration-result format that lets a central unit verify cross-RU consistency. The theoretical guarantees proven here are the minimum correctness conditions that any standard-compliant implementation must meet.

Reciprocity Calibration

Why Reciprocity Is a Claim About Hardware

Definition: TDD Reciprocity with RF-Chain Mismatch

Theorem: Relative Calibration Suffices

Substitute the observation model

Absorb the per-user scalar

Reduce to estimating C

Global-scalar ambiguity

Hardware Impairments in Large TDD Arrays: Relative Calibration

Definition: Argos Relative-Calibration Protocol

Argos Calibration: Pseudocode

Argos Reciprocity Calibration Flow

Example: Why the Reference Antenna Does Not Need to Be Central

Write the measurement model

Apply reciprocity of the air channel

Conclude

Definition: Calibration Residual and Update Period

Residual Calibration Error vs Update Period

Parameters

Temperature Drift and the Real-World Calibration Cadence

Common Mistake: Chasing Absolute Calibration

Common Mistake: Over-Frequent Calibration

Historical Note: From Van Trees to Argos: A 40-Year Calibration Journey

Key Takeaway

Quick Check

Why This Matters: Calibration and the O-RAN M-plane

Definition:
TDD Reciprocity with RF-Chain Mismatch

Definition:
Argos Relative-Calibration Protocol

Definition:
Calibration Residual and Update Period