Ferkans — Interactive Telecom Tutor

Beyond the Two-Way Relay

The two-way relay channel of §3 is the pedagogical gateway to compute-and-forward — the smallest nontrivial network where the framework beats classical strategies. The payoff is much broader. §5 surveys three major applications, each of which stretches the Nazer-Gastpar framework into new territory:

(1) The $K$ -user Gaussian interference channel (Nazer-Gastpar 2011 main application). CF gives an inner bound on the symmetric sum-capacity of the $K$ -user GIC that is within a bounded gap of the Etkin-Tse-Wang upper bound — a lattice alternative to Cadambe- Jafar interference alignment.

(2) The integer-forcing linear receiver (Ordentlich-Erez-Nazer 2014). CF as a MIMO receiver: instead of decoding the $K$ streams of a $K \times K$ MIMO channel independently, decode $K$ linearly independent integer combinations and invert the integer matrix. Approaches the MIMO sum-capacity at a fraction of ML complexity.

(3) Distributed source coding (Korner-Marton 1979, revisited). The Slepian-Wolf problem of encoding correlated sources admits a CF-like lattice structure: encode each source with a shared lattice, the joint decoder exploits the integer structure.

All three applications rest on the integer-coefficient search: given the channel or correlation structure, find the best $\mathbf{a}^*(\mathbf{h})$ . This optimisation is NP-hard in general but tractable via lattice reduction for modest $K$ — the concluding topic of this section and of Chapter 18.

,

Definition:
Integer-Forcing Linear Receiver

Consider a $K \times K$ MIMO channel $\mathbf{y} = \mathbf{H} \mathbf{X} + \mathbf{W}$ , where the $K$ rows of $\mathbf{X} \in \mathbb{R}^{K \times n}$ are lattice codewords from the shared Erez-Zamir codebook and each column of $\mathbf{W}$ is $\mathcal{N}(\mathbf{0}, \mathbf{I}_K)$ . The integer-forcing (IF) linear receiver chooses an integer matrix $\mathbf{A} \in \mathbb{Z}^{K \times K}$ with $\det(\mathbf{A}) \neq 0$ and decodes the $K$ integer linear combinations $\mathbf{V} \;=\; \mathbf{A} \mathbf{X} \;\in\; \Lambda_c^{K \times n}$ by applying compute-and-forward row-by-row: the $k$ -th row of $\mathbf{A}$ , $\mathbf{a}_k$ , is decoded from $\mathbf{y}$ at rate $R(\mathbf{H}^T \mathbf{b}_k^*, \mathbf{a}_k)$ where $\mathbf{b}_k^*$ is the optimal row-specific equaliser. The IF sum-rate is $R_{\rm IF}(\mathbf{H}, \mathbf{A}) \;=\; \sum_{k=1}^K \min_ {\mathbf{b} \in \mathbb{R}^K} R(\mathbf{H}^T \mathbf{b}, \mathbf{a}_k).$ Once the $K$ integer combinations are decoded, the original streams are recovered by inverting $\mathbf{A}$ : $\mathbf{X} = \mathbf{A}^{-1} \mathbf{V} \bmod \Lambda_s$ .

Think of IF as the lattice analogue of zero-forcing. Classical ZF decodes each stream independently after equalising to the channel inverse $\mathbf{H}^{-1}$ — incurring noise enhancement $\|\mathbf{H}^{-1}\|$ . IF avoids the inverse by decoding $K$ integer-coefficient linear combinations, each at its own MMSE- scaled CF rate, then inverting the integer matrix $\mathbf{A} \in \mathbb{Z}^{K \times K}$ instead of the channel. The integer inverse is exact (no noise enhancement). The tradeoff: the $K$ CF rates are constrained by the integer coefficient vectors, which may not be the best decomposition.

,

Theorem: Integer-Forcing Achieves a Constant Gap to MIMO Sum-Capacity

Consider the $K \times K$ real MIMO channel $\mathbf{y} = \mathbf{H} \mathbf{x} + \mathbf{w}$ with $\mathbf{w} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ , per-stream power $P$ , and MIMO sum-capacity $C(\mathbf{H}) = \tfrac{1}{2} \log_2 \det (\mathbf{I} + P \mathbf{H}^T \mathbf{H})$ . The integer-forcing receiver with optimised integer matrix $\mathbf{A}^* = \arg\max_ {\mathbf{A}: \det \neq 0} R_{\rm IF}(\mathbf{H}, \mathbf{A})$ achieves $R_{\rm IF}^*(\mathbf{H}) \;\ge\; C(\mathbf{H}) \;-\; K \log_2 \left(\tfrac{2K}{\pi e}\right)^{K/2} \cdot (1 + o(1))$ within a constant gap of the sum-capacity (independent of SNR), and strictly dominates zero-forcing at all $\mathbf{H}$ with $\mathrm{cond}(\mathbf{H}) > K^{1/4}$ .

At high SNR, MIMO sum-capacity grows like $K \tfrac{1}{2} \log_2 \text{SNR}$ . The integer-forcing constant gap is a fixed offset (in bits) that does not grow with SNR. Zero-forcing, by contrast, loses $\tfrac{K}{2} \log_2 \lambda_{\min}^{-1}$ where $\lambda_{\min}$ is the smallest eigenvalue of $\mathbf{H}^T \mathbf{H}$ — unbounded for ill-conditioned channels. IF avoids this noise enhancement by choosing integer rows that are short in the dual lattice of $\mathbf{H}^T \mathbf{H}$ .

Proof

Write the IF sum-rate as a lattice problem

For a fixed integer matrix $\mathbf{A} \in \mathbb{Z}^{K \times K}$ , the row $k$ of the IF receiver achieves CF rate $R(\mathbf{H}^T \mathbf{b}_k^*, \mathbf{a}_k) = \tfrac{1}{2} \log_2 (1/N_{\rm eff}^{(k)})$ where $N_{\rm eff}^{(k)} = \|\mathbf{a}_k\|^2 - P (\mathbf{H}^T \mathbf{a}_k)^T (P \mathbf{H}^T \mathbf{H} + \mathbf{I})^{-1} (\mathbf{H}^T \mathbf{a}_k) = \mathbf{a}_k^T (\mathbf{H}^T \mathbf{H} + \tfrac{1}{P} \mathbf{I})^{-1} \mathbf{a}_k / P$ — a quadratic form in $\mathbf{a}_k$ .

Minimise over $\mathbf{a}_k$: shortest-vector problem

The optimal row $\mathbf{a}_k^*$ minimises $\mathbf{a}_k^T \mathbf{G}_{\rm ch}^{-1} \mathbf{a}_k$ over $\mathbf{a}_k \in \mathbb{Z}^K \setminus \{\mathbf{0}\}$ , where $\mathbf{G}_ {\rm ch} = \mathbf{H}^T \mathbf{H} + \tfrac{1}{P} \mathbf{I}$ . This is the shortest-vector problem (SVP) on the integer lattice $\mathbb{Z}^K$ with the Gram matrix $\mathbf{G}_{\rm ch}^{-1}$ — an NP-hard problem in general.

Full integer matrix via lattice reduction

Extending the single-row SVP to $K$ linearly independent rows is the Shortest-Basis Problem (SBP) — also NP-hard, but with good polynomial-time approximations (LLL reduction). The LLL algorithm runs in $O(K^4 \log B)$ where $B$ is the largest coefficient, and gives an $\mathbf{A}$ with $\det \neq 0$ and $\|\mathbf{a}_k\|_{\mathbf{G}_{\rm ch}^{-1}}^2 \le 2^K \cdot \lambda_k(\mathbf{G}_{\rm ch}^{-1})$ where $\lambda_k$ are the successive minima.

Bound the IF sum-rate via Minkowski

Summing over $k$ , $\sum_k \log_2 \|\mathbf{a}_k\|^2_{\mathbf{G}_ {\rm ch}^{-1}} \le K \log_2 2^K + \sum_k \log_2 \lambda_k = K^2 + \log_2 \det \mathbf{G}_{\rm ch}^{-1} = K^2 - \log_2 \det (\mathbf{I} + P \mathbf{H}^T \mathbf{H})$ . Therefore $R_{\rm IF}^* \ge \tfrac{1}{2} \log_2 \det(\mathbf{I} + P \mathbf{H}^T \mathbf{H}) - \tfrac{K^2}{2}$ in the loose-bound form; tighter Minkowski analysis gives $K \log_2 (2K/\pi e)^{K/2}$ .

Achievability requires the AWGN-good Erez-Zamir lattice

Each row is decoded at the Nazer-Gastpar rate, which requires the shared nested lattice to be Poltyrev-good (Ch. 16). This is the standard achievability assumption and is met by lattice constructions from Construction A or D over capacity-achieving codes. $\blacksquare$

,

Optimal Integer Norm $\|\mathbf{a}^*\|^2$ vs. Channel $(h_1, h_2)$

Heatmap of the optimal integer coefficient norm $\|\mathbf{a}^* (\mathbf{h})\|^2$ over the $(h_1, h_2) \in [0.5, 2.5]^2$ plane for the $K = 2$ compute-and-forward problem. Bright cells mark channels where a short integer vector $\mathbf{a}^*$ aligns well with $\mathbf{h}$ : $(1, 1)$ , $(1, 2)$ , $(2, 1)$ , $(1, 3)$ , etc. Dark valleys mark channels with strong Diophantine mismatch, where $\|\mathbf{a}^*\|^2$ jumps and the CF rate drops sharply. The boundaries between integer regions are the loci where two integer vectors give equal CF rates — a Voronoi partition of the channel plane by integer coefficients. This visualises the shortest-vector problem as a function of the channel.

Parameters

SNR [dB]10

Computation Rate vs. Per-User MAC Rate

Comparison of the Nazer-Gastpar computation rate $R(\mathbf{h}, \mathbf{a}^*(\mathbf{h}))$ (at the optimum integer vector) against the per-user capacity of the Gaussian MAC with decode-and-forward relay strategy — i.e., $\tfrac{1}{2K} \log_2(1 + K \text{SNR})$ for symmetric channels. At low-to-moderate SNR, CF matches or exceeds DF for most channel directions; at very high SNR and when $\mathbf{h}$ admits a good integer approximation, CF equals the single-user cap $\tfrac{1}{2} \log_2(1 + \text{SNR} \|\mathbf{h}\|^2)$ . Plotted for $K = 2$ users and the channel $(h_1, h_2) = (1, 1)$ (symmetric), $(1, 1.5)$ (mild mismatch), $(1, \sqrt{2})$ (irrational ratio).

Parameters

Definition:
Integer Coefficient Search (the $\mathbf{a}^*$ Problem)

Given a channel vector $\mathbf{h} \in \mathbb{R}^K$ and SNR $\text{SNR}$ , the integer-coefficient search problem asks for the integer vector that maximises the CF rate: $\mathbf{a}^*(\mathbf{h}) \;=\; \arg\max_{\mathbf{a} \in \mathbb{Z}^K \setminus \{\mathbf{0}\}} R(\mathbf{h}, \mathbf{a}) \;=\; \arg\min_{\mathbf{a} \in \mathbb{Z}^K \setminus \{\mathbf{0}\}} N_{\rm eff}(\mathbf{h}, \mathbf{a}).$ Equivalently, $\mathbf{a}^*$ is the shortest vector in the quadratic form $N_{\rm eff}(\mathbf{h}, \mathbf{a}) = \mathbf{a}^T \left( \mathbf{I} - \frac{\text{SNR}\, \mathbf{h} \mathbf{h}^T}{1 + \text{SNR}\, \|\mathbf{h}\|^2} \right) \mathbf{a},$ i.e., the shortest non-zero vector of $\mathbb{Z}^K$ under the Gram matrix $\mathbf{G}_{\rm ch}(\mathbf{h}) = \mathbf{I} - \text{SNR}\, \mathbf{h} \mathbf{h}^T / (1 + \text{SNR}\, \|\mathbf{h}\|^2)$ .

The SVP under $\mathbf{G}_{\rm ch}$ is NP-hard in general (Ajtai 1996) but admits good approximation algorithms: LLL reduction (Lenstra-Lenstra-Lovász 1982) gives a vector within a $2^{K/2}$ -factor in $O(K^4 \log B)$ ; Fincke-Pohst (1985) finds the exact optimum in the worst case at exponential cost, but is fast in practice for $K \le 8$ . For large $K$ (say $K > 20$ ), the combinatorial explosion starts to bite, and heuristic approaches (stochastic local search, quantum-inspired SVP) are active research topics.

,

LLL-Based Integer-Coefficient Search

Complexity:

O(K^4 \log B)

where

K

is the number of users and

B

is the largest integer coordinate; polynomial in

K

but with constants that grow with

K^2

. Suitable for

K \le 20

; beyond that, heuristic SVP solvers (Schnorr-Euchner, block Korkine-Zolotarev) are preferred.

# Given: channel h in R^K, SNR in dB; target: a* in Z^K maximising R(h, a)

# Construct the Gram matrix G_ch

SNR_lin = 10^(SNR_dB / 10)

G_ch = I_K - (SNR_lin / (1 + SNR_lin * ||h||^2)) * h * h^T

# Cholesky factorise G_ch = L * L^T, so ||a||_{G_ch}^2 = ||L^T a||^2

L = cholesky(G_ch)

B = L^T # basis matrix

# LLL reduction on B (Lenstra-Lenstra-Lovasz 1982)

# delta = 0.75 gives good approximation; swap and Gram-Schmidt loop

repeat:

for k = 2, 3, ..., K:

for j = k-1, k-2, ..., 1:

q = round(<B_k, B_j*> / ||B_j*||^2)

B_k = B_k - q * B_j

if ||B_k*||^2 < (delta - mu_{k,k-1}^2) * ||B_{k-1}*||^2:

swap(B_k, B_{k-1})

restart

until no changes

# Extract shortest vector: first row of reduced basis

a_star = B[1] # after LLL, first basis vector approximates SVP

# Compute rate

N_eff = a_star^T * G_ch * a_star

R_star = 0.5 * log2(max(1/N_eff, 1))

return a_star, R_star

In practice one also runs Fincke-Pohst (exact SVP) to verify that the LLL output is optimal; for $K \le 5$ the Fincke-Pohst enumeration over the box $\|\mathbf{a}\|_\infty \le 5$ (at most $11^K = 161051$ candidates for $K = 5$ ) is faster than LLL and gives the guaranteed optimum.

Common Mistake: Integer-Coefficient Search Is NP-Hard in General

Mistake:

Assuming the integer-coefficient search for compute-and-forward can always be done in polynomial time. The problem reduces to the shortest-vector problem (SVP) in a specific Gram metric, which is NP-hard under randomised reductions (Ajtai 1996).

Correction:

For small $K$ (say $K \le 10$ ), LLL reduction and Fincke-Pohst enumeration are effective in practice. For $K > 20$ , the problem becomes computationally challenging and heuristics are required. This is why practical deployment of CF remains limited to low- $K$ settings — two-way relay ( $K = 2$ ), small-network uplink ( $K = 4$ - $8$ ). For massive-MIMO settings ( $K = 64$ or more), CF with full integer search is not tractable; instead, one uses reduced-search CF (restrict to $\|\mathbf{a}\|_\infty \le 2$ ) or integer-forcing with pre-specified $\mathbf{A}$ .

Why This Matters: Integer-Forcing and Massive MIMO

Massive MIMO (Ch. 20) deploys $M$ antennas at the base station with $M \gg K$ (typically $M = 64$ , $128$ ; $K = 16$ ). The integer-forcing framework of this section extends naturally: the IF receiver now chooses integer matrix $\mathbf{A} \in \mathbb{Z}^{K \times K}$ with $K$ moderate, and the $M$ -dimensional receive beamforming gives an effective channel $\tilde{\mathbf{H}} \in \mathbb{R}^{K \times K}$ that is typically well-conditioned (the $M/K$ asymptotic regime of Marzetta 2010). Integer-forcing then achieves near-sum-capacity at a fraction of the ML complexity — a natural fit for massive MIMO processing.

The Sakzad-Viterbo-Hong-Bakrani 2013 "Integer-forcing MIMO linear receivers" paper gave a $3$ -dB gain over zero-forcing for $K = 4$ spatially-multiplexed streams on a correlated Rayleigh channel. For $K = 16$ the gain would be larger, but the integer- coefficient search complexity grows as $K^4$ — tractable at the base station but far more expensive than the ZF $K^3$ baseline. Chapter 20 revisits this tradeoff in the massive-MIMO limit.

See full treatment in Chapter 20

🔧Engineering Note

CF Prototypes vs. Commercial Deployment

A decade after Nazer-Gastpar, compute-and-forward remains primarily a research tool. Academic prototypes exist: Nam-Ghassemi et al. 2013 (TWRC with lattice codes), Wübben-Rost 2015 (multi-hop satellite), Naderializadeh-Avestimehr 2017 (CF for interference networks). But no major commercial standard (LTE, 5G NR, Wi-Fi 7) has adopted CF. The obstacles are practical, not theoretical: (i) sample-level synchronisation across geographically distributed transmitters is hard; (ii) the integer-coefficient search and lattice decoding add receiver complexity beyond current ASIC budgets; (iii) the framework assumes quasi-static channels that are uncommon in vehicular and mmWave scenarios.

Practical Constraints

•
Timing alignment $< T_s/10$ (hard for distributed transmitters)
•
Frequency alignment $< 1/(nT_s)$ (GPS discipline required)
•
Quasi-static channel (coherence time $\gg$ block length)
•
Lattice decoder ASIC: $\sim 10\times$ the area of a ZF decoder

📋 Ref: None (no 3GPP or IEEE standard currently specifies CF)

,

Historical Note: Integer-Forcing: Ordentlich-Erez-Nazer 2014

2014

Three years after the Nazer-Gastpar framework, Or Ordentlich, Uri Erez, and Bobak Nazer published "The Approximate Sum Capacity of the Symmetric Gaussian $K$ -User Interference Channel" (IEEE Trans. IT 2014), which cast CF as a MIMO receiver technique rather than a relay-network primitive. The insight: a $K$ -user interference channel at the receiver is isomorphic to a $K \times K$ MIMO channel, and CF row-by-row on the integer matrix $\mathbf{A}$ gives an integer-forcing alternative to ML / ZF / MMSE receivers. The same framework covers the symmetric GIC sum- capacity, the multi-user MAC-relay, and standard MIMO detection — all unified under the integer-coefficient optimisation.

The paper was awarded the IEEE Information Theory Society Best Paper Award in 2016, recognising the Nazer-Gastpar framework's reach beyond the relay channel. It also sharpened the practical critique: the integer-coefficient search is the bottleneck, and LLL-based approximations suffice at moderate $K$ . The 2014 paper is the direct ancestor of the massive-MIMO extensions of Chapter 20.

Definition:
Distributed Source Coding via Compute-and-Forward

The Körner-Marton distributed source coding problem has two spatially separated encoders observing correlated binary sources $U_A$ and $U_B$ , communicating to a joint decoder that wants to recover $U_A \oplus U_B$ . The classical Slepian-Wolf region gives the separate-rate bounds $R_A, R_B \ge H(U_A \oplus U_B)$ , but the minimum sum rate $R_A + R_B$ can be as low as $H(U_A \oplus U_B)$ — not $H(U_A) + H(U_B)$ .

The CF interpretation: a lattice-coded version of this problem has each encoder $k$ quantise $U_k$ onto a shared lattice $\Lambda$ , transmit the quantisation index, and have the joint decoder compute $\sum_k a_k U_k \bmod \Lambda$ . The Nazer-Gastpar rate $R(\mathbf{h}, \mathbf{a})$ plays the role of the Slepian-Wolf sum-rate bound; the lattice codebook replaces the random-binning argument. Applications include sensor networks (fuse correlated measurements without exchanging them pairwise) and distributed storage (encode data across servers with computable combinations).

This reframing is conceptually elegant but has not yet yielded dominant practical protocols — mostly because pure Slepian-Wolf binning is simpler to implement, and the lattice approach is most advantageous when combined with a downstream channel coding step (as in the CF relay, where the channel is physically combining encoded bits). Body-sensor networks with joint classification are a natural emerging application.

,

Common Mistake: CF Requires Sample-Level Sync of All Transmitters

Mistake:

Assuming compute-and-forward can be deployed in a standard cellular uplink where UEs asynchronously access the channel via random-access preambles or even LTE-style orthogonal scheduling. The MA-phase lattice sum $\sum_k h_k \mathbf{x}_k$ only works if the transmitters are sample-aligned to within a fraction of a symbol period at the relay's sample clock.

Correction:

Practical CF deployment requires either: (a) a common master clock (GPS, femto-cell backhaul), (b) closed-loop TA (timing advance) adjustment as in 5G NR uplink, or (c) sample-rate fractional-delay compensation at the relay. In 5G NR, TA granularity is $16 \cdot T_c \approx 520$ ns per step — far coarser than the $T_s/10 \approx 3\,\mu$ s precision CF needs. Hence CF has not yet been adopted into the 3GPP stack; research prototypes rely on GPS or custom synchronisation.

Quick Check

Why does the integer-forcing receiver outperform zero-forcing on an ill-conditioned MIMO channel?

IF inverts an integer matrix $\mathbf{A}$ instead of the real channel $\mathbf{H}$ ; the integer inverse has no noise enhancement.

IF uses a lattice codebook, which has higher capacity than random Gaussian codebooks.

IF is equivalent to ML decoding, which is optimal.

IF only works when $\mathbf{H}$ is square.

Correction:

IF inverts an integer matrix

\mathbf{A}

instead of the real channel

\mathbf{H}

; the integer inverse has no noise enhancement.

Correct. IF chooses $\mathbf{A} \in \mathbb{Z}^{K \times K}$ with $\det(\mathbf{A}) \neq 0$ , and the integer inverse is exact. ZF applies $\mathbf{H}^{-1}$ , which amplifies noise by $\|\mathbf{H}^{-1}\|^2 \propto 1/\lambda_{\min}^2$ — the inverse of the smallest singular value.

Integer-forcing linear receiver

A MIMO receiver (Ordentlich-Erez-Nazer 2014) that decodes $K$ linearly independent integer combinations of the transmitted lattice codewords — one per row of an integer matrix $\mathbf{A} \in \mathbb{Z}^{K \times K}$ with $\det(\mathbf{A}) \neq 0$ — using compute-and-forward, then inverts the integer matrix to recover the original streams. Achieves within a constant gap of MIMO sum-capacity at much lower complexity than ML.

Integer-coefficient search

The optimisation $\mathbf{a}^*(\mathbf{h}) = \arg\max_{\mathbf{a} \in \mathbb{Z}^K} R(\mathbf{h}, \mathbf{a})$ that finds the best integer vector for compute-and-forward given channel $\mathbf{h}$ . Formulated as a shortest-vector problem (SVP) on a lattice with Gram matrix $\mathbf{G}_{\rm ch}(\mathbf{h}) = \mathbf{I} - \text{SNR}\, \mathbf{h} \mathbf{h}^T / (1 + \text{SNR}\, \|\mathbf{h}\|^2)$ . NP-hard in general; LLL and Fincke-Pohst give tractable approximations and exact solutions respectively at moderate $K$ .

Interference channel (CF inner bound)

The $K$ -user Gaussian interference channel: $y_k = \sum_{j=1}^K h_{kj} x_j + w_k$ for $k = 1, \ldots, K$ . Nazer-Gastpar 2011 showed that each receiver, treating the transmitters as a $K$ -user MAC-relay, can apply compute-and-forward with a receiver-specific integer vector $\mathbf{a}_k$ to achieve a per-user rate that scales with the single-user cut — within a constant gap of the Etkin-Tse-Wang sum-capacity upper bound. An alternative to Cadambe-Jafar interference alignment that does not require infinite channel diversity.

Key Takeaway

Compute-and-forward extends beyond the two-way relay channel to the $K$ -user interference channel, the integer-forcing MIMO receiver (Ordentlich-Erez-Nazer 2014 — a within-constant-gap sum-capacity achievement via lattice coding), and distributed source coding (lattice-binned Körner-Marton). All applications rest on the integer-coefficient search $\mathbf{a}^*(\mathbf{h}) = \arg\min_{\mathbf{a} \in \mathbb{Z}^K} \mathbf{a}^T \mathbf{G}_ {\rm ch}(\mathbf{h}) \mathbf{a}$ — a shortest-vector problem that is NP-hard in general but tractable via LLL reduction at moderate $K$ . Integer-forcing specifically approaches MIMO sum-capacity within a constant gap by decoding $K$ integer combinations (each at its own CF rate) and inverting the integer matrix — avoiding the noise enhancement that plagues zero-forcing.

Applications and Extensions

Beyond the Two-Way Relay

Definition: Integer-Forcing Linear Receiver

Theorem: Integer-Forcing Achieves a Constant Gap to MIMO Sum-Capacity

Write the IF sum-rate as a lattice problem

Minimise over $\mathbf{a}_k$: shortest-vector problem

Full integer matrix via lattice reduction

Bound the IF sum-rate via Minkowski

Achievability requires the AWGN-good Erez-Zamir lattice

Optimal Integer Norm ∥a∗∥2\|\mathbf{a}^*\|^2∥a∗∥2 vs. Channel (h1,h2)(h_1, h_2)(h1​,h2​)

Parameters

Computation Rate vs. Per-User MAC Rate

Parameters

Definition: Integer Coefficient Search (the a∗\mathbf{a}^*a∗ Problem)

LLL-Based Integer-Coefficient Search

Common Mistake: Integer-Coefficient Search Is NP-Hard in General

Why This Matters: Integer-Forcing and Massive MIMO

CF Prototypes vs. Commercial Deployment

Historical Note: Integer-Forcing: Ordentlich-Erez-Nazer 2014

Definition: Distributed Source Coding via Compute-and-Forward

Common Mistake: CF Requires Sample-Level Sync of All Transmitters

Quick Check

Integer-forcing linear receiver

Integer-coefficient search

Interference channel (CF inner bound)

Key Takeaway

Definition:
Integer-Forcing Linear Receiver

Optimal Integer Norm $\|\mathbf{a}^*\|^2$ vs. Channel $(h_1, h_2)$

Definition:
Integer Coefficient Search (the $\mathbf{a}^*$ Problem)

Definition:
Distributed Source Coding via Compute-and-Forward