PaperlayerView exact revision on arXiv

Source-linked AI summary

The Degrees of Freedom of MIMO Interference Channels without State Information at Transmitters

Yan Zhu, Dongning Guo

arXiv:1008.5196v2cs.IT

TL;DR

The paper asks for the DoF region of two-user MIMO interference channels when receivers know the channel but transmitters do not. It characterizes the high-SNR region for arbitrary antenna configurations under isotropic fading, using channel-independent random Gaussian codebooks. The result shows that beamforming and interference alignment provide no DoF gain without CSIT.

  • Problem

    The paper addresses the missing complete DoF characterization for two-user MIMO interference channels with receiver-only channel knowledge and isotropic fading.

  • Method

    The paper characterizes the DoF region and achieves it with random Gaussian codebooks independent of the channel states.

  • Results

    The DoF region is completely characterized, and beamforming or interference alignment provides no additional DoF gain without CSIT.

  • Takeaways & Limitations

    Without CSIT, structured signaling schemes cannot provide additional high-SNR DoF gains under the paper’s isotropic-fading model.

Abstract

from arXiv · show

This paper fully determines the degree-of-freedom (DoF) region of two-user interference channels with arbitrary number of transmit and receive antennas and isotropic fading, where the channel state information is available to the receivers but not to the transmitters. The result characterizes the capacity region to the first order of the logarithm of the signal-to-noise ratio (SNR) in the high-SNR regime. The DoF region is achieved using random Gaussian codebooks independent of the channel states. Hence the DoF gain due to beamforming and interference alignment is completely lost in absence of channel state information at the transmitters (CSIT).

I. INTRODUCTION

The paper studies two-user MIMO interference channels with receiver-only channel knowledge under isotropic fading and completely characterizes their DoF region. It shows that, without CSIT, beamforming and interference alignment provide no additional DoF gains, closing a prior inner–outer bound gap.

  • I. INTRODUCTION: The model reflects practical settings where channel states are measured at receivers but are difficult for transmitters to acquire accurately and timely.
  • I. INTRODUCTION: The DoF region is completely characterized for the two-user MIMO interference channel with isotropic fading and no CSIT.The result is identified as the paper’s main result.
  • I. INTRODUCTION: Without CSIT, beamforming and interference alignment achieve no additional DoF gains.This contrasts with the full-CSI case shown in.
  • I. INTRODUCTION: Earlier work studied related no-CSIT interference-channel models, but a gap remained between the inner and outer DoF bounds in [10]–.
  • I. INTRODUCTION: For antenna configuration (M1, N1, M2, N2) = (1, 2, 3, 4), the pair (1, 1.5) is not achievable, closing the gap between prior inner and outer bounds.The pair (1, 1) was previously achievable, while the best prior outer bound included (1, 1.5).

II. CHANNEL MODEL

The channel model gives receivers their channel realizations while transmitters know only channel statistics, with block-wise independent isotropic fading and aligned coherence blocks. Isotropic fading makes no signaling direction preferable in the absence of CSIT.

  • II. CHANNEL MODEL: The received signals combine each transmitter’s channel matrix with its transmitted vector and additive i.i.d. circularly symmetric complex-Gaussian noise.
  • II. CHANNEL MODEL: Receivers know their channel realizations, whereas transmitters know only the channel statistics under the no-CSIT assumption.
  • II. CHANNEL MODEL: The channel matrices remain constant for T consecutive slots, then change independently; T = 1 yields i.i.d. fading over time.All links have perfectly aligned coherence blocks.
  • II. CHANNEL MODEL: Fading is assumed full-rank, finite-average-power, and isotropic, meaning right multiplication by any compatible deterministic unitary matrix preserves its distribution.
  • II. CHANNEL MODEL: Isotropic fading is plausible without CSIT because no signaling direction should be preferred, and it includes important models such as Rayleigh fading studied in [10].

III. THE MAIN THEOREM AND ACHIEVABILITY PROOF

Theorem 1 characterizes the DoF region under the stated antenna ordering, and its achievability uses schemes that require no channel-state-dependent codebooks. The region is established through case-specific constructions including MAC intersections, time sharing, and common-message transmission.

  • Theorem 1 assumes N1 ≤N2 and gives the DoF region for the two-user interference channel.
  • Case (a): N1 ≥M2: For case (a), achievable DoF pairs lie in the intersection of two MAC regions and use independent random Gaussian codebooks with common messages only.The two MACs are formed by both transmitters with receiver 1 and receiver 2, respectively.
  • Case (b): M2 > N1 and M1 ≥N1: In case (b), M2 > N1 and M1 ≥N1, the region is a triangle achieved by time sharing between the single-user DoF pairs.The achievable corner points are (N1, 0) and (0, min(M2, N2)).
  • The exact DoF region agrees with the previous outer bound in cases (a) and (b), but the previous bound is strictly loose in case (c) [10]–.
  • The result can be achieved through TDMA or the Han-Kobayashi scheme with common messages only, using codebooks independent of the fading processes.

IV. PROOF OF THE CONVERSE OF THEOREM 1

The converse proof section introduces notation for channel sequences and channel matrices across time slots. It denotes the full channel state over n slots by Hn.

  • The sequence notation xn or {x}n represents x[1], . . . , x[n].
  • The aggregate channel state Hn contains all channel matrices H11, H12, H21, and H22 over n time slots.

A. Fading Statistics Revisited

The fading-statistics treatment replaces isotropic channel matrices by equivalent SVD-based representations without changing capacity. A uniformly randomized right-singular subspace is independent of the singular values and left-singular vectors.

  • An isotropic N ×M random matrix G is decomposed as G = W ΛV †, with K = min(M, N) and a compact SVD.
  • Scrambling the right singular vectors with an independent uniformly distributed unitary Q makes V uniformly distributed and independent of (W, Λ, V 1).
  • The randomized decomposition preserves the channel distribution: G and W ΛV † are identically distributed.
  • The construction extends across T aligned coherence blocks using block-diagonal matrices, preserving the same distributional relation.
  • For each physical channel matrix Hrt, the SVD-derived V rt is uniformly distributed and independent of Hrt, enabling replacement by the equivalent representation without changing capacity.

B. Preliminary Results

The preliminary results develop mutual-information bounds and structural lemmas used in the converse and achievability proofs. They establish that Gaussian inputs provide a sufficient high-SNR benchmark and that mutual information decreases with uniform-transform dimensionality.

  • Theorem 2 (Gaussian input is not too bad): Theorem 2 bounds relevant mutual informations using independent white Gaussian inputs, supporting Gaussian inputs as a converse benchmark.The proof applies these bounds to rate inequalities for both receivers.
  • Lemma 2: Lemma 2 upper-bounds mutual-information changes caused by replacing diagonal fading amplitudes with their element-wise minimum.The bound applies to independent input and Gaussian vectors under positive diagonal random matrices.
  • Lemma 4: Lemma 4 compares a fading channel with Gaussian-input counterparts and bounds block mutual information by n times a single-use conditional mutual information.The bound assumes isotropic channel matrices, independent Gaussian noises, and a power constraint.
  • Converse setup: The converse temporarily gives both receivers all channel-state information, which can only enlarge the capacity region.This strengthens the converse by proving the outer bound under a more favorable receiver-information assumption.

1) Proof of Cases (a) and (b):

For Cases (a) and (b), the converse uses isotropic-fading mutual-information bounds and a MAC interpretation to establish the relevant outer bound. The argument relies on i.i.d. Gaussian inputs and applies the dimensionality lemma to compare the remaining terms.

  • 1) Proof of Cases (a) and (b):: The converse establishes the outer bound for Cases (a) and (b) by comparing mutual-information terms and taking the blocklength to infinity.The final comparison invokes the corresponding bound after n →∞.
  • 1) Proof of Cases (a) and (b):: The mutual information of the isotropic-fading channel without CSIT is maximized by i.i.d. Gaussian inputs.This supplies the Gaussian benchmark used in the converse calculation.
  • 1) Proof of Cases (a) and (b):: The right-hand side of the resulting bound equals the sum ergodic capacity of the MAC formed by the two transmitters and receiver 1.The argument compares the two remaining mutual-information terms using Lemma 3.

2) Proof of Case (c):

For Case (c), the converse isolates the interference subspace using channel alignment and mutual-information comparisons. The resulting bound is established under M2 > N1 > M1, and the same DoF region extends to general coherence time T.

  • 2) Proof of Case (c):: The alignment construction decomposes the received signal into interference-containing and interference-free components, then compares their conditional mutual informations.The proof uses Markov relations, the chain rule, and Lemma 4 to obtain the required inequality.
  • 2) Proof of Case (c):: The equality used in the argument requires W12 to be invertible; it does not generally hold when W12 is column-rank-deficient because the interference-plus-noise is nonwhite.This is the explicit technical boundary of the Case (c) proof step.
  • 2) Proof of Case (c):: The Case (c) converse establishes the target bound under M2 > N1 > M1 by aligning and isolating the interference subspace.The proof concludes after applying Lemma 3 to the aligned subspaces.
  • 2) Proof of Case (c):: Because interference occupies only an M1-dimensional subspace of the N1-dimensional received space, the remaining N1 − M1 dimensions can be used by user 2 without interference.This is the geometric role of the alignment construction in the high-SNR argument.
  • D. Proof of the Converse of Theorem 1 with general T: For coherence time T, stacking T channel uses yields an equivalent block-diagonal channel, and the DoF region is identical to the T = 1 case.The proof repeats the i.i.d. argument with the block-diagonal corollaries replacing the single-use lemmas.

V. CONCLUDING REMARKS

The paper fully characterizes the DoF region for two-user MIMO interference channels without CSIT under its isotropic fading model. Independent Gaussian codebooks achieve the entire region, while beamforming and interference alignment provide no additional high-SNR DoF gains within this model.

  • The paper fully characterizes the DoF region of two-user isotropic-fading MIMO interference channels without CSIT.
  • Independent Gaussian single-user codebooks achieve the entire DoF region.
  • Beamforming and interference alignment cannot provide additional high-SNR DoF gains under the considered model.
  • The result applies to two-user interference channels with i.i.d. block fading, equal coherent times, and aligned coherence blocks across physical links.
  • Beyond this channel model, interference alignment might still provide additional gains without transmitter CSI.
  • With different coherence times or block alignments, interference alignment may achieve DoF pairs excluded by this region, such as (1, 1.5) in the cited configuration.

A. Proof of Theorem 2

The proof establishes key mutual-information inequalities for the MIMO interference-channel analysis by decomposing vector channels and exploiting sufficient statistics, Gaussian representations, and isotropic matrix structure. These steps extend scalar inequalities across channel realizations and time slots.

  • The proof derives the complex-vector inequality I(Hw + v; Hw) ≤ I(Hw + Hẽw; Hw) + I(Hẽw + v; Hẽw) for independent vectors.The argument uses sufficient statistics and mutual-information decompositions.
  • A full SVD reduces the transformed channel to min(M, N) parallel Gaussian channels with identical gains.The unitary transformations preserve the relevant structure while exposing parallel scalar channels.
  • The proof extends the single-channel inequality over time by stacking vectors and applying the result to a block-diagonal channel matrix.Averaging over the distribution of the time-varying channel yields the general result.
  • The mutual-information bounds rely on data processing, Markov chains, and conditioning arguments for independent random vectors.These tools control successive noise additions and establish the required inequalities.
  • Uniformity and identical-distribution arguments for random matrix transformations establish the remaining covariance and matrix inequalities.The proof constructs block-diagonal matrices and uses their distributional symmetry across time.

D. Proof of Lemma 4

The proof of Lemma 4 shows that Gaussian inputs with isotropic covariance maximize the relevant conditional mutual information under isotropic channel matrices. The optimizer is the equal-power covariance Q = (γ/M)I_M.

  • Replacing an input by a Gaussian vector with the same covariance cannot decrease the conditional MMSE used in the mutual-information bound.The argument compares the original input with x_Q ∼ CN(0, Q).
  • The proof uses MMSE–mutual-information relations to convert the optimization into a covariance-design problem over Gaussian inputs.
  • Q = (γ/M)I_M maximizes the relevant mutual information among admissible input covariance matrices.
  • Isotropy makes the objective invariant to the eigenvectors of Q, allowing the optimization to be restricted to diagonal covariance matrices.
  • Concavity and permutation symmetry force all diagonal entries of the optimal covariance to be equal.
Loading 1008.5196v2…