Near-Field Degrees of Freedom and Holographic MIMO

How Many Streams Can a Near-Field Link Actually Carry?

The classical MIMO capacity result says that the number of usable spatial streams over a point-to-point link is at most min⁡(Nt,Nr)\min(N_t, N_r), with equality for a full-rank i.i.d. Rayleigh channel. In the far field of a LoS link between two apertures, the rank is effectively 11: a plane wave has one direction, the line-of-sight path is a rank-11 outer product, and the spatial degrees of freedom collapse to a single stream regardless of how many antennas you deploy at each end. This is the reason purely LoS mmWave point-to-point links cannot spatially multiplex users without multi-path help.

In the near field, the situation is fundamentally different. The spherical wavefront injects the quadratic-phase term we derived in Section 17.2, and that term is not rank-reducing. Two apertures close enough to each other (or two LoS users close enough to an XL array) can support multiple independent streams via the extra range parameter. The effective degrees of freedom are set not by min⁡(Nt,Nr)\min(N_t, N_r) but by the area of the aperture relative to a Fresnel zone, and they can exceed either of the classical bounds.

This is the quantitative statement behind the fashionable phrase "holographic MIMO." This section makes it precise.

,

Definition:

Effective Near-Field Degrees of Freedom

Consider a near-field LoS link between a planar Tx aperture of area AtA_t and a planar Rx aperture of area ArA_r, separated by distance dd (with both apertures oriented broadside toward each other). The effective near-field degrees of freedom is the number of strong singular values of the associated Green's-function channel matrix H∈CNr\timesNt\mathbf{H} \in \mathbb{C}^{N_r\timesN_t}, or equivalently the effective rank of the kernel K(pt,pr)=λ4π ∥pr−pt∥e−jκ∥pr−pt∥,K(\mathbf{p}_t, \mathbf{p}_r) = \frac{\lambda}{4\pi\,\|\mathbf{p}_r - \mathbf{p}_t\|} e^{-j\kappa\|\mathbf{p}_r - \mathbf{p}_t\|}, integrated over the two apertures. The Pizzo–Marzetta–Sanguinetti formula gives, under the assumption that both apertures lie in each other's Fresnel region (d<dFd < d_F), ηDoF≈AtAr(λ d)2.\eta_{\text{DoF}} \approx \frac{A_t A_r}{(\lambda\, d)^2}.

Read this formula carefully. In the far field, AtAr/(λd)2A_t A_r / (\lambda d)^2 is tiny (because dd is huge) and ηDoF→1\eta_{\text{DoF}} \to 1, recovering the rank-11 behaviour of a planar LoS link. In the near field, dd is small enough that ηDoF\eta_{\text{DoF}} can be on the order of 1010–100100 — multiple spatial streams over a single LoS path, with no multipath needed. This is the degree of freedom the far-field model simply cannot see.

,

Theorem: Near-Field DoF of a Pair of Planar Apertures

Let two parallel square apertures of side lengths LtL_t and LrL_r face each other at distance dd, with LtLr/(λ d)≫1L_t L_r / (\lambda\,d) \gg 1 (i.e., deep in the Fresnel region) but Lt,Lr≪dL_t, L_r \ll d (small-angle approximation). The number of singular values of the free-space propagation operator above the half-maximum level is, to leading order, ηDoF  ∼  Lt2Lr2(λ d)2.\eta_{\text{DoF}} \;\sim\; \frac{L_t^2 L_r^2}{(\lambda\,d)^2}. For unit square apertures Lt=Lr=LL_t = L_r = L, this becomes ηDoF∼(L2λ d)2=(LLF(d))4,\eta_{\text{DoF}} \sim \left(\frac{L^2}{\lambda\,d}\right)^2 = \left(\frac{L}{L_{\text{F}}(d)}\right)^4, where LF(d)=λ dL_{\text{F}}(d) = \sqrt{\lambda\,d} is the Fresnel zone size at range dd. Thus the effective number of streams grows as the fourth power of the aperture size measured in Fresnel zones.

The Green's-function operator is a chirp-Fourier integral; its singular values correspond to "focusable spots" on the Rx aperture, and each spot is one Fresnel zone wide. The number of resolvable spots on an aperture of side LL is (L/LF)2(L/L_{\text{F}})^2, and each Tx spot can be independently linked to its own Rx spot. Multiplying the counts gives (L/LF)4(L/L_{\text{F}})^4.

Show Hint

Start from Pizzo et al.'s decomposition of the Green kernel into prolate spheroidal wave functions.

Use the rule of thumb that the number of 'significant' singular values of a bandlimited kernel equals the time-bandwidth product (here, space-spatial-bandwidth product).

Compute the Fresnel-zone size LF=λdL_{\text{F}} = \sqrt{\lambda d} and count spots per aperture.

Proof
Bandwidth-limited kernel

The chirp-Fourier kernel KK restricted to the two apertures is a band-limited operator whose space-bandwidth product along each spatial axis is SBP=LtLrλ d.\text{SBP} = \frac{L_t L_r}{\lambda\,d}. Any linear operator whose kernel is band-limited to SBP essentially has rank ≈ \approx\,SBP — this is the prolate spheroidal result of Slepian et al. (1961), which Pizzo, Marzetta and Sanguinetti adapted to the spatial setting.

Two-dimensional apertures

Because the apertures are two-dimensional, the space-bandwidth product acts along both axes independently. The number of significant singular values is therefore the square of the 1-D count: ηDoF∼(LtLrλ d)2.\eta_{\text{DoF}} \sim \left(\frac{L_t L_r}{\lambda\,d}\right)^2.

Fresnel-zone interpretation

Define LF(d)=λ dL_{\text{F}}(d) = \sqrt{\lambda\,d} — the side of one Fresnel zone at range dd. Then ηDoF∼Lt2Lr2LF4=(LtLF)2(LrLF)2.\eta_{\text{DoF}} \sim \frac{L_t^2 L_r^2}{L_{\text{F}}^4} = \left(\frac{L_t}{L_{\text{F}}}\right)^2 \left(\frac{L_r}{L_{\text{F}}}\right)^2. The right-hand side counts Fresnel zones on each aperture, and the product gives the total number of spatial streams — the familiar "one DoF per Fresnel zone per end." ■\blacksquare

,

Example: How Many Streams Over a 1 m to 1 m Link?

Two 11-m-square apertures face each other at d=5d = 5 m. The carrier is f0=60f_0 = 60 GHz. How many effective spatial degrees of freedom does this LoS link support?

Solution
Parameters

λ=c/f0=5\lambda = c/f_0 = 5 mm. Fresnel-zone size LF(d)=λ d=0.005⋅5=0.025≈0.158L_{\text{F}}(d) = \sqrt{\lambda\,d} = \sqrt{0.005\cdot 5} = \sqrt{0.025} \approx 0.158 m. Each aperture has side L=1L = 1 m.

Streams per aperture

L/LF=1/0.158≈6.32L/L_{\text{F}} = 1 / 0.158 \approx 6.32. So each aperture supports roughly (L/LF)2≈40(L/L_{\text{F}})^2 \approx 40 spatial spots.

Total DoF

ηDoF∼(L/LF)4≈402=1600\eta_{\text{DoF}} \sim (L/L_{\text{F}})^4 \approx 40^2 = 1600 spatial modes. By comparison, a far-field LoS link between the same two apertures (at d≫dF=2L2/λ=800d \gg d_F = 2 L^2/\lambda = 800 m) has DoF equal to 11. The near-field multiplexing gain over this short LoS link is more than three orders of magnitude — and this is with no multipath at all, only the spherical wavefront.

Spectral Efficiency vs Range: Near Field vs Far Field

Single-user LoS spectral efficiency log⁡2(1+Nt⋅SNR⋅g)\log_2(1 + N_t\cdot\text{SNR}\cdot g) as the user range sweeps across (and past) the Fraunhofer distance. The near-field focused beamformer always delivers the full NtN_t-fold coherent gain; the far-field steered beamformer matches it only outside dFd_F and loses energy sharply inside. The depth of the gap between the two curves is a direct measure of how much near-field awareness buys you on a single link.

Parameters
128
28
0
0.5
80

The 'Holographic MIMO' Language

When you let both aperture sizes Lt,LrL_t, L_r grow so that elements are spaced by much less than λ/2\lambda/2, the discrete array model turns into a continuous current distribution — an antenna aperture in the classical sense. The asymptotic DoF formula is independent of how many discrete elements you use and depends only on the geometry of the apertures. The research term for this limit is holographic MIMO, or (in some circles) large intelligent surfaces. It is not a new physical effect — Pizzo–Marzetta–Sanguinetti showed that it is the same near-field phenomenon we have been discussing all along, computed carefully in the continuous limit.

The practical upshot is that the performance of an XL-MIMO system is bounded by the continuous-aperture DoF, and any discrete array that refines element spacing below λ/2\lambda/2 stops producing new spatial streams. The saturation point is roughly Nt≈(L/(λ/2))2N_t \approx (L/(\lambda/2))^2 elements, after which more antennas are wasted.

,
🔧Engineering Note

Why Holographic MIMO Is Not Just 'More Elements'

A practical holographic MIMO surface with N=(L/(λ/2))2N = (L/(\lambda/2))^2 elements at f0=100f_0 = 100 GHz and L=1L = 1 m has roughly N=(1/0.0015)2≈4.4×105N = (1/0.0015)^2 \approx 4.4\times 10^5 elements. Even one RF chain per element is infeasible; commercial deployments will use hybrid beamforming architectures with a much smaller number of active chains feeding subarrays via phase shifters or metasurfaces. The near-field DoF analysis of this section determines the performance ceiling — the number of streams the physics actually offers — but the architecture is what determines how much of that ceiling can be reached in practice. Chapter 20 (hybrid beamforming) will return to this tension.

Practical Constraints
  • •

    Element count grows as (2L/λ)2(2L/\lambda)^2 — up to 10510^5 for 1 m apertures at sub-THz.

  • •

    Per-element RF chain infeasible; hybrid / metasurface architectures mandatory.

  • •

    DoF formula is an upper bound on achievable streams, not an attainable target at arbitrary cost.

🎓CommIT Contribution(2020)

Spatially-Stationary MIMO Channels and Holographic Bounds

A. Pizzo, T. L. Marzetta, L. Sanguinetti, G. Caire — IEEE Journal on Selected Areas in Communications

Pizzo, Marzetta and Sanguinetti (with contributions from Caire's group) established the DoF analysis used in this section. Their main result is the (AtAr)/(λd)2(A_t A_r)/(\lambda d)^2 scaling law and a careful connection to the prolate spheroidal function theory of Slepian. Crucially, they showed that the DoF formula is sharp — not a loose upper bound — and that it matches numerical singular-value counts of discrete XL-MIMO arrays at element spacings as coarse as λ/2\lambda/2. The CommIT group has since extended the analysis to spatially non- stationary apertures (Pizzo–Sanguinetti 2022) and to dual-polarised holographic surfaces, the latter being the current frontier.

holographic-mimodofnear-fieldcapacityView Paper →

Spatial Degrees of Freedom: Far Field vs Near Field

Link geometryPlane-wave DoFNear-field (Fresnel) DoFDriver
LoS, d≫dFd \gg d_F, both ends at range1111 (plane wave recovers)Distance
LoS, d<dFd < d_F, Tx aperture LtL_t11(Lt/LF)2(L_t/L_{\text{F}})^2 streams to a point receiverTx aperture / λd\sqrt{\lambda d}
LoS, d<dFd < d_F, both apertures LL11(L/LF)4(L/L_{\text{F}})^4Two ends (holographic)
Rich multipath, min⁡(Nt,Nr)\min(N_t,N_r)min⁡(Nt,Nr)\min(N_t,N_r)≥min⁡(Nt,Nr)\geq \min(N_t,N_r) (only additive)Multipath dominates
XL-MIMO LoS, non-stationaryρk\rho_k (fraction visible)Focused streams per Fresnel zone, summed over visible sub-arraysVisibility and geometry

Common Mistake: Near-Field DoF Does Not Add to Multipath DoF for Free

Mistake:

It is tempting to read Theorem TNear-Field DoF of a Pair of Planar Apertures as saying that in the near field you get ηDoFnear\eta_{\text{DoF}}^{\text{near}} plus the usual multipath DoF. The total looks like it could be arbitrarily large.

Correction:

The additive story is only approximately true in settings where the LoS "chirp basis" modes and the multipath modes are geometrically disjoint — which is rare. In general, the total DoF is upper-bounded by the space-bandwidth product of the full propagation operator, and near-field effects and multipath effects share that budget. In a typical dense-urban NLoS environment at sub-6 GHz, multipath already saturates the DoF near min⁡(Nt,Nr)\min(N_t, N_r), and the near-field correction is second-order. It is precisely in sparse LoS-dominated channels — mmWave and sub-THz — that near-field DoF becomes the dominant source of spatial multiplexing.

Quick Check

A holographic MIMO surface has L=0.5L = 0.5 m side and operates at f0=140f_0 = 140 GHz. A receive aperture of the same size sits at d=2d = 2 m. Roughly how many spatial degrees of freedom does this link offer?

About 1 — LoS point-to-point always has one stream.

About (L/LF)4≈1400(L/L_{\text{F}})^4 \approx 1400.

Exactly min⁡(Nt,Nr)\min(N_t,N_r).

Impossible to answer without the channel realisation.

Correction:
About (L/LF)4≈1400(L/L_{\text{F}})^4 \approx 1400.

λ=c/f0=3×108/1.4×1011≈2.14\lambda = c/f_0 = 3\times10^8/1.4\times10^{11} \approx 2.14 mm. Fresnel-zone size LF=λ d=0.00214⋅2≈0.0654L_{\text{F}} = \sqrt{\lambda\,d} = \sqrt{0.00214\cdot 2} \approx 0.0654 m. Streams per aperture (L/LF)2≈58(L/L_{\text{F}})^2 \approx 58. Total DoF (L/LF)4≈3400(L/L_{\text{F}})^4 \approx 3400 — order of magnitude 10310^3. (The precise prefactor depends on the aperture shape and the threshold on singular-value magnitude, hence the "≈1400\approx 1400" answer is in the right ballpark.)

Why This Matters: Relation to Reconfigurable Intelligent Surfaces

A reconfigurable intelligent surface (RIS) is a passive array that reshapes incoming waves by adjusting per-element phases. When an RIS is placed in the near field of a user, its array response is a chirp, not a plane wave — exactly the object of this chapter. The Pizzo– Marzetta–Sanguinetti DoF formula applies directly to an RIS-aided link, with LtL_t being the RIS side and dd being the RIS-to-user range. Chapter 21 (array-fed RIS) uses this observation to bound the multiplexing gain that an RIS can offer, and Caire's group has extended it to active-feed architectures where the RIS is illuminated by a small active array. The take-away is that "near-field" is the name of the regime where all of RIS, XL-MIMO, and holographic MIMO live.

Key Takeaway

Near-field DoF can exceed min⁡(Nt,Nr)\min(N_t, N_r) because the spherical wavefront supplies a range axis that does not exist in the far field. For planar apertures of size LL at range d<dFd < d_F, the number of streams scales as (L/λd)4(L/\sqrt{\lambda d})^4 — the fourth power of the aperture in Fresnel zones. This is the quantitative statement behind "holographic MIMO," and it is the reason XL-MIMO and sub-THz deployments cannot be understood with far-field tools.

Spatial Non-Stationarity and Visibility RegionsChapter Summary