Boundary Mutual Information (BMI)

• Boundary Mutual Information (BMI) is a specialization of mutual information that focuses on correlations organized by boundaries, interfaces, or shielding layers.
• In renormalization and holography, BMI is employed to diagnose transitions—such as connected versus disconnected configurations—by optimizing local coarse-graining or minimal surfaces.
• BMI frameworks in Gaussian field theory and detector-harvesting studies reveal that boundary conditions critically control correlation scaling, phase transitions, and entanglement features.

Boundary Mutual Information (BMI) denotes a family of mutual-information observables in which the relevant correlations are organized by a boundary, interface, or shielding layer. The phrase is not used uniformly across subfields. In classical lattice renormalization, BMI is the mutual information between a block $B$ and its immediate boundary $\partial B$; in holography it is the mutual information of disjoint boundary subregions; in double holography it refers to intervals on the intersection of a brane and a bath; in Gaussian statistical field theory it is mutual information whose structure is controlled by boundary degrees of freedom; and in detector-harvesting problems it is the mutual information extracted in the presence of a reflecting boundary (Bertoni et al., 2021, Fonda et al., 2014, Liu et al., 13 Feb 2026, Schröfl et al., 2023, Quan et al., 14 Apr 2026).

1. Information-theoretic core and scope

BMI inherits the standard information-theoretic definitions. For discrete random variables $X,Y$ with joint law $P_{XY}$, the Shannon entropy is

$H(X):= -\sum_x P(x)\log P(x),$

and the mutual information is

$I(X;Y):=H(X)+H(Y)-H(X,Y).$

The conditional mutual information is

$I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$

Chain rules imply

$I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$

and the data-processing inequality gives $I(f(X);Y)\le I(X;Y)$ and $I(f(X);Y|Z)\le I(X;Y|Z)$ for any possibly stochastic map $\partial B$0 (Bertoni et al., 2021).

What changes from one BMI literature to another is the meaning of the subsystems. In the renormalization setting, the boundary is the shielding shell $\partial B$1 surrounding a block. In AdS/CFT, it is the conformal boundary on which spatial regions $\partial B$2 and $\partial B$3 are chosen. In double holography, it is the intersection $\partial B$4, on which the relevant intervals live. In Gaussian field theory, the Markov decomposition isolates boundary degrees of freedom that control mutual information between separated regions. In detector problems, the boundary is a reflecting plane that modifies the Wightman function and hence the harvested correlations. A common source of confusion is therefore terminological rather than conceptual: BMI is not one universal observable, but a boundary-adapted specialization of mutual information.

2. BMI in real-space renormalization

For classical lattice spin systems in thermal equilibrium, one considers a Gibbs state $\partial B$5 generated by a local Hamiltonian $\partial B$6. The lattice $\partial B$7 is partitioned as $\partial B$8, where $\partial B$9 is a finite block, $X,Y$0 is the set of sites adjacent to $X,Y$1, and $X,Y$2. A real-space renormalization map is a possibly stochastic channel $X,Y$3, with $X,Y$4 a coarse-grained variable. In this setting, the Boundary Mutual Information is

$X,Y$5

the retained short-range mutual information is $X,Y$6, and the BMI loss is

$X,Y$7

by data processing (Bertoni et al., 2021).

The central structural fact is shielding. By the Hammersley–Clifford theorem, the Gibbs state of a local Hamiltonian is a Markov network on the lattice graph, so the boundary shields the block from the rest: $X,Y$8 After coarse-graining, data processing yields

$X,Y$9

hence $P_{XY}$0 as well. This immediately constrains the renormalized log-density $P_{XY}$1: no new couplings can connect $P_{XY}$2 directly to variables in $P_{XY}$3, and conditionally independent regions inside $P_{XY}$4 remain uncoupled. The nontrivial problem is therefore not the creation of arbitrary distant couplings, but the appearance of couplings that bridge the boundary.

In one dimension, with $P_{XY}$5 split into left and right halves $P_{XY}$6 and $P_{XY}$7, the induced boundary coupling is bounded by the BMI loss: $P_{XY}$8 The theorem shows that maximizing $P_{XY}$9, equivalently minimizing $H(X):= -\sum_x P(x)\log P(x),$0, directly controls the conditional dependence between opposite sides of the boundary. Through known stability bounds for conditional independence, small $H(X):= -\sum_x P(x)\log P(x),$1 implies that the renormalized law is close in total variation to one with conditional independence across the boundary, and therefore to a Hamiltonian with weak or absent boundary-spanning couplings. The higher-dimensional extension proceeds by strip decompositions in each Cartesian direction for isotropic lattices; anisotropic systems may require direction-dependent optimal maps.

This criterion is related to, but distinct from, real-space mutual information (RSMI) in the sense of Koch-Janusz and Ringel. Their objective is to maximize $H(X):= -\sum_x P(x)\log P(x),$2. Because $H(X):= -\sum_x P(x)\log P(x),$3, one has $H(X):= -\sum_x P(x)\log P(x),$4, and by data processing

$H(X):= -\sum_x P(x)\log P(x),$5

Maximizing long-range mutual information is therefore a relaxation of maximizing BMI. The source paper emphasizes the counterintuitive consequence: preserving short-range block-boundary information is the appropriate criterion for suppressing long-range couplings.

The optimization problem also simplifies sharply. For fixed $H(X):= -\sum_x P(x)\log P(x),$6, the map $H(X):= -\sum_x P(x)\log P(x),$7 is convex over stochastic channels, so the optimum is attained at an extreme point, namely a deterministic map. This permits brute-force enumeration for small blocks. In the two-dimensional square-lattice Ising model with a central $H(X):= -\sum_x P(x)\log P(x),$8 block, all $H(X):= -\sum_x P(x)\log P(x),$9 deterministic maps $I(X;Y):=H(X)+H(Y)-H(X,Y).$0 can be enumerated using a $I(X;Y):=H(X)+H(Y)-H(X,Y).$1 marginal estimated by Corner Transfer Matrix methods. The optimal map depends on inverse temperature $I(X;Y):=H(X)+H(Y)-H(X,Y).$2: decimation is optimal at high temperature $I(X;Y):=H(X)+H(Y)-H(X,Y).$3; majority vote with deterministic tie-breaking by decimation is optimal for $I(X;Y):=H(X)+H(Y)-H(X,Y).$4; majority vote with fixed tie-breaking is optimal near criticality $I(X;Y):=H(X)+H(Y)-H(X,Y).$5; and at low temperature $I(X;Y):=H(X)+H(Y)-H(X,Y).$6 a biased map favoring the dominant phase becomes optimal. Majority vote with random tie-breaking is never optimal. Only the local marginal $I(X;Y):=H(X)+H(Y)-H(X,Y).$7 is required, which is a major simplification relative to objectives involving $I(X;Y):=H(X)+H(Y)-H(X,Y).$8.

3. Holographic BMI for disjoint boundary domains in AdS$I(X;Y):=H(X)+H(Y)-H(X,Y).$9

In holography, BMI refers to the mutual information of disjoint spatial domains $I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$0 and $I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$1 on the boundary of AdS$I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$2,

$I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$3

with entanglement entropy evaluated by the Ryu–Takayanagi prescription

$I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$4

On a constant-time slice, the bulk geometry is $I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$5 with metric

$I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$6

and for an embedded surface $I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$7 the area functional is

$I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$8

For smooth $I(X;Y|Z):=H(X|Z)+H(Y|Z)-H(X,Y|Z)=I(X;YZ)-I(X;Z).$9, the UV-regularized area has expansion

$I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$0

while corners introduce an additional logarithmic term,

$I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$1

Numerically one works with the finite part $I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$2, so that

$I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$3

at fixed cutoff (Fonda et al., 2014).

The characteristic phenomenon is a classical phase transition in $I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$4. At small separation, a connected RT surface is globally minimal and $I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$5; above a critical distance $I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$6, the disconnected configuration wins and $I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$7. The transition curve is defined by the vanishing of BMI, and the first derivative is discontinuous there. The paper stresses that this is a leading classical effect: quantum corrections are known to smooth out the transition.

The AdS$I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$8 study uses triangulated surfaces evolved numerically with Surface Evolver and benchmarks the method against analytic solutions. For a disk of radius $I(X;YZ)=I(X;Y)+I(X;Z|Y)=I(X;Z)+I(X;Y|Z),$9, the minimal surface is a hemisphere with

$I(f(X);Y)\le I(X;Y)$0

so $I(f(X);Y)\le I(X;Y)$1. For an infinite strip of width $I(f(X);Y)\le I(X;Y)$2 and longitudinal size $I(f(X);Y)\le I(X;Y)$3,

$I(f(X);Y)\le I(X;Y)$4

For an annulus with $I(f(X);Y)\le I(X;Y)$5, connected minimal surfaces exist only for $I(f(X);Y)\le I(X;Y)$6, and the transition occurs at $I(f(X);Y)\le I(X;Y)$7. By conformal mapping, the corresponding case of two equal circles has critical separation $I(f(X);Y)\le I(X;Y)$8 and disappearance threshold $I(f(X);Y)\le I(X;Y)$9.

The main geometric result is the dependence of BMI on shape and orientation. The paper analyzes ellipses, superellipses, and two-dimensional spherocylinders. Elongation $I(f(X);Y|Z)\le I(X;Y|Z)$0 tends to increase the BMI at fixed $I(f(X);Y|Z)\le I(X;Y|Z)$1 and pushes the transition curve toward the infinite-strip value. Superellipses with $I(f(X);Y|Z)\le I(X;Y|Z)$2 approach the strip transition more closely than ellipses, and spherocylinders lie even closer still. For anisotropic shapes, the transition depends on relative orientation; this is explicit in the numerically constructed surfaces for disjoint squares. In this literature, BMI is therefore a diagnostic of entanglement-wedge connectivity and of the sensitivity of holographic correlations to shape, curvature, and orientation of boundary domains.

4. BMI in double holography

In double holography, the relevant composite system consists of AdS$I(f(X);Y|Z)\le I(X;Y|Z)$3 gravity coupled to a flat heat bath, realized geometrically by an asymptotically AdS$I(f(X);Y|Z)\le I(X;Y|Z)$4 bulk truncated by a codimension-one Planck brane $I(f(X);Y|Z)\le I(X;Y|Z)$5. The intervals whose BMI is studied lie on the intersection $I(f(X);Y|Z)\le I(X;Y|Z)$6, a straight line in the boundary Minkowski space. From the brane perspective, the entropy of a region $I(f(X);Y|Z)\le I(X;Y|Z)$7 is computed by a quantum extremal surface: $I(f(X);Y|Z)\le I(X;Y|Z)$8 Double holography maps this generalized entropy to a purely geometric problem in one higher dimension,

$I(f(X);Y|Z)\le I(X;Y|Z)$9

where $\partial B$00 is an AdS$\partial B$01 extremal surface ending on $\partial B$02 (Liu et al., 13 Feb 2026).

For a single interval of length $\partial B$03 on $\partial B$04, the entropy takes the form

$\partial B$05

The two terms have distinct interpretations. The geometric term $\partial B$06 comes from the QES area on the brane, while the term $\partial B$07 is the bulk-QFT contribution. The leading linear divergence is attributed to brane-bath entanglement near the AdS$\partial B$08 boundary.

For two disjoint intervals $\partial B$09 and $\partial B$10 of lengths $\partial B$11 and separation $\partial B$12, there are disconnected and connected QES phases: $\partial B$13

$\partial B$14

Because the UV divergences cancel, the BMI in the connected phase is finite and splits naturally as

$\partial B$15

with

$\partial B$16

A central numerical result is that $\partial B$17 always exceeds the total BMI in the connected phase: $\partial B$18 The interpretation given is geometric and entropic. When the quantum entanglement wedges merge, $\partial B$19 contains more bulk degrees of freedom than $\partial B$20, so the bulk entropy subtracted in $\partial B$21 is larger in the connected phase, producing a negative finite correction. The phase transition occurs at the critical separation $\partial B$22 where

$\partial B$23

Numerically, $\partial B$24 decreases approximately linearly with $\partial B$25 in the connected phase and vanishes beyond $\partial B$26, while $\partial B$27 increases as $\partial B$28 decreases and $\partial B$29 grows.

The surfaces are constructed numerically with Surface Evolver, anchored at $\partial B$30 and on the brane, with both connected and disconnected candidates explicitly compared. The same sign structure is reproduced in a random tensor network toy model. There the large bond-dimension formula

$\partial B$31

shows that the geometric minimal-cut term is non-negative, whereas a highly mixed bulk state with volume-law entropy yields a non-positive bulk contribution when wedges merge. In this setting BMI cleanly separates QES geometry from bulk-matter corrections.

5. BMI in Gaussian statistical field theory

For a real free massive scalar field in a bounded region $\partial B$32,

$\partial B$33

the covariance operator is

$\partial B$34

where $\partial B$35 specifies the boundary condition, such as Dirichlet, Neumann, periodic, local Robin, or free. The corresponding Green kernel is $\partial B$36. In this setting mutual information is defined as a relative entropy between Gaussian measures: $\partial B$37 where $\partial B$38 is the centered Gaussian field restricted to $\partial B$39, and $\partial B$40 are its marginals. The exact Gaussian relative-entropy formula is

$\partial B$41

whenever the measures are equivalent (Schröfl et al., 2023).

The basic theorem is that if $\partial B$42 and $\partial B$43 are bounded open sets with $\partial B$44, then $\partial B$45 is finite in any spatial dimension $\partial B$46. The exact one-dimensional result for intervals $\partial B$47 and $\partial B$48, with separation $\partial B$49, is

$\partial B$50

At large separation this behaves as $\partial B$51, while as $\partial B$52 at fixed $\partial B$53 it diverges. By contrast, for touching regions the mutual information diverges. The paper proves that for touching open $\partial B$54-rectangles sharing a planar face, $\partial B$55 in any $\partial B$56. The mechanism is not merely a short-distance singularity of $\partial B$57; it is a mutual-singularity statement originating in the mismatch between the form domain for the joined region and the direct-sum form domain for the separated regions.

The conceptual explanation is the Markov property of the free Euclidean scalar field. On a region $\partial B$58,

$\partial B$59

where $\partial B$60 is the Dirichlet field in $\partial B$61 and $\partial B$62 is a Gaussian field on the boundary. For separated regions, the Dirichlet parts factorize, so the mutual information is governed by boundary degrees of freedom. The source overview presents this as the formal identity

$\partial B$63

which is the field-theoretic origin of the area law. In this usage, BMI is not an extra observable added to the theory; it is the statement that mutual information between separated regions is effectively boundary-supported.

Boundary conditions then act directly on BMI through the covariance ordering

$\partial B$64

and, for rectangular $\partial B$65,

$\partial B$66

A standard inference consistent with the source is that Neumann boundary conditions enhance correlations relative to free space, while Dirichlet suppresses them. Likewise, the large-separation form

$\partial B$67

is presented as an inference from the paper’s area-law mechanism and Gaussian perturbation theory, rather than as a proved theorem. The robust statements are the finiteness for separated regions, divergence for touching regions, exact one-dimensional formula, and the Markov reduction of mutual information to boundary data.

6. BMI harvested by accelerated detectors near a reflecting boundary

In the detector-harvesting literature, BMI means the mutual information harvested by two Unruh–DeWitt detectors in the presence of a boundary. The setup consists of two identical detectors coupled locally to a massless scalar field in Minkowski spacetime with a perfectly reflecting planar boundary at $\partial B$68. The boundary condition in the Wightman function is Dirichlet, the detectors share a common rotational axis, have equal trajectory radii, equal proper accelerations, and move on coaxial circular trajectories parallel to the $\partial B$69-plane. Their $\partial B$70-coordinates differ by $\partial B$71, which sets the interdetector separation. A Gaussian switching function $\partial B$72 controls the interaction duration, and the calculation is performed perturbatively to $\partial B$73 (Quan et al., 14 Apr 2026).

At leading order, the two-detector reduced state in the computational basis $\partial B$74 is

$\partial B$75

where $\partial B$76 are local excitation probabilities and $\partial B$77 is the nonlocal cross correlator. The leading-order mutual information is

$\partial B$78

with

$\partial B$79

Thus, at $\partial B$80, the mutual information is fully determined by $\partial B$81, and $\partial B$82. The operational BMI is the harvested mutual information in the presence of the boundary, $\partial B$83, and the boundary-induced change is $\partial B$84.

The circular worldlines are

$\partial B$85

$\partial B$86

with

$\partial B$87

The reflecting plane modifies the Wightman function by the method of images,

$\partial B$88

so the boundary contribution enters both the local terms and the cross correlator. In the synchronous case,

$\partial B$89

with

$\partial B$90

The $\partial B$91 factor is the source of oscillatory behavior under fast rotation.

The reported parameter dependence is structured and nonmonotonic. As the interdetector separation increases, the mutual information may exhibit oscillatory behavior at large acceleration and small radius. For fixed radius, a larger acceleration leads to a larger peak value of the mutual information. Near the boundary, the mutual information may oscillate and the maximum can be obtained. As the acceleration increases, the mutual information at small interdetector separation first increases and then decreases, while at intermediate separation it may oscillate with acceleration. For not large separation, when acceleration is large and the radius is small, increasing the energy gap causes the mutual information first to decrease, then to oscillate, and finally to go to zero. The paper identifies the combination of large acceleration and small radius with fast rotation, which strongly modulates vacuum fluctuations and intensifies boundary-induced oscillations through coherent superposition of direct and reflected contributions.

Several threshold values are reported. For small $\partial B$92 and $\partial B$93, oscillations in $\partial B$94 appear when $\partial B$95; for larger $\partial B$96, they appear beyond the higher threshold $\partial B$97. The detector-boundary distance also separates regimes, with a critical value $\partial B$98 in the scans shown: below it, boundary effects are strong and oscillations are pronounced; above it, $\partial B$99 approaches $X,Y$00. In this literature, BMI quantifies how a physical boundary reshapes the sampled vacuum fluctuations and thereby redistributes both classical and quantum correlations between localized probes.

7. Comparative structure, misconceptions, and limitations

The shared acronym conceals substantial differences of ontology. In lattice renormalization, BMI is $X,Y$01 and the boundary is a shielding neighborhood. In AdS$X,Y$02 holography, BMI is $X,Y$03 for disjoint boundary domains. In double holography, it is $X,Y$04 for intervals on a defect line coupled to a bath. In Gaussian field theory, BMI is the boundary-supported content of $X,Y$05 implied by the Markov property. In detector harvesting, BMI is $X,Y$06, the mutual information obtained after boundary-induced image effects are included in the field correlators (Bertoni et al., 2021, Fonda et al., 2014, Liu et al., 13 Feb 2026, Schröfl et al., 2023, Quan et al., 14 Apr 2026).

Despite these differences, several structural themes recur. First, locality or shielding converts a nominally bulk quantity into a boundary-mediated one. This is explicit in the Gibbs-Markov shielding argument for $X,Y$07, in the formal identity $X,Y$08 for Gaussian fields, and in the QES or RT characterization of holographic BMI through surfaces anchored on boundaries. Second, many BMI observables exhibit sharp configurational changes: connected versus disconnected RT surfaces in AdS$X,Y$09, merged versus disconnected quantum entanglement wedges in double holography, and finite versus divergent mutual information for separated versus touching regions in Gaussian field theory. Third, the specific boundary condition matters: it appears as the choice of shielding set $X,Y$10, as the anchoring and regularization of holographic surfaces, as Dirichlet versus Neumann behavior in Gaussian fields, and as the image term in detector Wightman functions.

The limitations are equally context-dependent. The renormalization results rely on classical local Hamiltonians and Markov-network arguments and do not directly extend to quantum systems with entanglement. The AdS$X,Y$11 results are classical RT results, and the sharp transition is smoothed by quantum corrections; cornered regions also introduce regularization-sensitive $X,Y$12 terms. The double-holography study is carried out for a specific AdS$X,Y$13/AdS$X,Y$14 braneworld, in the semiclassical subcritical regime, with zero-temperature intervals on a straight defect. The Gaussian field-theory area-law interpretation is strongly supported by the Markov decomposition, but the source explicitly treats some large-distance BMI asymptotics as inference rather than theorem. The detector analysis is perturbative to $X,Y$15, assumes Gaussian switching and a perfect Dirichlet reflector, and attributes the strongest oscillatory effects to fast rotation.

Taken together, these literatures show that Boundary Mutual Information is best understood as a family of boundary-conditioned mutual informations rather than a single canonical quantity. Its technical role ranges from selecting coarse-graining maps that suppress long-range couplings, to diagnosing entanglement-wedge connectivity, to identifying boundary-supported correlations in continuum fields, to quantifying interference-enhanced harvesting in the presence of reflecting surfaces.