---
title: Spherical Boltzmann Machine
url: https://www.emergentmind.com/topics/spherical-boltzmann-machine-sbm
type: topic
---

# Spherical Boltzmann Machine

Spherical Boltzmann machines are Boltzmann-machine models in which continuous variables are subject to a spherical constraint or spherical prior, so that configurations lie on a fixed-radius sphere or thin spherical shell rather than on a discrete hypercube. In the recent literature, the term covers several related but non-identical constructions: fully visible quadratic energy models on $\|\boldsymbol{\sigma}\|^2=N$ [2604.17936], undirected energy-based models over $\mathbf x\in\mathcal S_N=\{\mathbf x\in\mathbb R^N:\|\mathbf x\|^2=N\}$ with symmetric coupling matrix $W$ [2605.09031], and restricted architectures in which a hidden layer is spherical while visible units are Gaussian [1910.14544]. A complementary line of work studies spherical and Gaussian priors in fully visible and restricted Boltzmann machines and proves a Legendre equivalence at the level of thermodynamic free energies [1910.14559]. Across these formulations, the spherical constraint is the source of both tractability and distinctive collective phenomena, notably condensation onto dominant spectral modes, explicit spectral phase boundaries, and closed-form or asymptotically exact analyses of learning and generation.

## 1. Terminology and model family

The expression “spherical Boltzmann machine” does not denote a single universally fixed architecture. In the fully visible formulation developed as an analytically solvable benchmark for ensemble learning, the model consists of continuous spins $\boldsymbol{\sigma}\in\mathbb R^N$ constrained by $\|\boldsymbol{\sigma}\|^2=N$ and governed by the quadratic energy
\[
E(\boldsymbol{\sigma};\boldsymbol{J})=-\frac{1}{2}\boldsymbol{\sigma}^{\top}\boldsymbol{J}\boldsymbol{\sigma},
\]
with $\boldsymbol J=\boldsymbol J^\top$ symmetric [2604.17936]. A closely related formulation writes the visible state as $\mathbf x\in\mathcal S_N$ and the energy as
\[
E(\mathbf x;W)=-\frac12\,\mathbf x^\top W\mathbf x,\qquad W\in \mathrm{Sym}_N,
\]
so that the induced Gibbs law on the sphere is a Bingham distribution [2605.09031].

Restricted variants impose spherical structure only on part of the system. In the Gaussian-spherical restricted Boltzmann machine, visible variables $\{s_i\}$ are real-valued with Gaussian priors, while hidden variables $\{\sigma_j\}$ are constrained by
\[
\sum_{j=1}^{N_h}\sigma_j^2=\bar{\sigma}^{\,2}L,\qquad L=\sqrt{N_vN_h},
\]
and interact through a bipartite weight matrix $W=(w_{ij})$ [1910.14544]. Another formulation treats both fully visible Hopfield-type models and bipartite RBMs with spherical priors, uniform on spheres $S_N(R)\subset\mathbb R^N$ of radius $R\sqrt N$, or on thin shells used in proofs [1910.14559].

A terminological caveat is necessary. In “Understanding Boltzmann Machine and Deep Learning via A Confident Information First Principle,” the acronym SBM is used for “single-layer Boltzmann machine without hidden units,” not for “spherical Boltzmann machine” [1302.3931]. That usage refers to a binary visible-only pairwise Boltzmann machine and belongs to a different line of work. This suggests that the acronym “SBM” is historically overloaded, and precise interpretation depends on context.

## 2. Canonical mathematical formulations

In the fully visible spherical model, the Gibbs distribution takes the form
\[
P_{\sigma\mid J}(\boldsymbol{\sigma}\mid\boldsymbol{J}) =\frac{1}{Z(\boldsymbol{J})}\, \delta(\|\boldsymbol{\sigma}\|^2-N)\, e^{\frac{1}{2T'}\boldsymbol{\sigma}^\top \boldsymbol{J}\boldsymbol{\sigma},
\]
with partition function
\[
Z(\boldsymbol{J}) = \int d\boldsymbol{\sigma}\; \delta(\|\boldsymbol{\sigma}\|^2-N)\, e^{\frac12 \boldsymbol{\sigma}^{\top}\boldsymbol{J}\boldsymbol{\sigma}}
\]
for $T'=1$ [2604.17936]. At large $N$, the model is controlled by the spectrum of $\boldsymbol J$ through a Lagrange multiplier $\mu$ solving
\[
1=\frac{1}{N}\sum_{j=1}^N\frac{1}{\mu-\lambda_j}, \qquad \mu>\lambda_{\max}.
\]
This spectral reduction is one of the central reasons the model is analytically manageable.

The closely related EBM formulation uses
\[
P(\mathbf x\mid W) = \frac{\exp\!\left(\frac{\beta}{2}\mathbf x^\top W\mathbf x\right)} {Z_\beta(W)} \mathbf 1(\mathbf x\in\mathcal S_N),
\]
with
\[
Z_\beta(W)=\int_{\mathcal S_N} \exp\!\left(\frac{\beta}{2}\mathbf x^\top W\mathbf x\right)\,d\mathbf x.
\]
Because the support is the sphere, the law is not Gaussian even though the energy is quadratic [2605.09031]. The paper emphasizes that the spherical constraint induces condensation transitions absent from the unconstrained Gaussian analogue.

In the Gaussian-spherical RBM, the energy is
\[
E(\mathbf{s},\boldsymbol{\sigma}) = - \sum_{i,j} w_{ij} s_i \sigma_j + \sum_i \frac{s_i^2}{2} - \sum_i \eta_i s_i - \sum_j \theta_j \sigma_j,
\]
and the joint distribution is
\[
p(\mathbf{s},\boldsymbol{\sigma}) = \frac{1}{Z} e^{-E(\mathbf{s},\boldsymbol{\sigma})} \, \delta\!\left(\sum_{j=1}^{N_h}\sigma_j^2-\bar{\sigma}^{\,2}L\right).
\]
Here the visibles remain unconstrained Gaussian variables, whereas the hidden layer does not conditionally factorize because the global $\delta$-constraint couples all hidden components [1910.14544].

The spherical-prior framework of Genovese and Tantari treats a layer of size $N$ under the uniform measure on $S_N(R)$ or, technically, the shell
\[
S_{N,R,\varepsilon} := \left\{ x\in\mathbb R^N: R-\varepsilon<\frac{|x|^2}{N}<R+\varepsilon \right\},
\]
and studies both the fully visible Hopfield Hamiltonian and the restricted Hamiltonian
\[
H^{RBM}_{N_1,N_2} := -\frac{1}{\sqrt{N_1+N_2}} \sum_{j=1}^{N_2}\sum_{i=1}^{N_1}\xi_{ij}x_i y_j
\]
in the thermodynamic limit [1910.14559]. In that setting, spherical constraints may be imposed on one or both layers.

## 3. Solvability, spectral reduction, and thermodynamic structure

A recurring feature of spherical Boltzmann machines is that the spherical constraint converts the partition function into a form dominated by spectral quantities. In the Gaussian-spherical RBM, singular value decomposition
\[
w_{ij}=\sum_{\alpha} w_\alpha\, u_i^\alpha v_j^\alpha
\]
diagonalizes the interaction, and after introducing a Fourier/Laplace representation of the spherical constraint, the partition function reduces to a one-dimensional contour integral
\[
Z = \frac{1}{2\pi i} \int_{a-i\infty}^{a+i\infty} dz\, e^{L\phi(z)},
\]
which can be analyzed by saddle-point methods [1910.14544]. The thermodynamics then depends only on the spectrum of $WW^\top$, especially edge behavior, isolated top modes, and spectral degeneracy.

In the fully visible spherical-prior models, free energies become explicit convex variational problems. For the spherical Hopfield model,
\[
\lim_{N_1,N_2\to\infty} A_{N_1,N_2}
= a\, \min_{2q>\beta(1-a)\lambda_+} \left[ qR^2 -\frac12\int P_{MP}(\lambda;a)\log(2q-\beta(1-a)\lambda)\,d\lambda -\log R -\frac12 \right],
\]
while for the doubly spherical RBM,
\[
\lim_{N_1,N_2\to\infty}A_{N_1,N_2}
\]
is given by a convex minimization formula involving the Marchenko–Pastur law and two Lagrange multipliers [1910.14559]. The paper stresses that the doubly spherical RBM yields a fully convex minimization principle.

In the 2026 visible-only SBM theory, the evidence over coupling matrices has a Coulomb-gas structure. Writing $W=\frac1N V\Lambda V^\top$, the posterior weight becomes
\[
\mathcal L_\Lambda(\Lambda;C)= -\frac N4\gamma\eta \Trace\Lambda^2 -K\eta\ln Z(\Lambda) +\ln J(\Lambda;C) +\sum_{i<j}\ln|\lambda_i-\lambda_j|,
\]
and the leading $O(N^2)$ contribution yields a Wigner semicircle bulk with half-width
\[
\sigma=\frac1{\sqrt{\gamma\eta}}.
\]
The bulk density is
\[
\rho(\lambda)\to \frac{\sqrt{4\sigma^2-\lambda^2}}{2\pi\sigma^2}\mathbf 1_{|\lambda|\le 2\sigma},
\]
with Stieltjes transform
\[
G(z)= \frac1{2\sigma}\left(\frac z\sigma-\sqrt{\frac{z^2}{\sigma^2}-4}\right)
\]
[2605.09031]. This spectral characterization controls alignment, condensation, and evidence asymptotics.

The replica treatment of SBM ensembles reaches a parallel conclusion from a different direction. There, the replicated free entropy closes in terms of a replica overlap matrix $Q$ and magnetizations along data modes $M$, producing a large-$N$ action
\[
\Phi(\boldsymbol Q,\boldsymbol M)
= \frac{1}{4\gamma T}\mathrm{tr}(\boldsymbol X^2)
+\frac{T}{4\gamma}\mathrm{tr}(\boldsymbol Q^2)
+\frac{1}{2\gamma}\mathrm{tr}(\boldsymbol M\boldsymbol D_\chi \boldsymbol M^\top)
+\frac12 \mathrm{tr}\ln(\boldsymbol Q-\boldsymbol M\boldsymbol M^\top)
-\frac12 \mathrm{tr}\big(\boldsymbol D_\mu(\boldsymbol Q-\boldsymbol I_n)\big),
\]
and the saddle-point equations can be solved analytically under a rotationally invariant ansatz [2604.17936]. A plausible implication is that the spherical constraint does not merely regularize the state space; it reorganizes the entire inference problem into a tractable spectral theory.

## 4. Learning formulations and parameter-space inference

Learning in spherical Boltzmann machines has been formalized in several distinct ways. One is a tempered ensemble over interaction matrices. Given data $\mathcal D=\{\boldsymbol{\xi}_k\}_{k=1}^K$ and a Gaussian/Wigner prior
\[
P_G(\boldsymbol{J})\propto e^{-\frac{N\gamma}{4}\mathrm{tr}(\boldsymbol{J}^2)},
\]
the model ensemble is
\[
P_T(\boldsymbol{J}\mid \mathcal D)\propto \left[ P_G(\boldsymbol{J}) \prod_{k=1}^K \frac{e^{-E(\boldsymbol{\xi}_k;\boldsymbol J)/T'}}{Z(\boldsymbol J)} \right]^{1/T}.
\]
The paper emphasizes that, unless $T=1$, this is not a Bayesian posterior because $Z(\boldsymbol J)^{1/T}\neq Z(\boldsymbol J/T)$, so $T$ is a genuine learning temperature [2604.17936]. In that framework, ensemble learning means averaging predictions over $P_T(\boldsymbol J\mid\mathcal D)$ rather than selecting a single minimizer.

A related but differently parametrized formulation defines a tempered posterior over $W$ through
\[
P_\eta(W\mid\mathcal D)=\frac{e^{\mathcal L(W;C)}}{Y(C)},
\qquad
\mathcal L(W;C) = \frac N2 \eta\,\Trace(CW) -\frac N4 \gamma\eta\,\Trace W^2 -K\eta \ln Z(W),
\]
with evidence
\[
Y(C)=\int_{W\in\mathrm{Sym}_N} e^{\mathcal L(W;C)}\,dW.
\]
Here $\eta=1$ corresponds to the standard Bayes posterior, $\eta\to\infty$ to the MAP limit, $\gamma>0$ to weight decay strength, and $1/\eta$ to temperature in parameter space [2605.09031]. The evidence acts as a partition function in parameter space and encodes global properties of the trained model.

The same paper also studies exact continuous-time stochastic training dynamics with exact negative phase,
\[
\partial_t W(t) = \frac12\left( C-\frac KN \langle \mathbf x\mathbf x^\top\rangle_{\mathbf x\sim P_{W(t)}} \right) -\frac\gamma2 W(t) +\frac{\Omega(t)}{\sqrt{\eta N}},
\]
and a persistent-MCMC variant,
\[
\partial_tW(t)= \frac12\left( C-\frac KN \mathbf x(t)\mathbf x(t)^\top \right) -\frac\gamma2 W(t) +\frac{\Omega(t)}{\sqrt{\eta N}},
\]
\[
\partial_t\mathbf x(t) = \nu W(t)\mathbf x(t)-\kappa(t)\mathbf x(t)+\sqrt{2\nu}\,\boldsymbol\xi(t),
\]
where $\nu$ is the sampling rate and $\kappa(t)$ enforces the sphere [2605.09031]. This yields an analytically tractable model of training coupled to a non-equilibrated negative phase.

In the Gaussian-spherical RBM, tractability changes the nature of training. Because response functions can be extracted with arbitrary precision, the learning dynamics of singular values and other mode-resolved quantities can be numerically integrated without relying on contrastive divergence or approximate mean field in the analyzed settings [1910.14544]. This supports a mode-by-mode interpretation of training as spectral emergence from a Marchenko–Pastur bulk.

## 5. Phase structure, condensation, and generation

The central physical phenomenon in spherical Boltzmann machines is condensation onto top spectral modes. In the visible-only SBM theory, introducing a Lagrange multiplier $\mu$ for the spherical partition function leads to the saddle equations
\[
1-h^2=G(\mu),\qquad (\lambda_i-\mu)h_i=0,
\]
with
\[
h^2=
\begin{cases}
0,& g_1\ge1,\\
1-g_1,& g_1\le1,
\end{cases}
\qquad
\mu=
\begin{cases}
G^{-1}(1),& g_1\ge1,\\
\lambda_1,& g_1\le1.
\end{cases}
\]
The threshold $g_1=1$ marks the transition between uncondensed and condensed sampling [2605.09031]. The paper explicitly contrasts representation and generation: $u_k>0$ means a mode is represented in $W$, whereas $h>0$ means it is active in generated samples.

That distinction produces a multiphase equilibrium diagram. The theory identifies a fully unaligned, uncondensed regime; an aligned but uncondensed regime in which top eigenvalues detach and align to data yet remain “silent” in generation; a condensed but unaligned regime in which samples condense along a random direction; and condensed, aligned regimes with either edge-pinned or outlier top eigenvalues [2605.09031]. The force balance on outliers clarifies the mechanism: HCIZ alignment pulls eigenvalues away from the bulk, whereas the sample-space partition function pushes them back through a condensation-dependent term.

The Gaussian-spherical RBM exhibits an analogous spectral phase structure. When the spectral edge exponent satisfies $\gamma>0$, the critical variance
\[
\bar\sigma_c^2 = \frac{1}{E_{\max}\left(\delta+\frac{\kappa}{\gamma}\right)}
\]
is finite, and for $\bar\sigma^2>\bar\sigma_c^2$ the top mode may acquire macroscopic occupation, interpreted as the analog of the ferromagnetic phase of the spherical model [1910.14544]. Degenerate top eigenspaces yield ordered phases with $O(n)$ symmetry, and in the doubly degenerate case the partition function admits an exact finite-size treatment that makes the analogy to Bose-Einstein condensation explicit.

The replica theory of SBM ensembles reaches similar conclusions at the level of learned couplings. The number $d$ of condensed data modes is selected by
\[
\frac{\gamma}{\chi_d}<\nu_0(d)<\min\left\{\frac{\gamma}{\chi_{d+1}},\sqrt{\frac{\gamma}{T}}\right\},
\]
yielding phases that are paramagnetic, condensed, or marginal, together with freezing of the negative-replica branch below a critical $n_c$ [2604.17936]. In the rank-one case, these phases map to the deformed-GOE picture with unaligned and aligned branches and edge-pinned marginal behavior.

Several generative consequences follow. Sampling-temperature tuning can rescue a learned but silent aligned mode by increasing $\beta$ so that generation crosses a condensation threshold, whereas lowering $\beta$ can de-condense a wrongly condensed random mode [2605.09031]. The same paper reports double descent of the typical reverse KL as a function of regularization strength $\gamma$, warm and cold posterior effects as a function of posterior temperature $\eta$, and out-of-equilibrium training biases at finite sampling rate $\nu$. This suggests that condensation is not a peripheral feature but the organizing principle behind the model’s generative behavior.

## 6. Relations to Gaussian models, RBMs, and broader Boltzmann-machine theory

A major theoretical result is that spherical and Gaussian or sub-Gaussian Boltzmann machines are related by Legendre transformation at the level of limiting free energy. For the fully visible model,
\[
A_p(\theta,a,\beta) = a\, \sup_{R>0} \left( a^{-1}A_\sigma(R,\beta) -\frac{R^2}{2\theta} +\log R -r(R) -\frac12(\log\theta-1) \right),
\]
and conversely
\[
A_\sigma(R,\beta) = a\, \inf_{\theta>0} \left( \frac{R^2}{2\theta} +a^{-1}A_p(\theta,a,\beta) -\log R +r(R) +\frac12(\log\theta-1) \right)
\]
[1910.14559]. Analogous formulas hold for mixed and doubly restricted models. The rigid radius $R^2$ and Gaussian variance $\theta$ are thus Legendre-conjugate variables in the thermodynamic limit.

This equivalence clarifies the role of the spherical constraint. A Gaussian prior decomposes over spherical shells, and the large-$N$ Laplace principle selects the dominant radius. In that sense, the Gaussian model is a superposition of spherical models over radii, while the spherical model can be viewed as a hard-constrained representative of the same free-energy landscape [1910.14559]. The 2026 solvable EBM paper sharpens the contrast by showing that condensed phases, temperature tuning, warm posterior effects, and the KL spike at the condensation boundary are specific to the spherical constraint because, in the unconstrained Gaussian model, $\ln Z_G$ contributes only $O(1)$ forces on top eigenvalues, whereas in the SBM the partition function contributes an $O(N)$ condensation force [2605.09031].

Restricted spherical models occupy an intermediate position. The Gaussian-spherical RBM preserves the bipartite structure of an RBM but loses hidden-unit conditional independence because the spherical hidden-layer constraint globally couples hidden coordinates [1910.14544]. By contrast, the fully visible SBM has no hidden variables and is closer in spirit to the spherical spin glass or a continuous Hopfield-type model, though the learning problem is cast in terms of inference over the coupling matrix rather than fixed Hebbian retrieval [2604.17936].

The older CIF paper underscores a different relation. Its “SBM” is a binary visible-only pairwise Boltzmann machine obtained by preserving first- and second-order expectation coordinates and setting higher-order natural coordinates to zero [1302.3931]. That model is neither spherical nor continuous. The overlap in acronym therefore reflects nomenclature rather than model continuity.

Taken together, the literature portrays spherical Boltzmann machines as a family of continuous, norm-constrained energy-based models whose analytical accessibility derives from spectral reduction, variational principles, and replica or random-matrix methods. Their main significance lies in providing solvable settings in which learning, evidence, mode alignment, condensation, and generation can all be studied explicitly [2605.09031].

Source: https://www.emergentmind.com/topics/spherical-boltzmann-machine-sbm