---
title: Self-Organizing Survival Manifolds (SOSM)
url: https://www.emergentmind.com/topics/self-organizing-survival-manifolds-sosm
type: topic
---

# Self-Organizing Survival Manifolds (SOSM)

Searching arXiv for the specified papers and closely related context.
Self-Organizing Survival Manifolds (SOSM) denotes a theoretical framework in which survival is modeled not as a supervised target but as an emergent property of latent biological geometry. In this formulation, biological samples in a high-dimensional observed space are assumed to lie near a smooth low-dimensional Riemannian manifold, biological progression is represented by trajectories on that manifold, and prognosis is associated with low-curvature, geodesic-like flow stability rather than externally supplied labels. A closely related line of work extends the geometric treatment of survival into viability-constrained dynamical systems, where state space is partitioned by mortality, ordering, and collapse manifolds into regions of qualitatively distinct survival outcomes; although that framework does not use the term SOSM, it supplies an explicit formalism for survival-organizing manifolds in single- and multi-agent ODE models [2508.06539, 2605.16753].

## 1. Conceptual foundations

SOSM is presented as a rejection of the premise that survival must be learned primarily from curated event labels, censoring-aware objectives, and fixed covariate-to-outcome mappings. Its central thesis is that survival should be treated as a geometric consequence of latent biological organization rather than as an externally annotated prediction target. The framework therefore opposes the standard supervised orientation of Cox proportional hazards, regularized Cox extensions, and deep survival models such as DeepSurv and DeepHit, arguing that such methods depend on outcome annotations that are often incomplete, noisy, delayed, cohort-biased, or sparse in large biological repositories [2508.06539].

Within this view, prognosis is presumed to be already encoded in the geometry of biological state space. Good prognosis corresponds to smooth, low-curvature, geodesic-like flows shaped by internal biological constraints, whereas poor prognosis corresponds to high-curvature, unstable, entropically costly deviations. The biological motivation is strongest in settings such as multi-omics analysis, where the data are high-dimensional and structured, latent progression may exist even when outcomes are weakly observed, and regulatory, metabolic, evolutionary, and homeostatic constraints plausibly induce coherent low-dimensional organization.

A common misconception is to interpret SOSM as merely a manifold-based variant of conventional risk prediction. The formal emphasis is different. SOSM does not begin by defining a hazard, partial likelihood, or direct time-to-event estimator; instead, it seeks a latent organization in which survival structure self-organizes through curvature minimization and flow stability. This suggests that the framework is best understood as a geometric theory of prognostic organization rather than as a routine replacement for standard supervised survival pipelines.

## 2. Geometric and mathematical formulation

The observed biological state space is taken to be $\mathcal{X} \subseteq \mathbb{R}^D$, with samples $\{x_i\}_{i=1}^N \subset \mathcal{X}$. The latent state space is a smooth Riemannian manifold $\mathcal{M}$ of intrinsic dimension $d \ll D$, and learned embeddings are written $z_i \in \mathcal{Z} \subset \mathbb{R}^d$. In continuous form, survival is represented by a scalar field $t(z)$ on $\mathcal{M}$, and biological progression is modeled by smooth curves $\gamma:[0,1]\to\mathcal{M}$, typically parameterized by arc length [2508.06539].

The core mathematical object is the Survival Energy Functional,
$$
E_{\text{SOSM}}(\gamma)=\int_0^1 \kappa(\gamma(s))^2\,ds,
$$
where geodesic curvature is defined by
$$
\kappa(s)=\left\|\frac{D^2\gamma}{ds^2}\right\|
$$
or equivalently $\kappa(s)=\|\nabla_{\dot\gamma}\dot\gamma\|$. In intrinsic coordinates, the squared curvature is written
$$
\kappa^2=g_{ij}\left(\frac{D\dot\gamma^i}{ds}\right)\left(\frac{D\dot\gamma^j}{ds}\right).
$$
The functional penalizes bending away from geodesic motion. Low energy therefore denotes smooth trajectory geometry, geodesic alignment, and prognostic stability; high energy denotes strong bending and unstable progression.

The discrete formulation introduces a survival-similarity kernel
$$
w(t_i,t_j)=\exp\left(-\frac{(t_i-t_j)^2}{2\sigma^2}\right),
$$
and a sample-level objective
$$
\mathcal{L}_{\text{SOSM}}=\sum_{i,j} w(t_i,t_j)\|\nabla^2(z_i-z_j)\|^2.
$$
The corresponding second-order path discretization is
$$
\Delta^2 z_i = z_{i+1}-2z_i+z_{i-1},
$$
with
$$
E_{\text{discrete}}=\sum_{i=2}^{N-1} w(t_i,t_{i+1})\|\Delta^2 z_i\|^2,
$$
and the appendix states the approximation
$$
\|\Delta^2 z_i\|^2 \approx (\delta s)^4\kappa(s_i)^2.
$$
In the dense-sample limit, the discrete loss is said to converge to a continuous manifold integral weighted by the embedding density $p(z)$.

This formal setup is accompanied by several caveats. The paper uses survival times $\{t_i\}_{i=1}^N$ in some formulations but emphasizes that they may be unavailable during training in the unsupervised setting. It repeatedly invokes “internal biological constraints,” yet it does not introduce a separate formal variable or operator encoding those constraints. Their role is conceptual: they justify the manifold hypothesis and the preference for low-curvature flow.

## 3. Emergence, stability, and convergence

The framework’s main theoretical claim is that minimizing the SOSM objective induces a latent embedding in which survival time $t(z)$ becomes a smooth, monotonic function along geodesic flows. This is the content of Theorem 1, “Survival Gradient Emergence,” which states that under manifold regularity and sufficient sample density, minimizing $\mathcal{L}_{\text{SOSM}}$ produces survival-aligned trajectories on $\mathcal{M}$. The paper’s proof strategy is based on the idea that penalizing local curvature among survival-similar points makes local reversals or disorder in survival ordering energetically unfavorable [2508.06539].

The dynamical content is extended through a variational expression. The appendix gives a functional derivative form
$$
\frac{\delta E_w}{\delta \gamma}
=
-2\left(\kappa''(s)+R[\gamma'(s),\gamma''(s)]\right)
+
\frac{\epsilon^2}{\sigma^2}(t'(s))^2\kappa(s),
$$
which is interpreted as combining geometric smoothing with an additional penalty where the survival gradient is large. The metric itself is then proposed to evolve under
$$
\frac{\partial g_{ij}}{\partial \tau}
=
-\alpha \cdot \frac{\delta \mathcal{L}_{\text{SOSM}}}{\delta g^{ij}},
$$
and, under approximation, this is linked to Ricci flow through
$$
\frac{\partial g_{ij}}{\partial \tau}\approx -2R_{ij}.
$$

Theorem 2, “Trajectory Stability,” states that if $\gamma$ is a geodesic path aligned with survival progression, then small perturbations $\gamma'=\gamma+\epsilon v$ increase the survival energy to second order in $\epsilon$. The main-text expansion is
$$
E_{\text{SOSM}}(\gamma')
=
E_{\text{SOSM}}(\gamma)
+
\epsilon^2\int_0^1 \|\nabla^2 v(s)\|^2\,ds
+
o(\epsilon^2),
$$
and the appendix gives the second variation
$$
\delta^2 E[\gamma;v]
=
\int_0^1
\left(
\|\nabla_s^2 v(s)\|^2
-
\langle R(\dot\gamma(s),v(s))\dot\gamma(s),v(s)\rangle
\right)ds.
$$
In Euclidean space, positivity is immediate:
$$
\delta^2 E[\gamma;v]=\int_0^1 \|\ddot v(s)\|^2 ds.
$$
This establishes a formal connection between prognostic quality and perturbative stability.

Theorem 3, “Manifold Convergence,” asserts that under a continuous curvature flow driven by minimization of $\mathcal{L}_{\text{SOSM}}$, the manifold converges toward a locally survival-optimized geometry. The supporting assumptions include bounded curvature, injectivity radius conditions, and smooth metric evolution. The paper is explicit, however, that this is not a fully rigorous new Ricci-flow theorem; it is an analogy-supported convergence argument grounded in a simplified geometric approximation. That limitation is central to the present status of SOSM as a formal paradigm rather than a finished differential-geometric theory.

## 4. Biological interpretation and relation to prior survival modeling

SOSM assigns biological meaning to geometric structure. Health is interpreted as a low-curvature regime in which trajectories are smooth, flows are coherent, entropy production is controlled, and state transitions are stable and efficient. Disease is interpreted as a curvature anomaly: a deviation from geodesic organization marked by abrupt state deviations, branching or fragmented trajectories, rising energetic cost, and instability. Aging is framed as accumulated geometric distortion, in which regulatory damage and molecular degradation deform the manifold, degrade geodesic survival structure, and raise mortality risk. Death is described as manifold collapse, a geometric phase transition or singularity-like event in which curvature distortions exceed the system’s capacity to preserve coherent survival flow [2508.06539].

The paper also draws links to thermodynamic efficiency, entropy flow, Ricci curvature, and optimal transport. These links do not all have the same status. The Ricci-flow relation is semi-formal and partially derived. The optimal transport interpretation introduces a curvature-weighted cost
$$
c(z,z')=\|\nabla^2(z-z')\|^2
$$
and an objective
$$
\mathrm{OT}_c(p,q)=\inf_{\gamma\in\Pi(p,q)}
\int_{\mathcal{M}\times\mathcal{M}} c(z,z')\,d\gamma(z,z'),
$$
with the suggestion that $\mathcal{L}_{\text{SOSM}}$ may be interpreted as a curvature-weighted transport quantity. By contrast, the thermodynamic and entropy relations are heuristic and conceptual rather than microscopically derived.

Relative to prior survival modeling, SOSM diverges from both classical survival analysis and supervised deep survival models. Traditional methods estimate quantities such as hazard $h(t\mid x)$, survival function $S(t\mid x)$, or risk score $r(x)$ under censoring-aware objectives. SOSM does not define a hazard function, partial likelihood, or censoring likelihood. It instead proposes that prognosis is represented by position on a survival-aligned trajectory, local curvature energy, flow stability, and manifold phase organization. A frequent misunderstanding is therefore to read SOSM as a direct hazard estimator. The formalism does not support that interpretation in the standard sense.

## 5. Relation to viability space decomposition

A distinct but closely related framework is viability space decomposition, which studies ODE models of agents with viability constraints. In that setting, the dynamics are only defined while the agent remains inside a bounded viability region, death occurs when trajectories reach the viability boundary with outward velocity, and the state space is partitioned into asymptotically viable and transiently viable sets. The central claim is that traditional attractors, basins, and separatrices are insufficient because trajectories may die before reaching asymptotic regimes, classical basins can be intercepted by the viability boundary, and neighboring initial conditions can differ not only by final attractor but by which viability constraint is violated, when it is violated, and in multi-agent systems which agent dies first [2605.16753].

Viability space decomposition introduces several manifold classes that organize these outcomes. Stable saddle manifolds remain relevant but are reinterpreted through the lens of viability. Mortality manifolds arise from tangency points on the viability boundary and typically form codimension-1 separators between robust asymptotically viable states and robust transiently viable states, or between different death modes. Ordering manifolds arise from intersections of guard conditions and separate outcomes distinguished by simultaneous versus alternative death conditions, or by which agent dies first. Collapse manifolds are specific to the multi-agent hybrid setting; they propagate lower-dimensional organizing structures backward through inverse collapse maps, thereby separating higher-dimensional initial conditions that share the same immediate death event but differ in the later fate of survivors.

The resulting global object is called a viability portrait: a complete decomposition of viability-constrained state space into robust regions of qualitatively similar survival outcomes. This supplies a mathematically explicit version of survival-organizing manifolds in dynamical systems. A plausible implication is that SOSM and viability space decomposition address complementary layers of the same general problem. SOSM treats survival as emergent geometry on a latent biological manifold, whereas viability space decomposition treats survival as a partition of viability-constrained state space by mortality, ordering, and collapse manifolds. The first is a theory of prognostic organization on latent manifolds; the second is a theory of existential outcome partitioning in explicit ODE and hybrid systems.

## 6. Operationalization, limitations, and research directions

SOSM is primarily theoretical and conceptual. It contains no experiments, no implementation section, and no empirical benchmark. To operationalize it on real datasets, the paper states that one would need a manifold-learning architecture mapping $x_i\mapsto z_i$, a discrete curvature estimator such as graph Laplacians or second-order finite differences, a survival-similarity proxy when labels are absent, an optimization procedure minimizing a sample-level curvature loss, longitudinal or neighborhood structure for trajectory inference, and a downstream validation strategy comparing learned geometry to survival outcomes even if those outcomes are not used in training [2508.06539].

Several assumptions are explicitly difficult. The smooth low-curvature manifold assumption may fail in severe pathology. Constructing $w(t_i,t_j)$ or a suitable proxy without supervision is nontrivial and may introduce bias. The link from molecular energetics to latent manifold curvature remains incomplete. Biological systems may exhibit discontinuities, singularities, topological changes, or strong stochasticity that lie outside the smooth-geometric regime. For these reasons, a common misconception—that SOSM is already a mature empirical method—should be rejected. Its current status is that of a formal framework seeking algorithmic realization.

The related viability-space program is similarly candid about computational limits. Although its definitions are dimension-agnostic, numerical manifold computation remains difficult in high dimensions, and in multi-agent settings the number of hybrid-automaton nodes scales as $2^{|a|}$. Future directions named there include novel bifurcation theory for viability portraits, chaotic dynamics and possible fractal survival boundaries, scalable algorithms, symmetry reduction in homogeneous multi-agent populations, and stochastic and nonautonomous generalizations [2605.16753].

Taken together, these lines of work define a broader research agenda in which survival is treated as a geometric and dynamical property rather than as a purely supervised label. SOSM contributes the low-curvature, survival-aligned manifold perspective; viability space decomposition contributes an explicit manifold-based partition of survival and death outcomes under viability constraints. Their combined significance lies in shifting survival analysis toward geometry, flow, and constraint-induced structure.

Source: https://www.emergentmind.com/topics/self-organizing-survival-manifolds-sosm