Why the Most Important Space in Quantitative Science Has No Name Until Now
Gregory Caldwell Beier
Susarb LLC, Cambridge, Massachusetts
March 2026
The covariance matrices underlying portfolio theory, risk management, and machine learning are composed predominantly of quantities whose statistical properties are generated by human decisions rather than physical law. This paper names the space of such quantities - Beierspace - and establishes its formal structure. Eigenspace quantities have covariance determined by physics, estimable from historical data with bounded error. Beierspace quantities have covariance that is endogenous to the observation and coordination behavior of the agents who participate in them. At any single moment, the eigendecomposition cannot distinguish the two. The distinction reveals itself dynamically: eigenspace components change when the physical world changes, while Beierspace components change whenever the decisions constituting them change. Three consequences follow for any continuous, strictly decreasing drift function. First, a phase transition at the stability threshold S = 1, above which the system self-stabilizes and below which it self-destabilizes. Second, bistability - two stable equilibria (functioning and collapsed) separated by a saddle-node bifurcation. Third, informational impossibility - a system in the collapse trap cannot determine from its own data whether the collapse is recoverable, because the parameters distinguishing the regimes are not identifiable from the degraded data the collapsed system produces. Autonomous AI agents amplify every term in the stability condition simultaneously, increasing learning capacity L while potentially decreasing disclosure rate D. The product LD determines stability, not either term alone. Four falsifiable predictions are stated, each incompatible with the standard assumption that covariance instability is exogenous.
Keywords: Beierspace, eigenvalue instability, agentic AI, reflexivity, covariance non-stationarity, ontological partition
Science distinguishes quantities by what generates them. Mass is generated by physics. Temperature is generated by thermodynamics. Orbital period is generated by gravity. The statistical properties of these quantities - their distributions, covariances, spectral decompositions - reflect the laws that produce them. When we compute the eigendecomposition of a covariance matrix composed of such quantities, the resulting structure is as stable as the physics it reflects. Change requires physical force.
Call this eigenspace.
Now consider a different class of quantities. The price of a stock. The spread on a credit default swap. A sovereign credit rating. The interest rate on a thirty-year mortgage. The exchange rate between the dollar and the yen. These quantities exist because humans collectively sustain them. Their statistical properties - distributions, covariances, spectral decompositions - reflect the current configuration of human decisions. Change requires only that the decisions change. No physical trigger is necessary because the quantities are constituted by beliefs about beliefs, not by physical law.
Call this Beierspace.
At any single moment, eigenspace and Beierspace quantities look identical under the spectral decomposition. A covariance matrix does not carry a label saying which of its entries are physically determined and which are human-decision-determined. The eigendecomposition cannot distinguish them. The difference reveals itself only dynamically: eigenspace components change when the physical world changes, while Beierspace components change whenever the decisions that constitute them change.
This distinction is not a metaphor. It is a structural property of the data-generating process that has formal consequences for stability, learnability, and prediction.
The distinction between physically-determined and human-decision-determined quantities has existed as long as markets have. Philosophers of social science have explored it (Searle, 1995; Bhaskar, 1975). Financial economists have documented its symptoms - correlation breakdown during crises, regime switches, factor rotation (Longin and Solnik, 2001; Forbes and Rigobon, 2002). Soros (1987) articulated the reflexivity principle verbally. Keynes (1936) described the beauty contest. MacKenzie (2006) demonstrated performativity. Lucas (1976) identified the instability of behavioral parameters under policy change; the present framework generalizes the observation to the full covariance structure and derives the dynamic consequences.
None of them formalized the space.
Why does formalization matter now? Because of what is about to enter the system.
Between 2020 and 2025, autonomous AI agents became increasingly significant participants in financial markets, supply chain logistics, pricing algorithms, and resource allocation. These agents are trained on historical data, deployed to act on current data, and periodically retrained on data that their own actions have reshaped. Every agentic AI system operating in economic space is a Beierspace participant. It does not observe an independent reality. It co-constitutes the statistical structure it is trying to learn.
The scale of this feedback is about to undergo a discontinuous expansion. Agentic AI systems are proliferating across trading desks, insurance pricing, credit allocation, supply chain optimization, dynamic pricing, and autonomous resource management. Each of these systems generates Beierspace quantities through its decisions. Each of them trains on the outputs of all the others. The covariance structure of the economy is becoming increasingly self-referential at a speed and scale that has no historical precedent.
Without a name for this space, and without a formal understanding of its dynamics, the field has no vocabulary to describe what is happening. Correlation breakdown is discussed as if it were a weather event. Model degradation is treated as a technical failure to be patched. Regime switches are modeled as exogenous - as if the switches arrive from outside the system rather than being generated by the participants whose models assume stability.
Beierspace is not a theoretical abstraction. It is the space in which the global economy operates, in which AI agents are being deployed at scale, and whose dynamics determine whether the statistical structures we depend on hold together or collapse.
The central formal result, derived in full in Beier (2026e), is a stability condition:
S = LD / (sigma_e + sigma_b(LD))
where L is aggregate learning capacity, D is aggregate disclosure rate, sigma_e is the rate of physical structural change, and sigma_b is the rate of human-generated structural change - the Beier drift.
The critical observation is that sigma_b is endogenous. It is a strictly decreasing function of LD. The agents whose learning and disclosure suppress Beier drift are the same agents whose decisions generate it. When participants learn more and disclose more, their beliefs converge, their actions coordinate, and the structural drift they collectively produce diminishes. When learning and disclosure degrade, beliefs diverge, actions scatter, and the drift accelerates.
This reflexive coupling produces three consequences that hold for any continuous, strictly decreasing drift function - not only for specific parameterizations.
Phase transition at S = 1. Above the threshold, the system self-stabilizes: better information produces more stability, which enables better information. Below the threshold, the system self-destabilizes: degraded information produces more drift, which degrades information further. The fixed point at S = 1 is unstable. Small perturbations above it lead to convergence. Small perturbations below it lead to collapse.
Bistability. With bounded re-entry forces (participants do not instantly return to the market after a crisis), the system exhibits two stable equilibria - a well-functioning state and a collapse trap - separated by an unstable fixed point. The boundary between these basins of attraction is governed by a saddle-node bifurcation at a computable critical parameter.
Informational impossibility. A system in the collapse trap cannot determine from its own data whether the collapse is permanent or recoverable. The parameters distinguishing these regimes are the structural parameters of the Beier drift mechanism - and these are precisely the parameters that are not identifiable from the degraded data the collapsed system can produce. The system is informationally blind to its own prognosis.
AI agents amplify every term in the stability condition.
Learning capacity (L). AI systems process information faster and across more dimensions than human participants. An agent monitoring 10,000 time series simultaneously operates at a learning capacity that no human trader approaches. In principle, this increases L and pushes S above the threshold.
Disclosure rate (D). Many AI systems operate as black boxes. Proprietary trading algorithms do not publish their models. Pricing engines do not explain their adjustments. If AI increases L while decreasing D - learning more but sharing less - the product LD may not increase. The stability condition depends on both.
Beier drift (sigma_b). AI agents whose strategies converge (because they are trained on similar data with similar architectures) will coordinate their actions, reducing Beier drift. AI agents whose strategies diverge (because of different training windows, objective functions, or data pipelines) will scatter their actions, increasing Beier drift. The net effect on sigma_b is not determined by the technology. It is determined by the information architecture - who learns what, and who shares what with whom.
Cycle time. The reflexive loop between decisions, outcomes, and retraining operates on a timescale that AI compresses by orders of magnitude. Human participants update their models over weeks and months. AI agents can retrain overnight. The phase transition dynamics that took years to unfold in historical crises could unfold in hours when the dominant participants are machines.
Three scenarios clarify the stakes.
Convergent AI. A market dominated by AI agents trained on the same data, using similar architectures, optimizing similar objectives. Learning capacity is enormous. Disclosure is unnecessary because the models are effectively public knowledge (everyone has GPT). Beier drift is low because actions coordinate. S is high. The system is stable - until the shared model encounters a regime it was not trained on, at which point all agents fail simultaneously. The phase transition is not gradual. It is a cliff.
Divergent AI. A market with heterogeneous AI agents using different data, architectures, and objectives. Each agent has high L individually but low D (proprietary systems do not share). The agents' decisions generate high Beier drift because they act on incompatible models. S may be low despite enormous total computational capacity. Stability depends on the disclosure architecture, not the intelligence of the participants.
Hybrid AI-human. The realistic near-term scenario. AI agents and human participants co-inhabit Beierspace. The AI agents compress the cycle time of the reflexive loop. The human participants provide (or fail to provide) the disclosure and interpretive capacity that the stability condition requires. The suppleness of the human participants - their capacity to absorb large model revisions without degraded learning - becomes the binding constraint on system stability.
The companion paper (Beier, 2026e) derives a suppleness multiplier from established results in cognitive psychology: the zone of tolerance framework (Sherif and Hovland, 1961), the biased assimilation literature (Lord et al., 1979), and the deliberate practice literature (Ericsson et al., 1993). Suppleness - the capacity to absorb large model revisions without degraded learning - unconditionally amplifies stability and conditionally expands the usable dimensionality of the eigenspace.
For human participants, suppleness is constrained by cognitive architecture. People resist belief revision. Disconfirming evidence produces defensive processing. The zone of tolerance is finite and expands slowly through deliberate practice.
For AI agents, suppleness is a design variable. It can be set, modified, and optimized. An AI system can be built to absorb arbitrary model revisions without flinching - or it can be built to anchor on prior beliefs and resist updating. The choice is engineering, not biology.
This means the stability of Beierspace systems in the AI age is partially a design problem. The suppleness of the AI agents participating in the system is not given. It is chosen. Regulators, architects, and engineers can influence the stability condition directly by specifying how AI agents respond to disconfirming evidence, how rapidly they update, and how completely they disclose their state.
The ceiling on this improvement is sharp. Regardless of suppleness, no participant - human or AI - can usably track dimensions whose Beier drift rate exceeds its information capacity. The extension theorem (Beier, 2026e) provides the boundary: K* = |{k : sigma_k < LD}|. Dimensions whose drift rates exceed the population's aggregate learning and disclosure capacity are structurally unlearnable. Suppleness expands the usable interior. It does not move the wall.
Without the Beierspace concept, practitioners and researchers face a vocabulary gap that produces systematic errors.
Risk management. Value-at-Risk assumes a covariance matrix with eigenstructure reflecting persistent features of the economy. In Beierspace, those features persist only as long as the decisions constituting them persist. A VaR model that cannot distinguish eigenspace from Beierspace components cannot assess its own fragility. The vocabulary enables a decomposition: which components of my risk model depend on physics, and which depend on the continuation of a particular pattern of human and AI decisions?
Machine learning. ML models trained on economic data assume approximate stationarity over the training window. In Beierspace, stationarity is not a property of the world. It is an achievement of the information architecture. A model trained during a period of high S (coordinated, well-disclosed decision-making) will encounter fundamentally different statistical structure when S degrades. The vocabulary enables the question: is my training data eigenspace-stable or Beierspace-contingent?
Regulation. Financial regulation after 2008 focused on capital requirements and stress testing. The Beierspace framework reveals that these interventions operate on the wrong variable. Capital buffers absorb losses after the covariance structure breaks. The stability condition identifies what prevents the break: learning capacity and disclosure rate. Mandating disclosure (increasing D) has a direct, formalizable effect on S. Mandating capital buffers does not.
AI safety. The alignment literature focuses on individual AI systems behaving as intended. The Beierspace framework identifies a collective risk that individual alignment cannot address: even perfectly aligned AI agents, acting optimally on their individual objectives, can collectively destabilize the covariance structure they all depend on. The risk is not that an AI goes rogue. The risk is that many AI systems, each functioning correctly, co-constitute a Beierspace whose drift rate exceeds the system's aggregate learning capacity.
A concept that cannot be tested cannot be adopted. The following predictions follow from the Beierspace framework and are incompatible with the standard assumption that covariance instability is exogenous.
Prediction 1. Markets with higher aggregate LD (measured by information quality indices, transparency scores, and disclosure compliance) will exhibit smaller Frobenius norm changes in their DCC-GARCH covariance estimates during stress events, controlling for shock magnitude. If covariance instability were exogenous, LD would be irrelevant to eigenstructure stability.
Prediction 2. Regulatory interventions that increase D (mandatory disclosure, transparency requirements) will reduce eigenvalue volatility with a lag of 6 to 18 months. Capital requirements that do not change D will not.
Prediction 3. Sectors dominated by AI agents with convergent architectures will exhibit lower baseline Beier drift but larger tail events (cliff-edge collapses) than sectors with diverse decision architectures. The Beierspace framework predicts this asymmetry. The exogenous-instability framework does not.
Prediction 4. As the fraction of market volume attributable to autonomous AI agents increases, the autocorrelation of covariance structure changes will decrease (faster cycling), but the amplitude will depend on D, not on L alone. Markets where AI agents are required to disclose their state will be more stable than markets where they are not, even if the non-disclosing agents are individually more capable.
Eigenspace and Beierspace coexist in every economic dataset. The eigendecomposition cannot distinguish them. The consequences of confusing them - assuming that human-decision-determined covariance structures are as stable as physically-determined ones - range from failed risk models to systemic collapse.
The AI agents now entering economic space at scale are not eigenspace observers. They are Beierspace participants. They consume statistical structure that they and their peers collectively produce, and they produce statistical structure that they and their peers will collectively consume. The reflexive loop that has always characterized human-decision-determined quantities is about to accelerate by orders of magnitude.
The framework developed here and in the companion papers provides the vocabulary, the stability condition, the phase transition dynamics, and the impossibility results necessary to analyze this system. The predictions are stated. The tests are specified. Whether the field adopts the vocabulary before or after the consequences arrive is the outstanding question.