Homeostasis is not stillness: How information changes the cost of staying stable

An illustrative stochastic-control and Bayesian-sensing model, with reproducible Wolfram Language calculations.

A stable system does not have to be motionless. A thermostat varies its response when conditions change; an organism likewise regulates variables while exchanging matter and energy with its surroundings. Regulation can therefore be understood as a dynamic process rather than an insistence on an unchanging state.

But there is an important trade-off: stronger regulation can reduce fluctuations, while regulation itself consumes resources. And to react appropriately, a controller must acquire information — which may also be costly and imperfect.

Question: When does environmental information improve regulation enough to justify its cost?

The following is a deliberately small mathematical example, not a model of an entire organism, a claim of evolutionary inevitability, or a new theorem. Its purpose is to make the trade-off explicit and testable.

1. Stability without stillness

Consider the Ornstein–Uhlenbeck model
dx_t = -k x_t dt + sqrt(v) dW_t, with k>0, v>0.
Here x_t is a deviation from a reference state, k the regulatory gain, and v the environmental noise intensity. Despite continuous stochastic motion, the stationary second moment is finite:
E[x_t^2] = v/(2k).
Thus higher gain reduces dispersion, but how much regulation is worth paying for?

2. Regulation is not free

Choose the illustrative normalized stationary cost
L(k,v) = alpha v/(2k) + beta k^2, with alpha,beta>0.
The first term penalizes state dispersion; the second is an assumed quadratic gain-maintenance cost, not a measured metabolic law and not the usual quadratic actuator-effort cost E[(kx)^2]. Differentiation gives a unique minimum:
k*(v) = (alpha v/(4 beta))^(1/3).
The optimal gain grows as v^(1/3) or as sigma^(2/3) if v=sigma^2. In this model, stronger disturbances call for stronger — but sublinearly increasing — regulation.

3. Knowing the environment changes the optimal decision

Suppose the positive noise intensity v is randomly drawn before a sufficiently long stationary episode. A blind controller knows its prior distribution but not the actual value. An informed controller observes v before choosing its gain.
Set C=3 beta^(1/3) (alpha/4)^(2/3). Optimizing the gain gives
J_blind = C (E[v])^(2/3), J_perfect = C E[v^(2/3)].
Since v^(2/3) is concave, Jensen’s inequality yields
J_perfect = J_blind.
This is the simple point: knowing a changing environment can make regulation less costly even when the physics of the regulated system is unchanged. The inequality is strict for a nondegenerate positive prior in this particular model.

4. An imperfect sensor: useful information, not perfect information

Suppose a sensor returns y=v+epsilon, where its Gaussian measurement error is independent of v. The controller updates to m(y)=E[v|y]. Its optimal gain is
k*(y) = (alpha m(y)/(4 beta))^(1/3).
Applying conditional Jensen, followed by ordinary Jensen, yields
J_perfect = J_noisy = J_blind.
The difference J_blind - J_noisy is the gross value of the sensor’s information, in units of the specified objective. This is not automatically a thermodynamic energy saving.

5. A small reproducible numerical example

Let v=1 or 4, with equal prior probabilities, and alpha=beta=1. For a Gaussian sensor with standard deviation s, Bayes’ rule gives
P(v=4|y)=1/(1+exp[-3(y-2.5)/s^2]).
Numerical integration provides:

  • Blind optimum: 2.193013303660
  • Perfect-information optimum: 2.095275394488
  • Maximum value of information: 0.097737909172
  • Value of information when s=1: 0.077621150439
    For an illustrative constant sensor operating cost C_s=0.05 per unit time, the break-even measurement-noise level is
    s_c approximately 1.556459225823.
    The decision rule is to use the sensor when its gross value exceeds C_s. The threshold is specific to these normalized assumptions; it is not a universal physical constant.

6. What changes when the environment itself switches?

Real environments may alternate between regimes before a system reaches stationarity. A two-state continuous-time Markov chain and a noisy continuous observation can be described with a Wonham filter. In the corresponding conditional-second-moment model, the joint conditional moments for each regime are needed in addition to the current regime probability; a single unconditional variance is generally insufficient to close the filtered dynamics.
This leads to a partially observed stochastic-control problem. We have its formal filtering structure, not a verified numerical solution of its three-dimensional Hamilton–Jacobi–Bellman equation. This extension is an open computational direction, not one of the demonstrated results above.

Why this matters

The elementary model separates three claims that are easily conflated:

  • Regulation can maintain bounded fluctuations without eliminating motion.
  • Information can improve the allocation of regulatory effort.
  • Adaptation and evolution require additional dynamics and evidence; neither is established by these equations alone.
    The conceptual proposal is modest but useful: homeostasis need not mean stillness; it may be achieved through ongoing regulation whose cost depends on what the system can learn about its surroundings.

Question for the community: What is the most informative next test when environmental switching, sensor noise, and regulatory response times are comparable?

Reproducibility and attribution. The Wolfram Language source for the stationary optimization, Bayesian quadrature, break-even threshold, and figures has been prepared and is available from the author on request. Computations have been independently cross-checked in a Wolfram kernel. This post uses established mathematics; it does not claim a novel control theorem or a new biological law. Drafting, code preparation, and mathematical auditing were performed with substantial assistance from OpenAI ChatGPT.

Reproducibility update: The initial omission of the accompanying Wolfram Language code was due to an operational misunderstanding. The code has now been executed in Wolfram Language, and its checks passed (stationarity 0; curvature True; ordering True; break-even residual 1.35031*^-14). The mathematical conclusions remain unchanged.

(* JSM Wolfram Community — audited stationary model *)
(* Not a CTMC/HJB numerical solver. Assumptions alpha,beta>0, v>0, k>0. *)
ClearAll[alpha,beta,v1,v2,costFactor,kStar,L,blind,perfect,fullVoI,posterior,meanPosterior,normalPDF,noisy,voi,sensCost,sStar,z,s];
alpha=1; beta=1; v1=1; v2=4;
L[k_,v_]:=alpha v/(2 k)+beta k^2;
kStar[v_]:=(alpha v/(4 beta))^(1/3);
costFactor=3 beta^(1/3) (alpha/4)^(2/3);
blind=costFactor ((v1+v2)/2)^(2/3);
perfect=costFactor (v1^(2/3)+v2^(2/3))/2;
fullVoI=blind-perfect;
posterior[y_?NumericQ,s_?NumericQ]:=1/(1+Exp[-(v2-v1)(y-(v1+v2)/2)/s^2]);
meanPosterior[y_?NumericQ,s_?NumericQ]:=v1+(v2-v1) posterior[y,s];
normalPDF[z_?NumericQ]:=Exp[-z^2/2]/Sqrt[2 Pi];
noisy[s_?NumericQ]/;s>0:=costFactor NIntegrate[
 ((meanPosterior[v1+s z,s]^(2/3)+meanPosterior[v2+s z,s]^(2/3))/2)*normalPDF[z],
 {z,-10,10},AccuracyGoal->9,PrecisionGoal->9,
 Method->{"GlobalAdaptive","SymbolicProcessing"->0}];
voi[s_?NumericQ]/;s>0:=N[blind-noisy[s]];
voi[0]=N[fullVoI];
sensCost=0.05;
sStar=s/.FindRoot[voi[s]==sensCost,{s,1.5,1.7}];
checks=
 "symbolic stationarity"->FullSimplify[D[alpha v/(2 s)+beta s^2,s]/.s->(alpha v/(4 beta))^(1/3),{alpha>0,beta>0,v>0}],
 "symbolic curvature positive"->FullSimplify[D[alpha v/(2 s)+beta s^2,{s,2}]>0,{alpha>0,beta>0,v>0,s>0}],
 "blind value"->N[blind,15],"perfect value"->N[perfect,15],
 "VoI at s=1"->voi[1.],"sensor break-even s"->sStar,
 "ordering"->TrueQ[N[perfect]