What Does It Mean for a Concentration Process to Be Nonnegative?
A concentration is physically interpreted as a quantity that cannot fall below zero. But a stochastic model for concentration does not automatically respect that interpretation: Brownian fluctuations can push a mathematical state into negative values. The relevant question is therefore not whether the model has a positive average, but whether its sample paths remain in the concentration state space.
Consider a scalar Itô SDE on a time interval \([0,T]\):
Here \(C_t\) is concentration, \(W_t\) is standard Brownian motion, \(b\) is the drift coefficient, and \(\sigma\) is the diffusion coefficient. The starting value \(c_0\) is nonnegative. The model has a nonnegative state space when its paths stay in \([0,\infty)\), rather than merely having nonnegative summaries.
The phrase “for every \(t\)” matters. A probability statement at each fixed time, such as \(P(C_t\geq 0)=1\) separately for every \(t\), is weaker in general than a single probability-one statement covering the whole path. For continuous SDE solutions, pathwise nonnegativity is the natural requirement when the concentration state space is \([0,\infty)\).
Zero Is a Boundary, Not an Ordinary State
For a real-valued SDE, zero is an interior point: a path can move from positive to negative values continuously by crossing it. For a concentration model restricted to \([0,\infty)\), zero is instead the boundary of the permitted state space. The behavior of both coefficients there is consequential. If the diffusion coefficient is nonzero at zero, Brownian fluctuations act in both directions at the boundary. A drift pointing into negative values also conflicts with the concentration interpretation.
A useful sufficient condition can be stated precisely for a common setting. Suppose that, on each finite time interval, \(b\) and \(\sigma\) are continuous in time and globally Lipschitz in the state, with a Lipschitz constant \(L\) uniform in time. These assumptions ensure a unique, non-explosive strong solution. If, for every \(t\),
then zero is an invariant boundary: a solution started at \(c_0\geq0\) remains nonnegative. The drift at zero does not point outward, and the diffusion vanishes there. The conditions are sufficient, not a complete characterization for every possible coefficient or boundary mechanism.
Proof. Write \(C_t^-=\max(-C_t,0)\) for the negative part and set \(Y_t=(C_t^-)^2\). The generalized Itô formula applies to \(x\mapsto(x^-)^2\), which is continuously differentiable with locally absolutely continuous derivative. Up to a localization time \(\tau_n=\inf\{t:|C_t|\geq n\}\), it gives
When \(C_t=-u<0\), where \(u=C_t^->0\), Lipschitz continuity and the boundary assumptions imply \(b(t,-u)\geq b(t,0)-Lu\geq-Lu\) and \(|\sigma(t,-u)|\leq L u\). Consequently, the drift of \(Y_t\) is bounded above by \((2L+L^2)Y_t\). The stopped stochastic integral has expectation zero on a finite interval, since its integrand is bounded after localization. Thus, with \(K=2L+L^2\),
Since \(Y_0=0\), Gronwall’s inequality gives \(E[Y_{t\wedge\tau_n}]=0\). Nonnegativity of \(Y\) then implies \(Y_{t\wedge\tau_n}=0\) almost surely. Letting \(n\) increase, non-explosion gives \(C_t^-=0\) throughout every finite interval, almost surely. Continuity of the paths extends the result from fixed times to all times simultaneously. \(\square\)
The proof shows why the boundary conditions work: below zero, Lipschitz coefficients can generate at most a drift in the squared negative part proportional to that negative part. Starting from zero negative part, it cannot grow. This argument establishes a sufficient condition; it does not say that every nonnegative model must have coefficients of exactly this form.
Worked Examples: Positive Paths and Positive Moments
Worked Example: Multiplicative Noise Gives Strictly Positive Paths
Let \(C_0=2\) and consider \(dC_t=-0.4C_t\,dt+0.3C_t\,dW_t\). This geometric diffusion has an explicit solution. Applying Itô’s formula to \(\log C_t\) gives
The exponential is positive for every finite value of \(W_t\); continuity of Brownian paths therefore makes \(C_t>0\) for all finite \(t\), almost surely. At \(t=2\), \(W_2\sim N(0,2)\), so \(\log C_2\) is normal with mean \(\log 2+(-0.4-0.3^2/2)2=\log 2-0.89\) and variance \(0.3^2(2)=0.18\). The mean calculation is \(-0.4-0.045=-0.445\) per time unit, giving \(-0.89\) over two units. Here strict positivity follows from the solution’s exponential form, not from the signs of its moments.
Worked Example: Positive Mean Does Not Prevent Negative Concentrations
Take \(C_0=1\) and \(dC_t=0.5(1-C_t)\,dt+1.5\,dW_t\). This is an Ornstein–Uhlenbeck process with long-run mean one. Its solution at time \(t\) is normal, with
At \(t=1\), the variance is \(2.25(1-e^{-1})\approx 1.4223\), and the standard deviation is approximately \(1.1926\). Therefore \(P(C_1<0)=\Phi(-1/1.1926)=\Phi(-0.8385)\approx0.201\), where \(\Phi\) is the standard normal distribution function. The mean is positive and the variance is, as always, nonnegative; nevertheless, there is about a 20% chance of a negative concentration at this time. Thus the process is not pathwise nonnegative.
Worked Example: Checking the Coefficients at Zero
Consider \(dC_t=(1-C_t)\,dt+0.2C_t\,dW_t\), starting from \(C_0\geq0\). The drift and diffusion are globally Lipschitz in concentration. At zero, \(b(t,0)=1\geq0\) and \(\sigma(t,0)=0\). The theorem therefore guarantees pathwise nonnegativity. It does not by itself claim strict positivity or give a distribution for \(C_t\).
Now compare the deterministic boundary-starting model \(dC_t=-1\,dt\), \(C_0=0\). Its solution is \(C_t=-t\), so it immediately leaves the nonnegative state space: the drift at zero points outward. If instead a nonzero additive diffusion were present at zero, the boundary condition on \(\sigma\) would fail; the Brownian term would permit fluctuations on either side of zero. These checks identify why zero requires explicit attention in model specification.
Model Choice, Diagnostics, and Limitations
Three ideas should not be conflated. First, \(E[C_t]\geq0\) concerns an average over possible outcomes. Second, \(\operatorname{Var}(C_t)\geq0\) is a general property of variance and says nothing about which values \(C_t\) can take. Third, pathwise nonnegativity is a support restriction on the process itself. A distribution can have a positive mean and substantial probability below zero, as the Gaussian example demonstrates.
When assessing a concentration SDE, inspect the behavior at zero before interpreting its trajectories. Ask whether the diffusion vanishes there and whether the drift points into or along the permitted state space. Then check the assumptions behind any claimed invariance result: for the theorem above, coefficients must be uniformly Lipschitz in the state on finite time intervals, and the solution must be non-explosive. A coefficient check without its hypotheses is not a proof of pathwise nonnegativity.
The main modeling alternatives have different meanings. A model with multiplicative noise, such as the geometric diffusion above, can stay strictly positive through its exponential solution. A Gaussian diffusion with additive noise allows the real line as its state space and may generate negative concentrations. One may instead specify a process whose coefficients preserve \([0,\infty)\), subject to suitable conditions. Clipping negative values to zero or reflecting them at the boundary changes the model; neither operation follows automatically from an unconstrained SDE.
The boundary criterion is deliberately limited. It is a practical sufficient result under regularity assumptions, not a classification of all possible boundary behaviors. More specialized coefficient structures and boundary tests are studied separately; in particular, the criterion here does not replace an analysis of whether a particular diffusion can reach zero or remain strictly positive.
Check Your Understanding
Answer each question, then compare with the solutions below.
- State the probability event that defines pathwise nonnegativity on \([0,T]\).
- Why do a positive mean and a nonnegative variance fail to establish pathwise nonnegativity?
- For \(dC_t=(2-C_t)\,dt+0.5C_t\,dW_t\), check the boundary conditions in the theorem and state its conclusion when \(C_0\geq0\).
- For the Ornstein–Uhlenbeck example with mean one and variance \(2.25(1-e^{-t})\), give \(P(C_1<0)\) in terms of \(\Phi\), using the computed standard deviation.
Solutions. (1) \(P(C_t\geq0\text{ for every }t\in[0,T])=1\). (2) Moments summarize a distribution but do not restrict its support; the worked Gaussian process has positive mean and a substantial negative tail. (3) Here \(b(t,0)=2\geq0\), \(\sigma(t,0)=0\), and both coefficients are globally Lipschitz, so the theorem gives pathwise nonnegativity on every finite interval. (4) The probability is \(\Phi(-1/1.1926)=\Phi(-0.8385)\), approximately \(0.201\).