What happens when we don't actually know the system's present state?

01 / The problem comes before prediction

Prospective health begins with a present state. How do we know that state?

We've returned to this definition throughout the series: a system must presently realize its organization and have sufficient capacity for viable continuation under the specified scenario and horizon.

We have to infer it from recorded stem identities, measurements, dates, apparent survival, and missing information. Errors or gaps in that inference flow into everything we later say about continuation.

02 / The observed site

A 35-hectare forest, observed over many years rather than on one day.

The empirical study uses the Harvard Forest ForestGEO plot in Massachusetts, a 700 × 500 meter research area. Researchers recorded large numbers of woody stems during two census periods, but the observations were not all made at the same instant.

Measurements taken years apart do not describe the same point in time. Before a future simulation can begin on January 3, 2020, the method must reconcile differing observation dates and reconstruct a plausible common-time state.

The research footprint and dates describe the source study, not a new field survey undertaken for this lesson.

03 / What was actually recorded?

The record is extensive. It is also incomplete in consequential ways.

These counts describe different record categories and should not be added together as though they were mutually exclusive groups of trees. Each matters for a different reason.

stem_gone is not proof of biological death. A missing diameter introduces uncertainty in physical structure. A stem appearing only in E1 may be a genuine entrant, a previously missed stem, or another identity-recording outcome. The raw census is evidence about the forest, rather than a complete, error-free state description.

04 / Observed evidence and unobserved state

I need to distinguish what was seen from what I am inferring.

Let Y contain the two census records. Let X denote the modeled forest state at our common starting date. In statistical language, X is a latent state: the state of interest is not fully and directly observed.

Imagine a purely illustrative census of 100 trees. Seventy are confirmed alive later, ten are known dead, and twenty have unresolved fates. Neither “only seventy survive” nor “all twenty unresolved trees survive” follows from those observations.

Our data have not changed when the slider moves. The statistical assumption has. That difference is exactly why the resulting inferred state cannot be treated as a direct observation.

05 / Compressing the forest

Hundreds of thousands of records become 64 cohort cells.

Following each individual stem through every demographic mechanism would require an extremely detailed system. The paper defines a computational representation with three grouping dimensions, four groups in each dimension.

Each cohort cell c stores a stem count n_c and a root-mean-square diameter d_c. One spatial/protocol category is a wet-candidate proxy; the study does not independently validate a corresponding hydrological classification.

As we saw in Paper III, compression must be judged against the questions the representation will later need to answer. A summary that retains one useful quantity may erase another.

06 / What the compression preserves

Root-mean-square diameter keeps the basal-area calculation intact.

Tree basal area scales with the square of trunk diameter. For diameters d₁ through d_n, the root-mean-square (RMS) diameter is the square root of the average squared diameter:

With diameter measured in centimeters and basal area expressed in square meters, the cohort-state basal area is:

This exact aggregation protects the basal-area calculation, given the grouped stem values. It does not recover every original diameter. Two cohorts may share n_c and d_c while containing different counts of stems below the study's 10-centimeter continuation threshold. That limitation becomes important later in Paper V.

07 / Different stems, different observation windows

A modeled annual mortality hazard converts elapsed time into a survival probability.

Because individual stems were observed on different dates, each stem has its own elapsed observation interval. Paper V uses a conditional constant-hazard survival model.

For a positive constant hazard, a longer interval yields a lower modeled probability of survival. This allows the reconstruction to respect actual observation dates rather than treating all intervals as equally long.

It remains a model assumption. The real hazard is not known in advance, and a constant-hazard law is not established by writing the formula. The method must estimate its parameters and test how the resulting predictions behave.

08 / Sampling possible present states

Instead of one certain present, we generate possible present forests.

In the primary computational design, the paper samples 256 reconstructed states. Combining those draws with parameter variations and two growth-model variants gives 1,024 outer computational units, each with 64 simulated future histories per scenario. These are conditional computational units, not 1,024 separately observed real forests.

Every possible future begins downstream of this uncertain starting point. The simulation pipeline can be executed exactly as designed while the scientific accuracy of its assumptions remains an open question.

10 / A deeper boundary than numerical uncertainty

The later census cannot by itself tell us how every later-only stem entered the record.

The identity audit finds 6,992 living stem IDs appearing only in E1, and none can be certified from the available two-census identity evidence as an unequivocal biological threshold entrant. This result does not mean that no biological entry occurred.

Both histories may produce the same later-only observation. When two underlying explanations give indistinguishable evidence, no deterministic function of that evidence alone identifies which one occurred. This is non-identifiability, and it is the representation problem from Paper III encountered in an empirical study.

We can introduce an assumption-based entry model. We cannot turn an unobserved event into an observed fact by adding a parameter. Additional identifying evidence would be needed.

11 / Test the distinction

One stem is recorded in E1 but is absent from E0. What follows from that fact alone?

12 / What kind of knowledge do we have?

There are three different levels of inference.

Each stage inherits the limits of the previous one and introduces additional modeling commitments. Correct execution of the declared mathematics and reliable measurement of the real forest are distinct achievements. Paper V documents both the construction and the points at which empirical evidence does not yet justify a stronger conclusion.

This study does not establish an operational health classification of Harvard Forest. It establishes an auditable path from recorded evidence toward model-conditional continuation estimates, together with substantial calibration and identification limits.