5D Library › Architectural Acoustics, Reverberation Time and the Environmental Psychology of Built Space · 中文
Chapter 4 Trees outside the Window: Fractal Dimension and Biophilic Design
The story of a window
Evidence
Ulrich, 1984. Roger Ulrich examined the records of patients who had undergone cholecystectomy between 1972 and 1981 (from May to October each year) in a suburban hospital in Pennsylvania. On the same side of a ward, some windows looked out on a small stand of deciduous trees and others on a brick wall. He matched patients into 23 pairs on sex, age, smoking, obesity and other conditions. Patients with a view of trees stayed on average 7.96 days after surgery, those with the wall 8.70 days; nurses’ notes contained on average 1.13 negative comments per patient for the first group and 3.96 for the second; on days 2–5 after surgery, patients with the tree view received fewer moderate and strong analgesic doses and more weak ones, while on the day of surgery, day 1 and days 6–7 there were no significant differences (Ulrich 1984).
This short paper in Science is often cited as the founding evidence of “biophilic design”. Its value lies in asking the right question; its limitations should be stated with it: it is a retrospective comparison of records, not random allocation; it involves only 23 pairs of patients, one hospital and one operation; and the comparison “wall” was a monotonous brick wall, so that, as Ulrich himself cautioned, the conclusion need not extend to all built views.
Two explanations and an uncertain review
Why might natural views help? There are two main theories. The stress recovery theory of Ulrich and colleagues holds that natural scenes calm stress responses more quickly. They showed 120 participants a stressful film followed by video of natural or urban settings and found that recovery, both in self-rated mood and in physiological measures such as heart rate, muscle tension and skin conductance, was faster and more complete with the natural scenes (Ulrich et al. 1991). Kaplan’s attention restoration theory holds that natural settings engage “soft fascination” and so let effortful, directed attention rest (Kaplan 1995). Wilson coined “biophilia” for the human tendency to affiliate with other forms of life (Wilson 1984).
Evidence
A systematic review of attention restoration. Ohly and colleagues in 2016 reviewed 31 studies and ran 13 meta-analyses, of which only 3 were significant: digit span forward, digit span backward and Trail Making Test B; the other 10 showed no clear benefit. The authors judged that the evidence gave “some support” to attention restoration theory, but that studies were few and small, quality was inconsistently reported and outcomes varied widely, so that it remained uncertain which aspects of attention are affected (Ohly et al. 2016).
Fractals and box-counting dimension
Is there a visual feature that natural scenes share? One answer is fractals. Mandelbrot called shapes that repeat similar structure across scales fractals, and described how fully they fill space by their fractal dimension (Mandelbrot 1982): a smooth curve has , a filled area , and a curve in between has a the closer to 2 the more convoluted it is.
Calculation
Box counting. Cover the figure with boxes of side and count the boxes that the curve passes through. For a fractal, , so
in practice one fits a straight line to against over a range of and takes its slope. The Koch curve replaces each segment by 4 segments of one third the length at every step, so . The profile of fractional Brownian motion with Hurst exponent has dimension : the smaller , the rougher the profile.
The program below has two parts. The first tests box counting on curves of known dimension; the second is read in the last section of this chapter.
Calculation
# Box-counting dimension: cover a curve in the unit square with boxes of side e = 2^-k,
# count occupied boxes N(e), fit log N against log(1/e). Part 1: test curves of known D
# (a line, the Koch curve, fractional-Brownian profiles with D = 2 - H). Part 2: four
# synthetic window views.
import math, random
def box_dim(pts, kmin, kmax):
xs, ys = [], []
for k in range(kmin, kmax + 1):
e = 2.0 ** -k
top = 2 ** k - 1
boxes = {(min(int(x / e), top), min(int(y / e), top)) for x, y in pts}
xs.append(k * math.log(2)); ys.append(math.log(len(boxes)))
mx, my = sum(xs) / len(xs), sum(ys) / len(ys)
sxy = sum((a - mx) * (b - my) for a, b in zip(xs, ys))
return sxy / sum((a - mx) ** 2 for a in xs)
def densify(pts, step=1e-3): # add points so no gap exceeds `step`
out = [pts[0]]
for (x0, y0), (x1, y1) in zip(pts, pts[1:]):
n = max(1, int(math.hypot(x1 - x0, y1 - y0) / step) + 1)
out += [(x0 + (x1 - x0) * i / n, y0 + (y1 - y0) * i / n) for i in range(1, n + 1)]
return out
def koch(level):
pts = [(0.0, 0.0), (1.0, 0.0)]
for _ in range(level):
new = [pts[0]]
for (x0, y0), (x1, y1) in zip(pts, pts[1:]):
dx, dy = (x1 - x0) / 3, (y1 - y0) / 3
a, b = (x0 + dx, y0 + dy), (x0 + 2 * dx, y0 + 2 * dy)
r3 = math.sqrt(3) / 2
peak = (a[0] + dx / 2 - dy * r3, a[1] + dy / 2 + dx * r3)
new += [a, peak, b, (x1, y1)]
pts = new
return [(x, y * 3 + 0.05) for x, y in pts] # stretch into the unit square
def fbm(H, seed, n=11): # random midpoint displacement, scaled to 0..1
rng = random.Random(seed)
N = 2 ** n
y, step, scale = [0.0] * (N + 1), N, 1.0
while step > 1:
half = step // 2
scale *= 0.5 ** H
for i in range(half, N, step):
y[i] = (y[i - half] + y[i + half]) / 2 + rng.gauss(0, scale)
step = half
lo, hi = min(y), max(y)
return [(v - lo) / (hi - lo) for v in y]
line = lambda y, y0=0.05, span=0.9: densify([(i / (len(y) - 1), y0 + span * v)
for i, v in enumerate(y)])
CURVES = [("straight line", 1.0, densify([(0.0, 0.1), (1.0, 0.9)])),
("Koch curve", math.log(4) / math.log(3), densify(koch(7)))]
CURVES += [(f"fBm profile H = {H}", 2 - H, line(fbm(H, 31, 12)))
for H in (0.9, 0.7, 0.5, 0.3)]
print(f"{'curve':20s} true D box-count D, e = 1/4..1/128 e = 1/16..1/256")
for name, D, p in CURVES:
print(f"{name:20s} {D:.3f} {box_dim(p, 2, 7):.3f}"
f" {box_dim(p, 4, 8):.3f}")
rng = random.Random(7) # block skyline: random widths and heights
B, i = [0.0] * 2049, 0
while i < 2049:
w, h = rng.randint(60, 260), rng.uniform(0.25, 0.85)
B[i:i + w] = [h] * len(B[i:i + w]); i += w
TREES = [0.1 + 0.5 * v for v in fbm(0.3, 5)]
VIEWS = {"block skyline": B,
"mountain ridge (H 0.7)": [0.2 + 0.6 * v for v in fbm(0.7, 3)],
"rough tree line (H 0.3)": TREES,
"blocks behind trees": [max(b, t) for b, t in zip(B, TREES)]}
print(f"{'window view':25s} D (e = 1/8..1/128) D (e = 1/32..1/512)")
for name, y in VIEWS.items():
p = line(y)
print(f"{name:25s} {box_dim(p, 3, 7):.2f} {box_dim(p, 5, 9):.2f}")
# curve true D box-count D, e = 1/4..1/128 e = 1/16..1/256
# straight line 1.000 1.001 1.000
# Koch curve 1.262 1.300 1.245
# fBm profile H = 0.9 1.100 1.210 1.135
# fBm profile H = 0.7 1.300 1.293 1.261
# fBm profile H = 0.5 1.500 1.379 1.399
# fBm profile H = 0.3 1.700 1.520 1.537
# window view D (e = 1/8..1/128) D (e = 1/32..1/512)
# block skyline 1.04 1.00
# mountain ridge (H 0.7) 1.28 1.21
# rough tree line (H 0.3) 1.48 1.45
# blocks behind trees 1.13 1.09The straight line is estimated almost exactly (1.001 and 1.000); the Koch curve gives 1.300 and 1.245 over the two ranges of scale, either side of the true 1.262; for fractional Brownian profiles, the smooth one is overestimated (, true value 1.1, estimated 1.21 and 1.14) and the rough one underestimated (, true value 1.7, only 1.52 and 1.54). At finite resolution, box counting “compresses” estimates towards the middle, and the result depends on the range of scales used in the fit. A value of means something only once the method and the range of scales are stated.
Taylor and “mid-range complexity”
Evidence
Natural fractal dimensions do not all lie between 1.3 and 1.5. The 2011 review by Taylor and colleagues lists cloud outlines at about 1.30–1.33, coastlines at about 1.05–1.52 and woody plants and trees at about 1.28–1.90, and notes that “many of nature’s fractals cluster around ” (Taylor et al. 2011). Taylor and colleagues in 2005 brought together their series of studies of perceptual and physiological responses to fractal patterns (Taylor et al. 2005).
Evidence
Preference. Spehar and colleagues in 2003 had 220 participants choose between pairs of patterns, including mathematical fractals, natural fractals and Pollock paintings; for all three categories preference peaked at of about 1.3–1.5 (Spehar et al. 2003); Hägerhäll and colleagues in 2004 obtained a similar result with landscape skyline silhouettes, the most preferred being about 1.3 (Hägerhäll et al. 2004). But Spehar, Walker and Taylor showed in 2016 that the group-average peak conceals marked individual differences: in their first experiment about half the participants preferred intermediate complexity, about a fifth preferred smoother patterns and about a fifth rougher ones (Spehar et al. 2016).
Evidence
Physiology. Taylor reported a preliminary study in 2006: 24 participants did mental tasks in a simulated space-station cabin with one image on the wall, and the pattern with was associated with the smallest rise in skin-conductance “stress” (Taylor 2006). Hägerhäll and colleagues in 2008 showed 32 participants (31 with usable EEG data) skyline silhouettes with = 1.14, 1.32, 1.51 and 1.70 for one minute each; the pattern with produced the most marked EEG changes, with the highest frontal alpha and the highest parietal beta (Hägerhäll et al. 2008).
Note
Correction to the author’s framework. First, “the box-counting dimension of natural landscapes clusters at 1.3–1.5” should read: different natural forms have widely spread values of (coastlines 1.05–1.52, trees 1.28–1.90), many clustering around 1.3; 1.3–1.5 is the range where group preference peaks, with clear individual differences. Second, the physiological evidence comes from small preliminary studies (skin conductance in 24 people, EEG in 32), with short-term physiological changes as outcomes rather than health outcomes, and has not been replicated at scale. Third, no direct evidence was found that “ or increases the load on visual cortex”; existing studies show that preference and some EEG measures are higher near intermediate values, which cannot be written as “visual cortical load”. Fourth, Ulrich’s 1984 study did not measure fractal dimension; “trees outside the window” and “” are two independent lines of evidence, and linking them is a hypothesis.
What is the dimension of your view?
The second part of the chapter’s first program generates four synthetic window views as silhouette lines: a row of box-like buildings, a mountain ridge, a rough tree line, and buildings partly hidden by trees, and measures their box-counting dimensions in the same way.
The box-like skyline has a dimension of about 1.0 (1.04 and 1.00), almost a smooth line; the ridge 1.28 and 1.21; the rough tree line 1.48 and 1.45; with a row of trees in front of the buildings the value rises only to 1.13 and 1.09, because the buildings hide most of the tree outline. Over the finer range of scales every estimate is somewhat lower. The lesson: anyone who wants to describe a real window view by “” must specify the viewing distance, image resolution, method of extracting the outline and range of the fit; otherwise numbers measured by different people cannot be compared.
Hypothesis · not established
Natural elements with outline dimensions of 1.3–1.5 in the view from a desk speed recovery from brief stress more than built views with dimensions of about 1.0 (not established). A testable consequence: in a within-participant randomised crossover design, after a standardised mental-arithmetic stressor, participants view for 5 minutes a window scene with , or measured by a common method; recovery of skin conductance and heart-rate variability should be fastest at , and the difference should remain after controlling for whether the scene is natural. If recovery depends only on whether the scene is natural and not on , or the three conditions do not differ, the hypothesis is refuted.
Chapter summary
Ulrich’s 1984 comparison of records (23 pairs; 7.96 versus 8.70 days in hospital) opened research on window views, but it was not a randomised trial and has not been directly replicated; a systematic review gives attention restoration theory only partial support. Fractal dimension can be measured by box counting, but at finite resolution rough curves are underestimated and the result depends on the range of scales. Natural forms have widely spread values of ; 1.3–1.5 is where group preference peaks, with clear individual differences; the skin-conductance and EEG evidence comes from preliminary studies of 24 and 32 people. In the illustrative calculation a box-like skyline has of about 1.0 and a rough tree line about 1.5. The next chapter discusses two other visible spatial variables: light and ceiling height.
References
Fan, Y. (2026). Architectural Acoustics, Reverberation Time and the Environmental Psychology of Built Space: From Sabine's Equation and Fractal Visual Dimensions to Cognitive Performance, Read Beside Classical Chinese Spatial Design [建筑声学、混响时间与空间环境心理学的物理模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23782090 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).