I am trying to run a nls() to interpolate height growth of a tree given a series of samples, where each sample has a value of age, and height. I am using the Richards (1959) method:

$$ \textrm{Height}=x\left(1-e^{-y\cdot\textrm{Age}}\right)^z $$

my age and height values:

> age
[1]  0  3  6 17 29 47 57 63 72
> height
[1]   37   70  199  464  736 1008 1280 1462 1643

the script I am using to calculate parameters x, y, and z, is:

b <- environment(growCurve[["m"]][["incr"]])[["internalPars"]][1]
y <- environment(growCurve[["m"]][["incr"]])[["internalPars"]][2]
z <- environment(growCurve[["m"]][["incr"]])[["internalPars"]][3]


when I run this however, I get the error

singular gradient matrix at initial parameter estimates

I tried using another fix from this post, but I still get identical results so I am not really sure on how to proceed.

  • 2
    $\begingroup$ $(0,0,0)$ is a particularly bad starting value for this model: it sits in one corner of the space of possible values. Choose something better. My answer in the thread you reference provides some ways to come up with better starting values. $\endgroup$
    – whuber
    Commented Dec 14, 2022 at 20:00
  • $\begingroup$ @whuber I did try that solution you suggested, however when I take those new output values using that solution and put them back into the nls() I still get the exact same error. $\endgroup$ Commented Dec 14, 2022 at 20:03
  • 1
    $\begingroup$ Plotting these data reveals that your model cannot be identified: it isn't distinguishable from a model of the form Height = b * Age. You could fit it with lm(height ~ age - 1). To check whether a nonlinear model might be an improvement, compare it to lm(height ~ age + I(age^2)) Neither the intercept nor the quadratic term improve things. $\endgroup$
    – whuber
    Commented Dec 14, 2022 at 20:05
  • 1
    $\begingroup$ @kjetilbhalvorsen A Flexible Growth Function for Empirical Use, Richards (1959). Journal of Experimental Botany, Volume 10, Issue 2, June 1959, Pages 290–301, doi.org/10.1093/jxb/10.2.290 $\endgroup$ Commented Dec 14, 2022 at 21:29
  • 1
    $\begingroup$ @Sal Or perhaps the age is in days and the heights are in millimeters :-). $\endgroup$
    – whuber
    Commented Dec 16, 2022 at 14:42

3 Answers 3


There are some things we can do. These are generally useful to know for any kind of optimization problem and are not limited to the nls function in R.

The first is to parameterize the model to make it everywhere differentiable in its coefficients and automatically constrain the parameters to reasonable ranges. These steps make it harder for the algorithm to have numerical problems or wander into forbidden territory. One way (of many possibilities) is to write $x$ as an exponential (making it positive), $y$ as a square (ditto), and $z$ -- which in the Richards paper is close to $1$ -- in the form $1 + \cos(t)$ for some angle $t.$

The second is to find a good set of starting values. I discovered one set by experimenting with a fit by eye, shown as the dotted curve in the plot below.

The third is to prevent nls from exiting, so even if it fails to find a solution, you can inspect the solution.

The fourth is to use suitable weights. I speculate that higher heights ought to be fit with larger error and so use weights proportional to height. This makes little difference here, but can sometimes be helpful and more statistically appropriate.

This procedure finds the solution plotted in red. You can see it follows the data closely and is slightly better than the initial eyeball fit.

enter image description here

The fitted values (in terms of the original parameters) are $x = 134571,$ $y = 0.000118,$ and $z = 0.926.$

The problem is evident: these data trend almost linearly, which is not enough to determine all three parameters in the model. Given that the model forces the curve through the origin, a linear trend uses just one parameter, making the other two parameters superfluous.

To verify this, analyze the correlation matrix of the estimates. In this case the eigenvectors are approximately $2.96,$ $0.039,$ and $10^{-8}.$ The last two are essentially zero compared to the first. The corresponding eigenvectors correspond roughly to (a) a contrast between $x$ and the other parameters (increases in $x$ are almost perfectly offset by decreases in $y$ and $z$), (b) the $z$ parameter, and (c) the $x$ and $y$ parameters, in that order. This tells us the model is confident about a particular combination of $x,$ $y,$ and $z$ (which determines the slope for small ages) but can glean essentially no specific information about $z$ or $y$ from these data, apart from a tiny hint of curvature suggested by the fact the second eigenvalue is not truly zero.

Thus, a least squares fit ought to work about as well. Using fit <- lm(height ~ 0 + age) I found a slope of $22.89.$ Plotting this line (through the origin) gives a curve barely distinguishable from the red curve above, so I haven't shown it.

Here is the R code showing the details.

# The model.
f <- function(age, beta) {
  x <- beta["x"]
  y <- beta["y"]
  z <- beta["z"]
  exp(x) * (1 - exp(-y^2 * age))^(1 + sin(z))
# The data.
age <- scan(text = "0  3  6 17 29 47 57 63 72")
height <- scan(text = "37   70  199  464  736 1008 1280 1462 1643")
plot(age, height)
# Draw an eyeball fit.
# These parameters correspond to an exponential curve (z == 0) that over the
# range of ages in the data hardly departs from linearity (age * y is always
# tiny).  Only `x` has to be estimated; it determines the slope, which is
# easy to gauge from the plots.
s <-  c(x = 9, y = sqrt(1/300), z = 0)
curve(f(x, s), add = TRUE, lwd = 2, col = gray(0.3), lty = 3)
# Polish this fit.
# The `control` object shows how to specify some of the most useful aspects
# of the search.
obj <- nls(height ~ exp(x) * (1 - exp(-y^2 * age))^(1 + sin(z)), 
    start = s, weights = height, 
    control = list(minFactor = 2^(-16), maxiter = 1e4, warnOnly = TRUE))
# Plot the fit.
b <- coefficients(obj)
curve(f(x, b), add = TRUE, lwd = 2, col = hsv(0.02, 1, 0.85))
# Report the estimates in terms of the original parameterization.
c(x = exp(b["x"]), y = b["y"]^2, z = 1 + sin(b["z"]))
# Analyze the correlations among the estimates.
V <- vcov(obj)
v <- sqrt(diag(V))
V <- t(V/v)/v
eigen(V) # Alternatively, `eigen(cov2cor(vcov(obj)))`
# Look at a simpler alternative model.
fit <- lm(height ~ 0 + age)
abline(fit, col = "Blue")
  • 1
    $\begingroup$ If the parameters have some biological/chemical/physical interpretation, they might have priors, possibly even strong priors, and a Bayesian non-linear model might also be an option. $\endgroup$
    – Roland
    Commented Dec 15, 2022 at 7:01
  • 3
    $\begingroup$ Two minor comments: i) In the code, you define b (the coefficients) after converting them in the original parametrization, which leads to an error because b has not yet been defined. ii) An easier (but probably less instructive) way to convert V to a correlation matrix would be the built-in function cov2cor. $\endgroup$ Commented Dec 15, 2022 at 7:03
  • 2
    $\begingroup$ @COOLSerdash I apologize for the error--I reordered the code before posting it and neglected to rerun the entire block. I'll fix it. Thanks for the cov2cor tip. $\endgroup$
    – whuber
    Commented Dec 15, 2022 at 14:33

The problem seems to be in convergence of the z parameter so fix it at each of a sequence of values and optimize on the others for each and then take the result with the lowest residual sum of squares. We use the plinear algorithm of nls which does not require starting values for the parameters that enter linearly -- x enters linearly so its starting value can be omitted. Since we are fixing z and since x enters linearly we only have to specify a starting value for y. To use plinear the right hand side should be a column that multiplies x but x itself is not specified (in a similar way to lm).

age <- c(0, 3, 6, 17, 29, 47, 57, 63, 72)
height <- c(37, 70, 199, 464, 736, 1008, 1280, 1462, 1643)

fo <- "height ~ cbind(x = (1 - exp(-y * age))^z)"
zseq <- seq(0, 2, 0.01)
sse <- sapply(zseq, function(z) {
  res <- try(nls(fo, start = list(y = 1), algorithm = "plinear"))
  if (inherits(res, "try-error")) NA else deviance(res)

z <- zseq[which.min(sse)]
z # z value which minimizes the SSE
## [1] 1

fm <- nls(fo, start = list(y = 1), algorithm = "plinear")


Nonlinear regression model
  model: height ~ cbind(x = (1 - exp(-y * age))^z)
   data: parent.frame()
        y    .lin.x 
3.384e-03 7.468e+03 
 residual sum-of-squares: 18004

Number of iterations to convergence: 9 
Achieved convergence tolerance: 1.103e-06

Plotting the solution

plot(height ~ age)
lines(fitted(fm) ~ age)

(continued after chart) screenshot

Note that a simple linear regression actually gives a better fit (lower residual sum of squares) and has fewer parameters (2 vs 3).

fm.lin <- lm(height ~ age)
## [1] 13546.89

## [1] 18003.86

The following is just to show that the deviance function in R is equivalent to taking the residual sum of squares for both nls and lm.

## [1] 13546.89

## [1] 18003.86
  • $\begingroup$ The simple linear regression is not nested within the original model (which has no additive intercept term), so it might be misleading to compare the residual sums of squares. I'm not sure these two deviances are even measuring the same thing--the reported reduction looks too great. $\endgroup$
    – whuber
    Commented Dec 15, 2022 at 14:53
  • $\begingroup$ It doesn't matter whether they are nested or not unless you are doing likelihood ratio tests. They are both measuring the sum of squares of height - fitted so they are certainly measuring the same thing. If you are unsure about the R deviance function you can get the exact same results using sum(resid(fm)^2) and sum(resid(fm.lin)^2) $\endgroup$ Commented Dec 15, 2022 at 15:12
  • $\begingroup$ I did that before posting my comment and got different results, with much closer deviances. The reason why the comparison might be misleading is that the two models have different meanings: being able to reduce the deviance by introducing a behavior (the intercept) that was not reflected in the original model might be scientifically inappropriate. In my experiments yesterday I could not even establish that the intercept was statistically significant. $\endgroup$
    – whuber
    Commented Dec 15, 2022 at 15:22
  • $\begingroup$ You must have made some computer error. I have added the results to the end of the answer and you can see they are identical. There is not much data here so I would not rely on statistical significance. This is more exploratory. $\endgroup$ Commented Dec 15, 2022 at 15:28
  • $\begingroup$ I made no error. My solution has a smaller deviance than yours and I used my solution. $\endgroup$
    – whuber
    Commented Dec 15, 2022 at 15:36

Minor point: Don't try to recreate the exponential function in your code

The excellent answer by whuber shows how you can use the data to choose better starting values for your iteration and how you can constrain the search space appropriately. This is something you should usually do when fitting nonlinear models using an iterative method. One other minor thing I notice in your code is that you define your own constant e and then take powers of this in later steps to yield your own exponential function. That is unecessary and it introduces additional rounding error. Instead of trying to create the exponential function yourself, use the exp function built into the program. (You will see that whuber has done this in his own code.)

  • 1
    $\begingroup$ This is a good point, but I don't see how using a specified constant for e introduces any rounding error to be concerned about in the modeling. The reason is that I would expect e^y to be implemented as $\exp(y\log(e)),$ thereby requiring an additional multiplication, but there's really no harm done except perhaps in the 52nd binary digit of the mantissa. Now, if by "rounding error" you mean in the reporting of the estimates, that's spot on due to the error in equating $\exp(1)$ with $2.71828,$ which is off by almost one part per million. $\endgroup$
    – whuber
    Commented Dec 14, 2022 at 23:01
  • $\begingroup$ I agree with your description of the implementation and effect. Thus, why add a rounding error of one-part-in-a-million in $\log(e)$? Why not use exp instead and not have that rounding error? Also, if someone sees the coding method in OP and thinks this is how to do things, maybe they will think it's similarly okay to use e <- 2.718 (error now one-part-in-ten-thousand) or e <- 2.72 (error now six-parts-in-ten-thousand). Better to nip this method in the bud, no? $\endgroup$
    – Ben
    Commented Dec 14, 2022 at 23:17
  • $\begingroup$ I agree -- it's a useful coding comment. $\endgroup$
    – whuber
    Commented Dec 15, 2022 at 14:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.