# Derivative at the Edge

My preferred way to approximate a derivative is the central difference, its more accurate than forward difference or backward difference, and I'm too lazy to go higher-order. But the central difference requires a data point on either side of point you are evaluating. Normally this means you end up not having a derivative at either endpoint. To solve it, I want you to switch to forward and backward difference at the edges:

Specifically, I want you to use a forward difference for the first point, a backward difference for the last point, and a central difference for all the points in the middle. Also, you can assume x values are evenly spaced, and focus only on y. Use these formulas: Good luck, I'm looking forward to seeing if someone comes up with a simple rule which reproduces all 3 derivatives in the right places!

EX INPUT:

0.034  9.62    8.885   3.477   2.38


I will use FD, CD, and BD to denote which algorithm to use in which spot, so above 5 points are used to approximate derivatives using

FD     CD      CD      CD     BD


And then the calculated values would be:

9.586  4.4255 -3.0715 -3.2525 -1.097


You can assume that there will always be at least 3 input points, and you can calculate using single or double precision.

And as always, shortest answer wins.

• Just a nitpick, central/forward/backward differences are just approximations of derivatives at a point, not the derivatives themselves. – Liam Jun 21 '16 at 22:08
• I don't understand what each input and output number corresponds to. – xnor Jun 21 '16 at 23:26
• @xnor, i put a brief description between the input and output explaining what algorithm to use for which data point. Does it make sense now? – Tony Ruth Jun 21 '16 at 23:29
• Yes, I think it makes sense. For 5 inputs, you'd do [a,b,c,d,e] -> [b-a,(c-a)/2,(d-b)/2,(e-c)/2,e-d]. Can there be fewer that 3 input points? – xnor Jun 21 '16 at 23:35
• @xnor, that's right. And, I updated so you can assume at least 3 input points. – Tony Ruth Jun 21 '16 at 23:37

# Jelly, 13 10 bytes

I.ịṚjI+2\H


Try it online!

### How it works

I.ịṚjI+2\H  Main link. Argument: A (array)

I           Increments; compute the deltas of consecutive values.
For [a, b, c, d, e], this yields [b-a, c-b, d-c, e-d].
.ị         At-index 0.5; get the the last and first element.
This yields [e-d, b-a].
Ṛ        Reverse the pair.
This yields [b-a, e-d].
jI      Join, separating by the increments.
This yields [b-a, b-a, c-b, d-c, e-d, e-d].
+2\   Add the values of all overlapping pairs.
This yields [2(b-a), c-a, d-b, e-c, 2(e-d)].
H  Halve all resulting numbers.
This yields [b-a, (c-a)/2, (d-b)/2, (e-c)/2, e-d].


# MATL, 21 15 bytes

2/d1)6Mh8Mt0)h+


TryItOnline

Halves the input vector, and takes successive differences, to give d=[i(2)-i(1) i(3)-i(2) ... i(end)-i(end-1)]/2 and then makes two modified vectors, [d(1) d] and [d d(end)], and adds them.

The older version was better (because convolution), but 21 bytes

d1j)6M1)6MTT2/H3\$Y+bv

• I see, pretty clever. So you take a list of forward differences, and a list of backwards differences and average them to get the central difference. Then the endpoints are fixed by instead averaging 2 forward differences or 2 backward differences (in the same spot). Since the forward and backward difference are just shifted from each other by one spot, you get to reuse a lot of structure. – Tony Ruth Jun 22 '16 at 21:41
• Just forward differences, otherwise yes. Doing (y(i)-y(i-1))+(y(i+1)-y(i)) gives y(i+1)-y(i-1), which is twice the centred difference. – David Jun 22 '16 at 22:35

## Python with NumPy, 29 bytes

import numpy;numpy.gradient


This happens to be the default behaviour of NumPy's gradient function. The bytes were counted according to this consensus.

## 05AB1E, 201917 14 bytes

¥Ð¦øO;s¤s0èŠ)˜


Explained

¥Ð              # triplicate deltas of list
[9.585999999999999, -0.7349999999999994, -5.4079999999999995, -1.097]
¦øO;          # get central difference (fold addition over deltas and divide by 2)
[4.4254999999999995, -3.0714999999999995, -3.2524999999999995]
s¤        # get backwards difference
-1.097
s0è     # get forwards difference
9.585999999999999
Š)˜  # reorder differences, merge to list and flatten
[9.585999999999999, 4.4254999999999995, -3.0714999999999995, -3.2524999999999995, -1.097]


Try it online

Saved 2 bytes thanks to @Adnan

# Julia, 8 bytes

gradient


Inspired by @MartinEnder's Python answer. Try it online!

• This also works in Matlab and in Octave – Luis Mendo Jun 22 '16 at 17:05

# Pyth, 14 bytes

.OM>L2._seB-Vt


Try it online: Demonstration

### Explanation:

.OM>L2._seB-VtQQ   implicitly add two Qs (input arrays) at the end
-VtQQ   get all neighbored differences
seB        get the last element of ^ and append it to ^
._           compute all prefixes
>L2             reduce all prefixes to the last two elements
.OM                compute the average of each ^


# J, 21 bytes

[:((,{:)+{.,])2-~/\-:


Similar to the approach used in @David's solution.

## Usage

   f =: [:((,{:)+{.,])2-~/\-:
f 0.034 9.62 8.885 3.477 2.38
9.586 4.4255 _3.0715 _3.2525 _1.097


## Explanation

[:((,{:)+{.,])2-~/\-:  Input: list A
-:  Halve each value in A
2   \    Select each overlapping sublist of size 2 in A
-~/     Reduce it using subtraction to get the difference
[:(          )         Operate on the list of differences, call it D
]          Identity function, returns D
{.            Get the head of D
,           Join them to get [head(D), D]
( {:)               Get the tail of D
,                  Join them to get [D, tail(D)]
+              Add them together elementwise to get the derivatives and return


# Pyth - 29 bytes

Stupid simple approach.

s[_-F<Q2mc-@Qhd@Qtd2tUtQ_-F>2


## JavaScript (ES6), 62 bytes

a=>a.map((_,i)=>i&&i--<a.length-2?(a[i+2]-a[i])/2:a[i+1]-a[i])


# Pyth, 272423 21 bytes

.bcF_-VNYK++]hJ,VUQQJ]eJttK
.bcF-VYN+]hJ,VQUQJ+tJ]eJ
++hJ-V+tQeQ+hQQcR2PtJeJ
*V++1*].5ttlQ1-V+tQeQ+h
*V++1m.5ttQ1-V+tQeQ+h


Try it online!